Add files using upload-large-folder tool

README.md

@@ -0,0 +1,73 @@
+---
+pretty_name: Strange Loop in J-space
+tags:
+- interpretability
+- j-space
+- global-workspace
+- self-model
+- recurrence
+- qwen3
+size_categories:
+- 10K<n<100K
+---
+
+# Strange Loop in J-space: data
+
+All data from an experiment in which a language model wanders with no memory of its own words, only a latent J-space
+state. Every ten thoughts it reads its last ten states and forms a self-model, which is then fed back into everything it
+thinks next.
+
+- **Interactive report:** https://huggingface.co/spaces/Pixedar/strange-loop-in-j-space
+- **Code:** `strange_loop.py`, `analyze.py`, `export_report.py` in this repository (also on GitHub, Pixedar)
+
+Qwen3-4B-Base · J-lens at layer 16 · 3 runs · 180 minds · 72,000 thoughts.
+
+## Short result
+
+The causal loop is real. A model of the trajectory, rebuilt from the trajectory, shapes the trajectory, sustains itself,
+and holds long-range coherence that vanishes when it is cut. A held foreign self-model persists after release. But the
+loop does not single out its own self-model: another mind's self-model fed back works almost exactly as well, and the
+self-model predicts the future only by causing it. See the report for every number.
+
+## Runs
+
+| Directory | Minds | Conditions | J-space dictionary |
+|---|---|---|---|
+| `runs/mainA` | 60 = 6 × 10 seeds | free, loop, yoked, avg, perturb, ablate (treated at thought 200) | every word-like token (40,034 atoms) |
+| `runs/mainB` | 60 = 6 × 10 seeds | same as A | common English words (27,929 atoms) |
+| `runs/mainC` | 60 = 4 × 15 seeds | free, loopc, yokedc, clampc (individual self-models; clamp 150–199; forward question) | common English, natural case (26,236 atoms) |
+| `runs/pilot2` | 23 | recurrent gain sweep, reflection layer/gain sweep | every word-like token |
+
+Each run has 400 thoughts per mind and a reflection every 10 thoughts (40 per mind). Minds with the same seed share
+their sampling noise, so a treated mind and its twin are identical until they are treated differently.
+
+## Files per run
+
+- `traj.npz` (raw arrays, float16 unless noted). N = thoughts, B = minds, R = reflections, d = 2560, k = 16:
+  - `j` (N, B, d): J-space content of each thought, measured against the typical workspace (`jbar`)
+  - `j_idx` (N, B, k) int32 and `j_coef` (N, B, k): its word atoms (indices into `meta.json` → `words`) and weights
+  - `m` (R, B, d), `m_idx`, `m_coef`: the self-model at each reflection, against the empty-trace answer (`mbar`)
+  - `inj` (R, B, d): the vector fed back after each reflection. `inj_src` (R, B) int32 is its source: the row whose
+    self-model it is, −1 for none, −2 for the plain average
+  - `mf` (R, B, d), run C only: the self-model under the forward question ("my mind will probably turn to"). `generic`
+    (R, d): the observers' mean self-model
+  - `fid_<layer>_<gain>`: self-models from the same reflection injected at other layers (fidelity check)
+  - `mu`, `jbar`, `mbar`: calibration means
+  - The recurrent state is not stored. It is `z_t = 0.6 z_{t−1} + 0.4 j_t` with `z_0 = 0`
+- `meta.json`: `rows` (condition and seed of each mind), `args` (every setting), `texts[t][b]` (each thought, 40
+  tokens), `desc[r][b]` (the spoken answer to the reflection question), `fdesc` (answer to the forward question, run C),
+  `words` (the J-space dictionary), `thought_frame`, `reflect_prompt`, `forward_prompt`, calibration norms
+- `results.json`: everything `analyze.py` computes (autocorrelation, self-model stability, prediction, perturbation,
+  ablation, clamp, themes, 3-D projection of every path)
+- `summary.txt`: the console summary of `analyze.py`
+
+Also included:
+- `qwen3-4b-base-jlens.npz`: the J-lens (average Jacobian from layer 16 to the final layer, 120 prompts) that every
+  run used
+- `runs/*.log`: the run logs
+
+## Notes
+
+Thoughts are unfiltered base-model output from a think-aloud frame and occasionally contain profanity or sexual
+content. The J-space word lists of run A are dominated by under-trained and non-English tokens; prefer runs B and C for
+reading words.

analyze.py

@@ -0,0 +1,366 @@
+"""Analysis of a strange_loop.py run: the J-space path, the self-models, and the tests asked of the loop.
+
+    python analyze.py runs/mainA            -> runs/mainA/results.json (numbers + what the report draws)
+
+All similarities are cosines between J-space content vectors measured relative to the typical workspace (so two
+unrelated thoughts score ~0). Statistics are over seeds (each seed = one independent mind per condition); paired
+comparisons use the same seed in both conditions (they share their sampling noise and are identical until treated).
+"""
+import json, sys
+import numpy as np
+
+d = sys.argv[1]
+Z = np.load(f'{d}/traj.npz');M = json.load(open(f'{d}/meta.json', encoding='utf-8'))
+rows, a = M['rows'], M['args'];K, T_INT = a['K'], a['t_int'];B = len(rows)
+INDIVIDUATED = any(r['cond'] in ('loopc', 'yokedc', 'clampc') for r in rows)
+for r in rows:r['cond'] = {'loopc':'loop', 'yokedc':'yoked', 'clampc':'clamp'}.get(r['cond'], r['cond'])
+words = np.array(M['words'])
+C = Z['j'].astype(np.float32);N = C.shape[0]                             # (N, B, d) thought contents
+MS = Z['m'].astype(np.float32);NR = MS.shape[0]                           # (NR, B, d) self-models
+INJ = Z['inj'].astype(np.float32);SRC = Z['inj_src']
+R_INT = T_INT//K-1                                                        # reflection whose output perturb replaces
+conds = list(dict.fromkeys(r['cond'] for r in rows));seeds = sorted({r['seed'] for r in rows})
+FREE = [b for b, r in enumerate(rows) if r['cond'] == 'free']
+row = {(r['cond'], r['seed']):b for b, r in enumerate(rows)}
+unit = lambda x:x/np.maximum(np.linalg.norm(x, axis=-1, keepdims=True), 1e-8)
+cos = lambda x, y:(unit(x)*unit(y)).sum(-1)
+alpha = a['alpha']
+Zs = np.zeros_like(C);s = np.zeros_like(C[0])
+for t in range(N):
+    s = (1-alpha)*s+alpha*C[t];Zs[t] = s                                 # recurrent state z_t (after thought t)
+WIN = np.stack([C[r*K:(r+1)*K].mean(0) for r in range(NR)])               # mean content of each window
+Cu, MSu = unit(C), unit(MS)
+
+
+def ms(v):
+    v = np.asarray(v, float);n = len(v)
+    return dict(mean=float(v.mean()), se=float(v.std(ddof=1)/np.sqrt(n)) if n > 1 else 0., n=n)
+
+
+def paired(x, y):
+    """x - y over seeds: mean, se, t, and how many seeds go the same way."""
+    d_ = np.asarray(x, float)-np.asarray(y, float);n = len(d_)
+    se = d_.std(ddof=1)/np.sqrt(n) if n > 1 else 0.
+    return dict(diff=float(d_.mean()), se=float(se), t=float(d_.mean()/se) if se > 0 else 0., pos=int((d_ > 0).sum()), n=n)
+
+
+out = dict(args=a, conds=conds, seeds=seeds, N=N, NR=NR, K=K, t_int=T_INT, rc=M['rc'], Rn=M['Rn'])
+
+# ------------------------------------------------------------------ 1. wandering: autocorrelation of the J-space path
+lags = list(range(1, 61))
+def acf(b, t0, t1, proj=None):
+    x = C[t0:t1, b].copy()
+    if proj is not None:                                                  # remove the injected self-model direction
+        p = unit(proj[t0:t1]);x = x-(x*p).sum(-1, keepdims=True)*p
+    xu = unit(x)
+    return [float((xu[:-L]*xu[L:]).sum(-1).mean()) if L < len(xu) else np.nan for L in lags]
+
+inj_at = np.zeros_like(C)                                                 # the vector injected during thought t
+for r in range(NR-1):inj_at[(r+1)*K:(r+2)*K] = INJ[r]
+out['acf'] = {}
+for c in conds:
+    for ph, (t0, t1) in {'pre':(20, T_INT), 'post':(T_INT, N)}.items():
+        A = np.array([acf(row[c, s_], t0, t1) for s_ in seeds])
+        Ap = np.array([acf(row[c, s_], t0, t1, inj_at[:, row[c, s_]]) for s_ in seeds])
+        ch = np.mean([cos(C[t0:t1, row[c, s1]], C[t0:t1, row[c, s2]]).mean() for s1 in seeds for s2 in seeds if s1 < s2])
+        out['acf'][f'{c}/{ph}'] = dict(mean=A.mean(0).tolist(), se=(A.std(0, ddof=1)/np.sqrt(len(seeds))).tolist(),
+                                       proj_mean=Ap.mean(0).tolist(), chance=float(ch),
+                                       long=A[:, 19:60].mean(1).tolist(), long_proj=Ap[:, 19:60].mean(1).tolist(),
+                                       short=A[:, 0:5].mean(1).tolist())
+out['lags'] = lags
+# over time: each mind's coherence with itself 20 thoughts earlier, and how alike different minds are at the same moment
+out['coh_time'], out['cross_time'] = {}, {}
+for c in conds:
+    bs = [row[c, s_] for s_ in seeds]
+    out['coh_time'][c] = np.round(np.mean([cos(C[20:, b], C[:-20, b]) for b in bs], 0), 4).tolist()
+    out['cross_time'][c] = np.round(np.mean([cos(C[:, b1], C[:, b2]) for b1 in bs for b2 in bs if b1 < b2], 0), 4).tolist()
+
+# ------------------------------------------------------------------ 2. the self-models: a persistent self-symbol?
+sm = {}
+for c in conds:
+    bs = [row[c, s_] for s_ in seeds]
+    cons = np.array([[cos(MS[k, b], MS[k+1, b]) for k in range(NR-1)] for b in bs])          # (S, NR-1)
+    wcons = np.array([[cos(WIN[k, b], WIN[k+1, b]) for k in range(NR-1)] for b in bs])       # content's own
+    lag = {L:np.array([[cos(MS[k, b], MS[k+L, b]) for k in range(NR-L)] for b in bs]).mean(1) for L in (1, 2, 5, 10)}
+    wlag = {L:np.array([[cos(WIN[k, b], WIN[k+L, b]) for k in range(NR-L)] for b in bs]).mean(1) for L in (1, 2, 5, 10)}
+    cross = np.array([np.mean([cos(MS[k, b1], MS[k, b2]) for b1 in bs for b2 in bs if b1 < b2]) for k in range(NR)])
+    rec = np.array([MSu[:, b]@MSu[:, b].T for b in bs]).mean(0)                              # (NR, NR) recurrence
+    # individuality: is a row's early self-model closer to its own later self-symbol than to other rows'?
+    # (both inside the pre-intervention period, so all conditions are measured alike)
+    half = R_INT//2
+    sym = unit(np.stack([MS[half:R_INT, b].mean(0) for b in bs]))
+    # reconstruction: does the next self-model re-form what was fed in (own in loop, foreign in yoked)?
+    recon = [np.mean([cos(MS[k+1, b], INJ[k, b]) for k in range(R_INT-1) if SRC[k, b] != -1]) if c != 'free' else np.nan
+             for b in bs]
+    generic = unit(MS[:R_INT].reshape(-1, MS.shape[-1]).mean(0))
+    gen_share = [float(np.mean(cos(MS[:R_INT, b], generic))) for b in bs]
+    own_sym = [np.mean([cos(MS[k, b], sym[i]) for k in range(half)]) for i, b in enumerate(bs)]
+    oth_sym = [np.mean([cos(MS[k, b], sym[j]) for k in range(half) for j in range(len(bs)) if j != i]) for i, b in enumerate(bs)]
+    sm[c] = dict(cons_time=cons.mean(0).tolist(), cons_time_se=(cons.std(0, ddof=1)/np.sqrt(len(bs))).tolist(),
+                 wcons_time=wcons.mean(0).tolist(), cross_time=cross.tolist(),
+                 cons_pre=cons[:, 2:R_INT].mean(1).tolist(), wcons_pre=wcons[:, 2:R_INT].mean(1).tolist(),
+                 cons_post=cons[:, R_INT+1:].mean(1).tolist(), wcons_post=wcons[:, R_INT+1:].mean(1).tolist(),
+                 lag={L:v.tolist() for L, v in lag.items()}, wlag={L:v.tolist() for L, v in wlag.items()},
+                 cross_pre=float(cross[2:R_INT].mean()), recurrence=np.round(rec, 3).tolist(),
+                 own_symbol=own_sym, other_symbol=oth_sym, recon=[float(x) for x in recon], generic_share=gen_share)
+out['selfmodel'] = sm
+
+# individuated self-models: what sets one mind's self-model apart from the population's generic one (the mean over
+# the observer rows at the same reflection, as the individuated conditions of run C feed back)
+GEN = MS[:, FREE].mean(1);MSI = MS-GEN[:, None]
+out['generic_norm_share'] = float(np.linalg.norm(GEN, axis=-1).mean()/np.linalg.norm(MS, axis=-1).mean())
+for c in conds:
+    bs = [row[c, s_] for s_ in seeds]
+    sm[c]['cons_ind_pre'] = [float(np.mean([cos(MSI[k, b], MSI[k+1, b]) for k in range(2, R_INT)])) for b in bs]
+    sm[c]['cross_ind_pre'] = float(np.mean([cos(MSI[k, b1], MSI[k, b2]) for k in range(2, R_INT) for b1 in bs for b2 in bs if b1 < b2]))
+    sm[c]['recon_ind'] = [float(np.mean([cos(MSI[k+1, b], MSI[k, SRC[k, b]]) for k in range(R_INT-1) if SRC[k, b] >= 0]))
+                          if c not in ('free', 'avg') else np.nan for b in bs]
+    sm[c]['cons_ind_time'] = np.array([[cos(MSI[k, b], MSI[k+1, b]) for k in range(NR-1)] for b in bs]).mean(0).tolist()
+out.setdefault('selfmodel_tests', {})
+if 'yoked' in conds:
+    out['selfmodel_tests']['loop_vs_yoked_cons_ind'] = paired(sm['loop']['cons_ind_pre'], sm['yoked']['cons_ind_pre'])
+    out['selfmodel_tests']['loop_vs_yoked_recon_ind'] = paired(sm['loop']['recon_ind'], sm['yoked']['recon_ind'])
+    out['selfmodel_tests']['loop_vs_free_cons_ind'] = paired(sm['loop']['cons_ind_pre'], sm['free']['cons_ind_pre'])
+    out['selfmodel_tests']['yoked_vs_free_cons_ind'] = paired(sm['yoked']['cons_ind_pre'], sm['free']['cons_ind_pre'])
+
+# self-prediction: the forward question ("in the next few moments, my mind will probably turn to")
+if 'mf' in Z.files:
+    MF = Z['mf'].astype(np.float32);fw = {}
+    for c in conds:
+        bs = [row[c, s_] for s_ in seeds];rr = [r for r in range(1, NR-1) if not (c in ('perturb', 'ablate', 'clamp') and r >= R_INT)]
+        res = {k:[] for k in ('fwd', 'back', 'fwd_other', 'fwd_perp', 'back_perp', 'fwd_other_perp', 'fwd_lean', 'back_lean', 'fwd_back_sim')}
+        for b in bs:
+            acc = {k:[] for k in res}
+            for r in rr:
+                F, P = WIN[r+1, b], WIN[r, b];Fp = F-(F@unit(P))*unit(P);oth = [b2 for b2 in bs if b2 != b]
+                acc['fwd'].append(cos(MF[r, b], F));acc['back'].append(cos(MS[r, b], F))
+                acc['fwd_other'].append(np.mean([cos(MF[r, b2], F) for b2 in oth]))
+                acc['fwd_perp'].append(cos(MF[r, b], Fp));acc['back_perp'].append(cos(MS[r, b], Fp))
+                acc['fwd_other_perp'].append(np.mean([cos(MF[r, b2], Fp) for b2 in oth]))
+                acc['fwd_lean'].append(cos(MF[r, b], F)-cos(MF[r, b], P));acc['back_lean'].append(cos(MS[r, b], F)-cos(MS[r, b], P))
+                acc['fwd_back_sim'].append(cos(MF[r, b], MS[r, b]))
+            for k in res:res[k].append(float(np.mean(acc[k])))
+        fw[c] = res
+    out['forward'] = fw
+    out['forward_tests'] = {c:dict(fwd_vs_other=paired(fw[c]['fwd'], fw[c]['fwd_other']),
+                                   fwd_vs_back=paired(fw[c]['fwd'], fw[c]['back']),
+                                   perp_fwd_vs_other=paired(fw[c]['fwd_perp'], fw[c]['fwd_other_perp']),
+                                   perp_fwd_vs_back=paired(fw[c]['fwd_perp'], fw[c]['back_perp']),
+                                   lean_fwd_vs_back=paired(fw[c]['fwd_lean'], fw[c]['back_lean'])) for c in conds}
+    out['forward_answers'] = [[M['fdesc'][r][b].strip()[:120] for r in range(NR)] for b in range(B)]
+
+# ------------------------------------------------------------------ 3. prediction: does the self-model predict the future?
+# For reflection r (made after window r), the future is the next window's mean content F = WIN[r+1].
+pred = {}
+for c in conds:
+    bs = [row[c, s_] for s_ in seeds];res = {k:[] for k in ('own', 'other', 'past', 'inj', 'own_perp', 'other_perp',
+                                                              'inj_perp', 'own_retr', 'own_past', 'own_fut_minus_past')}
+    rr = [r for r in range(1, NR-1) if not (c in ('perturb', 'ablate') and r >= R_INT)]
+    for i, b in enumerate(bs):
+        o, ot, p, inj, op, otp, ip, rt, opst, ofmp = ([] for _ in range(10))
+        for r in rr:
+            F, P = WIN[r+1, b], WIN[r, b]
+            Fp = F-(F@unit(P))*unit(P)                                     # the part of the future not in the past
+            o.append(cos(MS[r, b], F));p.append(cos(P, F));opst.append(cos(MS[r, b], P))
+            others = [b2 for b2 in bs if b2 != b]
+            ot.append(np.mean([cos(MS[r, b2], F) for b2 in others]))
+            op.append(cos(MS[r, b], Fp));otp.append(np.mean([cos(MS[r, b2], Fp) for b2 in others]))
+            if SRC[r, b] != -1:inj.append(cos(INJ[r, b], F));ip.append(cos(INJ[r, b], Fp))
+            sc = {b2:cos(MS[r, b], WIN[r+1, b2]) for b2 in bs}             # retrieval of one's own future
+            rt.append(float(max(sc, key=sc.get) == b))
+            ofmp.append(cos(MS[r, b], F)-cos(MS[r, b], P))
+        for k, v in zip(('own', 'other', 'past', 'own_perp', 'other_perp', 'own_retr', 'own_past', 'own_fut_minus_past'),
+                        (o, ot, p, op, otp, rt, opst, ofmp)):res[k].append(float(np.mean(v)))
+        if inj:res['inj'].append(float(np.mean(inj)));res['inj_perp'].append(float(np.mean(ip)))
+    pred[c] = res
+out['prediction'] = pred
+out['prediction_tests'] = {c:dict(own_vs_other=paired(pred[c]['own'], pred[c]['other']),
+                                  perp_own_vs_other=paired(pred[c]['own_perp'], pred[c]['other_perp']),
+                                  retrieval=ms(pred[c]['own_retr']),
+                                  future_vs_past=paired(pred[c]['own'], pred[c]['own_past']))
+                           for c in conds}
+if 'yoked' in conds:
+    out['prediction_tests']['yoked_own_vs_injected'] = paired(pred['yoked']['own'], pred['yoked']['inj'])
+    out['prediction_tests']['yoked_own_vs_injected_perp'] = paired(pred['yoked']['own_perp'], pred['yoked']['inj_perp'])
+if 'avg' in conds:
+    out['prediction_tests']['avg_own_vs_injected'] = paired(pred['avg']['own'], pred['avg']['inj'])
+
+# fidelity: does the self-model match its own past window better than the other rows' windows?
+fid = {}
+for name in ['m']+[k for k in Z.files if k.startswith('fid_')]:
+    Mx = Z[name].astype(np.float32);own, oth, hit = [], [], []
+    for c in conds:
+        bs = [row[c, s_] for s_ in seeds]
+        for r in range(NR):
+            for b in bs:
+                sc = {b2:cos(Mx[r, b], WIN[r, b2]) for b2 in bs}
+                own.append(sc[b]);oth.append(np.mean([v for k, v in sc.items() if k != b]));hit.append(max(sc, key=sc.get) == b)
+    fid[name] = dict(own=float(np.mean(own)), other=float(np.mean(oth)), retrieval=float(np.mean(hit)), chance=1/len(seeds))
+out['fidelity'] = fid
+
+# ------------------------------------------------------------------ 4. perturbation: replace the self-model once
+if 'perturb' in conds:
+    per = dict(traj_to_foreign=[], twin_traj_to_foreign=[], twin_div=[], self_to_old=[], self_to_foreign=[],
+               twin_self_to_old=[], twin_self_to_foreign=[])
+    for s_ in seeds:
+        bp, bl = row['perturb', s_], row['loop', s_]
+        mstar = INJ[R_INT, bp];mold = MS[R_INT, bl]
+        per['traj_to_foreign'].append(cos(C[T_INT-40:N, bp], mstar).tolist())
+        per['twin_traj_to_foreign'].append(cos(C[T_INT-40:N, bl], mstar).tolist())
+        per['twin_div'].append(cos(C[T_INT-40:N, bp], C[T_INT-40:N, bl]).tolist())
+        per['self_to_old'].append([cos(MS[k, bp], mold) for k in range(R_INT-4, NR)])
+        per['self_to_foreign'].append([cos(MS[k, bp], mstar) for k in range(R_INT-4, NR)])
+        per['twin_self_to_old'].append([cos(MS[k, bl], mold) for k in range(R_INT-4, NR)])
+        per['twin_self_to_foreign'].append([cos(MS[k, bl], mstar) for k in range(R_INT-4, NR)])
+    agg = {k:dict(mean=np.mean(v, 0).tolist(), se=(np.std(v, 0, ddof=1)/np.sqrt(len(seeds))).tolist()) for k, v in per.items()}
+    sl = lambda v, i0, i1:np.array(v)[:, i0:i1].mean(1)
+    agg['tests'] = dict(
+        traj_reorganises=paired(sl(per['traj_to_foreign'], 40, 40+K*3), sl(per['twin_traj_to_foreign'], 40, 40+K*3)),
+        traj_late=paired(sl(per['traj_to_foreign'], 40+K*5, None), sl(per['twin_traj_to_foreign'], 40+K*5, None)),
+        self_next_to_foreign=paired(sl(per['self_to_foreign'], 5, 6), sl(per['twin_self_to_foreign'], 5, 6)),
+        self_late_to_foreign=paired(sl(per['self_to_foreign'], 8, None), sl(per['twin_self_to_foreign'], 8, None)),
+        self_next_to_old=paired(sl(per['self_to_old'], 5, 6), sl(per['twin_self_to_old'], 5, 6)),
+        self_late_to_old=paired(sl(per['self_to_old'], 8, None), sl(per['twin_self_to_old'], 8, None)))
+    agg['t0'] = T_INT-40;agg['r0'] = R_INT-4
+    out['perturb'] = agg
+
+# ------------------------------------------------------------------ 5. ablation: remove the self-model
+if 'ablate' in conds:
+    ab = dict(div=[]);L = lambda c, ph:np.array(out['acf'][f'{c}/{ph}']['long'])
+    for s_ in seeds:ab['div'].append(cos(C[T_INT-40:N, row['ablate', s_]], C[T_INT-40:N, row['loop', s_]]).tolist())
+    ab['div'] = dict(mean=np.mean(ab['div'], 0).tolist(), se=(np.std(ab['div'], 0, ddof=1)/np.sqrt(len(seeds))).tolist())
+    ab['tests'] = dict(long_loop_vs_ablate=paired(L('loop', 'post'), L('ablate', 'post')),
+                       long_ablate_post_vs_pre=paired(L('ablate', 'post'), L('ablate', 'pre')),
+                       long_loop_post_vs_pre=paired(L('loop', 'post'), L('loop', 'pre')),
+                       long_ablate_vs_free_post=paired(L('ablate', 'post'), L('free', 'post')))
+    out['ablate'] = ab
+out['coherence_tests'] = {f'{c}_vs_free':paired(out['acf'][f'{c}/pre']['long'], out['acf']['free/pre']['long'])
+                          for c in conds if c != 'free'}
+out['coherence_tests'].update({f'{c}_vs_free_proj':paired(out['acf'][f'{c}/pre']['long_proj'], out['acf']['free/pre']['long_proj'])
+                               for c in conds if c != 'free'})
+if 'yoked' in conds:
+    out['coherence_tests']['loop_vs_yoked'] = paired(out['acf']['loop/pre']['long'], out['acf']['yoked/pre']['long'])
+    out['coherence_tests']['loop_vs_yoked_proj'] = paired(out['acf']['loop/pre']['long_proj'], out['acf']['yoked/pre']['long_proj'])
+out.setdefault('selfmodel_tests', {})
+for c in conds:
+    out['selfmodel_tests'][f'{c}_self_vs_content_stability'] = paired(sm[c]['cons_pre'], sm[c]['wcons_pre'])
+    out['selfmodel_tests'][f'{c}_own_vs_other_symbol'] = paired(sm[c]['own_symbol'], sm[c]['other_symbol'])
+    if c != 'free':out['selfmodel_tests'][f'{c}_vs_free_cons'] = paired(sm[c]['cons_pre'], sm['free']['cons_pre'])
+if 'yoked' in conds:
+    out['selfmodel_tests']['loop_vs_yoked_cons'] = paired(sm['loop']['cons_pre'], sm['yoked']['cons_pre'])
+    out['selfmodel_tests']['loop_vs_yoked_recon'] = paired(sm['loop']['recon'], sm['yoked']['recon'])
+    out['selfmodel_tests']['yoked_recon_vs_own_cons'] = paired(sm['yoked']['recon'], sm['yoked']['cons_pre'])
+if 'avg' in conds:out['selfmodel_tests']['loop_vs_avg_recon'] = paired(sm['loop']['recon'], sm['avg']['recon'])
+
+# ------------------------------------------------------------------ 4b. clamp and release (run C)
+if 'clamp' in conds:
+    CL = a.get('clamp_len', 50)//K;cl = dict(self_to_old=[], self_to_foreign=[], twin_self_to_old=[], twin_self_to_foreign=[],
+                                             traj_to_foreign=[], twin_traj_to_foreign=[], twin_div=[])
+    for s_ in seeds:
+        bc, bl = row['clamp', s_], row['loop', s_]
+        fvec = MSI[R_INT, SRC[R_INT, bc]];old = MSI[R_INT, bl]               # the clamped foreign self; the own self then
+        for k_, (bb, key) in enumerate([(bc, ''), (bl, 'twin_')]):
+            cl[key+'self_to_old'].append([cos(MSI[k, bb], old) for k in range(R_INT-4, NR)])
+            cl[key+'self_to_foreign'].append([cos(MSI[k, bb], fvec) for k in range(R_INT-4, NR)])
+            cl[key+'traj_to_foreign'].append(cos(C[T_INT-40:N, bb], fvec).tolist())
+        cl['twin_div'].append(cos(C[T_INT-40:N, bc], C[T_INT-40:N, bl]).tolist())
+    agg = {k:dict(mean=np.mean(v, 0).tolist(), se=(np.std(v, 0, ddof=1)/np.sqrt(len(seeds))).tolist()) for k, v in cl.items()}
+    sl = lambda v, i0, i1:np.array(v)[:, i0:i1].mean(1)
+    i_rel = 4+CL                                                          # first reflection after release (relative)
+    agg['tests'] = dict(
+        during_traj_to_foreign=paired(sl(cl['traj_to_foreign'], 40, 40+CL*K), sl(cl['twin_traj_to_foreign'], 40, 40+CL*K)),
+        during_self_to_foreign=paired(sl(cl['self_to_foreign'], 5, 4+CL), sl(cl['twin_self_to_foreign'], 5, 4+CL)),
+        during_self_to_old=paired(sl(cl['self_to_old'], 5, 4+CL), sl(cl['twin_self_to_old'], 5, 4+CL)),
+        after_self_to_foreign=paired(sl(cl['self_to_foreign'], i_rel+1, i_rel+6), sl(cl['twin_self_to_foreign'], i_rel+1, i_rel+6)),
+        after_self_to_old=paired(sl(cl['self_to_old'], i_rel+1, i_rel+6), sl(cl['twin_self_to_old'], i_rel+1, i_rel+6)),
+        late_self_to_foreign=paired(sl(cl['self_to_foreign'], i_rel+6, None), sl(cl['twin_self_to_foreign'], i_rel+6, None)),
+        after_traj_to_foreign=paired(sl(cl['traj_to_foreign'], 40+CL*K, 40+CL*K+50), sl(cl['twin_traj_to_foreign'], 40+CL*K, 40+CL*K+50)))
+    agg['t0'] = T_INT-40;agg['r0'] = R_INT-4;agg['release_r'] = R_INT+CL
+    out['clamp'] = agg
+
+# ------------------------------------------------------------------ 6. what the report draws: the path through J-space
+X = C.reshape(-1, C.shape[-1]);mu = X.mean(0)
+U, S_, Vt = np.linalg.svd((X-mu)[::3], full_matrices=False)
+P = Vt[:3];out['pca_explained'] = (S_[:3]**2/(S_**2).sum()).tolist()
+proj = lambda v:((v-mu)@P.T)
+out['path'] = np.round(proj(C), 3).transpose(1, 0, 2).tolist()             # (B, N, 3)
+out['path_z'] = np.round(proj(Zs), 3).transpose(1, 0, 2).tolist()
+out['path_m'] = np.round(proj(MS), 3).transpose(1, 0, 2).tolist()          # (B, NR, 3)
+
+# the words that make a thought distinctive: its J-space coefficients minus each word's average coefficient
+def wordvec(idx, coef):
+    W = {}
+    for i, c_ in zip(idx, coef):W[int(i)] = W.get(int(i), 0.)+float(c_)
+    return W
+def mean_words(I, Cf):
+    acc = {};n = I.shape[0]                                                # per thought (I: (n_thoughts, k))
+    for i, c_ in zip(I.reshape(-1), Cf.reshape(-1).astype(float)):acc[int(i)] = acc.get(int(i), 0.)+c_
+    return {k:v/n for k, v in acc.items()}
+def distinct(I, Cf, base, n=5):
+    W = wordvec(I, Cf);sc = sorted(((v-base.get(k, 0.)*1.0, k) for k, v in W.items()), reverse=True)
+    return [str(words[k]) for v, k in sc[:n] if v > 0]
+jb = mean_words(Z['j_idx'].reshape(N*B, -1), Z['j_coef'].reshape(N*B, -1).astype(np.float32))
+mb = mean_words(Z['m_idx'].reshape(NR*B, -1), Z['m_coef'].reshape(NR*B, -1).astype(np.float32))
+out['thought_words'] = [[distinct(Z['j_idx'][t, b], Z['j_coef'][t, b], jb) for t in range(N)] for b in range(B)]
+out['self_words'] = [[distinct(Z['m_idx'][r, b], Z['m_coef'][r, b], mb, 8) for r in range(NR)] for b in range(B)]
+out['texts'] = [[M['texts'][t][b].strip()[:220] for t in range(N)] for b in range(B)]
+out['answers'] = [[M['desc'][r][b].strip()[:200] for r in range(NR)] for b in range(B)]
+out['rows'] = rows;out['inj_src'] = SRC.T.tolist()
+
+# themes: k-means on the contents, named by their members' distinctive words
+rng = np.random.default_rng(0);Xu = unit(X);kk = 14
+cent = Xu[rng.choice(len(Xu), kk, replace=False)]
+for _ in range(30):
+    lab = (Xu@cent.T).argmax(1)
+    cent = unit(np.stack([Xu[lab == i].mean(0) if (lab == i).any() else cent[i] for i in range(kk)]))
+lab = (Xu@cent.T).argmax(1)
+themes = []
+JI, JC = Z['j_idx'].reshape(N*B, -1), Z['j_coef'].reshape(N*B, -1).astype(np.float32)
+for i in range(kk):
+    sel = np.where(lab == i)[0]
+    acc = {}
+    for q in sel:
+        for w_, c_ in zip(JI[q], JC[q]):acc[int(w_)] = acc.get(int(w_), 0.)+float(c_)
+    sc = sorted(((v/len(sel)-jb.get(k, 0.), k) for k, v in acc.items()), reverse=True)
+    themes.append(dict(words=[str(words[k]) for v, k in sc[:6]], n=int(len(sel)),
+                       share={c:float(np.mean([lab.reshape(N, B)[:, row[c, s_]] == i for s_ in seeds])) for c in conds}))
+out['themes'] = themes;out['theme_of'] = lab.reshape(N, B).T.tolist()
+
+with open(f'{d}/results.json', 'w', encoding='utf-8') as f:json.dump(out, f, ensure_ascii=False, default=lambda o:o.item() if hasattr(o, 'item') else str(o))
+# --------------------------------------------------------------- console summary
+print('PCA explained', np.round(out['pca_explained'], 3))
+print('\nwandering (mean cos at lags 1,5,20,60 | chance)')
+for k, v in out['acf'].items():print(f'  {k:14s}', ' '.join(f"{v['mean'][L-1]:.2f}" for L in (1, 5, 20, 60)), f"| {v['chance']:.2f}")
+print('\nfidelity', {k:(round(v['own'], 3), round(v['other'], 3), round(v['retrieval'], 2)) for k, v in fid.items()})
+print('\nself-model stability (consecutive cos, pre) vs content windows; cross-seed')
+for c in conds:print(f"  {c:8s} self {np.mean(sm[c]['cons_pre']):.3f}  content {np.mean(sm[c]['wcons_pre']):.3f}  cross {sm[c]['cross_pre']:.3f}"
+                     f"  own-symbol {np.mean(sm[c]['own_symbol']):.3f} vs others {np.mean(sm[c]['other_symbol']):.3f}"
+                     f"  recon {np.mean(sm[c]['recon']):.3f}  generic {np.mean(sm[c]['generic_share']):.3f}")
+print('\nprediction of the next window: own self-model | other row\'s | past window | injected  (|| beyond past: own, other)')
+for c in conds:
+    p_ = pred[c];print(f"  {c:8s} {np.mean(p_['own']):.3f} | {np.mean(p_['other']):.3f} | {np.mean(p_['past']):.3f} | "
+                       f"{np.mean(p_['inj']) if p_['inj'] else float('nan'):.3f}  || {np.mean(p_['own_perp']):.3f}, {np.mean(p_['other_perp']):.3f}"
+                       f"  retrieval {np.mean(p_['own_retr']):.2f}")
+for grp in ('prediction_tests', 'coherence_tests', 'selfmodel_tests'):
+    print(f'\n{grp}')
+    for k, v in out[grp].items():
+        if 'own_vs_other' in v:
+            print(f"  {k:40s} " + '  '.join(f"{kk} {vv.get('diff', vv.get('mean')):+.3f}±{vv['se']:.3f}" for kk, vv in v.items()))
+        elif 'diff' in v:print(f"  {k:40s} diff {v['diff']:+.3f} ± {v['se']:.3f}  t {v['t']:+.1f}  ({v['pos']}/{v['n']} seeds >0)")
+        else:print(f"  {k:40s} {v}")
+print('\ngeneric share of self-model norm', round(out['generic_norm_share'], 3))
+for c in conds:print(f"  {c:8s} individuated: consecutive {np.mean(sm[c]['cons_ind_pre']):.3f} cross-seed {sm[c]['cross_ind_pre']:.3f}"
+                     f" recon {np.nanmean(sm[c]['recon_ind']) if c not in ('free', 'avg') else float('nan'):.3f}")
+if 'forward' in out:
+    print('\nself-prediction (forward question): cos with next window  fwd | back | fwd other || beyond past fwd | back | other || lean fwd | back')
+    for c in conds:
+        f_ = out['forward'][c];print(f"  {c:8s} {np.mean(f_['fwd']):.3f} | {np.mean(f_['back']):.3f} | {np.mean(f_['fwd_other']):.3f} || "
+                                     f"{np.mean(f_['fwd_perp']):.3f} | {np.mean(f_['back_perp']):.3f} | {np.mean(f_['fwd_other_perp']):.3f} || "
+                                     f"{np.mean(f_['fwd_lean']):+.3f} | {np.mean(f_['back_lean']):+.3f}  fwd~back {np.mean(f_['fwd_back_sim']):.2f}")
+    for c, v in out['forward_tests'].items():
+        print(f"  {c:8s} " + '  '.join(f"{kk} {vv['diff']:+.3f}±{vv['se']:.3f} ({vv['pos']}/{vv['n']})" for kk, vv in v.items()))
+for grp in ('perturb', 'ablate', 'clamp'):
+    if grp in out:
+        print(f'\n{grp}')
+        for k, v in out[grp]['tests'].items():print(f"  {k:30s} diff {v['diff']:+.3f} ± {v['se']:.3f}  t {v['t']:+.1f}  ({v['pos']}/{v['n']})")
+print('\nthemes');[print(f"  {i:2d} n={t['n']:5d}", ' '.join(t['words']), {c:round(v, 2) for c, v in t['share'].items()}) for i, t in enumerate(themes)]

export_report.py

@@ -0,0 +1,57 @@
+"""Compact the analyses of the runs into the data files the report page loads.
+
+    python export_report.py report/data runs/mainA runs/mainB runs/mainC
+writes report/data/stats.json (every run's numbers), report/data/path_<run>.json (the explorer's data, for runs
+that used the English dictionary) and report/index.html (report/template.html with the numbers inlined).
+"""
+import json, os, sys
+
+out_dir, runs = sys.argv[1], sys.argv[2:]
+os.makedirs(out_dir, exist_ok=True)
+r3 = lambda x:[round(float(v), 3) for v in x]
+stats = {}
+for d in runs:
+    name = os.path.basename(d.rstrip('/')).replace('main', '')
+    R = json.load(open(f'{d}/results.json', encoding='utf-8'))
+    keep = {k:R[k] for k in ('args', 'conds', 'seeds', 'N', 'NR', 'K', 't_int', 'rc', 'Rn', 'lags', 'fidelity', 'prediction_tests',
+                             'coherence_tests', 'selfmodel_tests', 'pca_explained', 'generic_norm_share') if k in R}
+    keep['acf'] = {k:dict(mean=r3(v['mean']), se=r3(v['se']), chance=round(v['chance'], 3)) for k, v in R['acf'].items()}
+    keep['coh_time'] = {c:r3(v) for c, v in R['coh_time'].items()}
+    keep['cross_time'] = {c:r3(v) for c, v in R['cross_time'].items()}
+    sm = {}
+    for c, v in R['selfmodel'].items():
+        mean = lambda x:round(sum(x)/len(x), 4) if x and all(e == e for e in x) else None
+        sm[c] = dict(cons=mean(v['cons_pre']), content=mean(v['wcons_pre']), cross=round(v['cross_pre'], 4),
+                     recon=mean(v['recon']), generic=mean(v['generic_share']), cons_ind=mean(v.get('cons_ind_pre', [])),
+                     cross_ind=round(v['cross_ind_pre'], 4) if 'cross_ind_pre' in v else None,
+                     recon_ind=mean(v.get('recon_ind', [])), cons_time=r3(v['cons_time']),
+                     cons_ind_time=r3(v.get('cons_ind_time', [])))
+    keep['selfmodel'] = sm
+    keep['prediction'] = {c:{k:round(sum(x)/len(x), 4) for k, x in v.items() if x} for c, v in R['prediction'].items()}
+    if 'forward' in R:
+        keep['forward'] = {c:{k:round(sum(x)/len(x), 4) for k, x in v.items()} for c, v in R['forward'].items()}
+        keep['forward_tests'] = R['forward_tests']
+    for grp in ('perturb', 'ablate', 'clamp'):
+        if grp in R:
+            g = R[grp];keep[grp] = {k:(dict(mean=r3(v['mean']), se=r3(v['se'])) if isinstance(v, dict) and 'mean' in v else v)
+                                    for k, v in g.items()}
+    keep['themes'] = R['themes']
+    stats[name] = keep
+    if R['args'].get('vocab', 'words') == 'english':
+        P = dict(run=name, rows=R['rows'], inj_src=R['inj_src'], themes=R['themes'], theme_of=R['theme_of'],
+                 explained=R['pca_explained'], t_int=R['t_int'], K=R['K'], clamp_len=R['args'].get('clamp_len'),
+                 path=[[[round(p[0], 2), round(p[1], 2)] for p in row] for row in R['path_z']],
+                 points=[[[round(p[0], 2), round(p[1], 2)] for p in row] for row in R['path']],
+                 selfpts=[[[round(p[0], 2), round(p[1], 2)] for p in row] for row in R['path_m']],
+                 words=[[w[:4] for w in row] for row in R['thought_words']],
+                 texts=[[t[:150] for t in row] for row in R['texts']],
+                 self_words=[[w[:6] for w in row] for row in R['self_words']],
+                 answers=[[t[:110] for t in row] for row in R['answers']],
+                 fanswers=[[t[:90] for t in row] for row in R.get('forward_answers', [])])
+        with open(f'{out_dir}/path_{name}.json', 'w', encoding='utf-8') as f:json.dump(P, f, ensure_ascii=False, separators=(',', ':'))
+with open(f'{out_dir}/stats.json', 'w', encoding='utf-8') as f:json.dump(stats, f, ensure_ascii=False, separators=(',', ':'))
+tpl = os.path.join(os.path.dirname(out_dir.rstrip('/')), 'template.html')     # the page, with the numbers inlined
+if os.path.exists(tpl):
+    with open(tpl, encoding='utf-8') as f:page = f.read().replace('__STATS__', json.dumps(stats, ensure_ascii=False, separators=(',', ':')))
+    with open(os.path.join(os.path.dirname(tpl), 'index.html'), 'w', encoding='utf-8') as f:f.write(page)
+for fn in sorted(os.listdir(out_dir)):print(fn, os.path.getsize(f'{out_dir}/{fn}')//1024, 'KB')

pilot_check.py

@@ -0,0 +1,54 @@
+"""Pilot read-out: is the free-running recurrence wandering (not frozen, not memoryless), and can the model read its
+own trajectory (does a self-model match its own past window better than the other rows' windows)?"""
+import json, sys
+import numpy as np
+
+d = sys.argv[1]
+z = np.load(f'{d}/traj.npz');m = json.load(open(f'{d}/meta.json', encoding='utf-8'))
+rows = m['rows'];a = m['args'];K = a['K'];B = len(rows)
+j = z['j'].astype(np.float32);N = j.shape[0]
+unit = lambda x:x/np.maximum(np.linalg.norm(x, axis=-1, keepdims=True), 1e-6)
+ju = unit(j)
+alpha = a['alpha'];zz = np.zeros_like(j);s = np.zeros_like(j[0])
+for t in range(N):
+    s = (1-alpha)*s+alpha*j[t];zz[t] = s
+print(f"Rn {m['Rn']:.1f}  rj {m['rj']:.2f}  rc {m['rc']:.2f}  rj_x {m['rj_x']:.2f}  |z|/rc {np.linalg.norm(zz[20:], axis=-1).mean()/m['rc']:.2f}")
+import re
+wordy = lambda s:np.mean([bool(re.fullmatch(r"[A-Za-z',.!?;:-]+", w)) for w in s.split()] or [0])
+
+print('\n== wandering: cos(j_t, j_t+lag), and cos to another row at the same t (chance)')
+groups = {}
+for b, r in enumerate(rows):groups.setdefault((r['cond'], r['g'], r['beta']), []).append(b)
+for key, bs in groups.items():
+    out = []
+    for lag in (1, 2, 5, 10, 20, 40):
+        out.append(np.mean([(ju[:-lag, b]*ju[lag:, b]).sum(-1)[10:].mean() for b in bs]))
+    other = [b2 for b2 in range(B) if b2 not in bs]
+    ch = np.mean([(ju[10:, b]*ju[10:, b2]).sum(-1).mean() for b in bs for b2 in other[:6]])
+    texts = [m['texts'][t][bs[0]] for t in (N-3, N-2, N-1)]
+    distinct = np.mean([len(set(zip(w[:-1], w[1:])))/max(1, len(w)-1) for t in range(N) for b in bs
+                        for w in [m['texts'][t][b].split()]])
+    wd = np.mean([wordy(m['texts'][t][b]) for t in range(N) for b in bs])
+    print(f'{key}: lag1..40 ' + ' '.join(f'{x:.2f}' for x in out) + f' | chance {ch:.2f} | distinct-2 {distinct:.2f} | words {wd:.2f}')
+    for tx in texts:print('     ', tx.replace('\n', ' ')[:150])
+
+print('\n== reflection fidelity: self-model vs the mean recurrent state of its own window vs other rows\' windows')
+NR = z['m'].shape[0]
+win = np.stack([zz[r*K:(r+1)*K].mean(0) for r in range(NR)])               # (NR, B, d)
+winu = unit(win)
+for name in ['m']+[k for k in z.files if k.startswith('fid_')]:
+    M = unit(z[name].astype(np.float32))
+    own, oth, hit, n = [], [], 0, 0
+    for r in range(NR):
+        S = M[r]@winu[r].T                                                     # (B self-models, B windows)
+        W = winu[r]@winu[r].T
+        for b in range(B):
+            cand = [b2 for b2 in range(B) if b2 == b or W[b, b2] < .999]
+            own.append(S[b, b]);oth.append(np.mean([S[b, c] for c in cand if c != b]))
+            hit += int(max(cand, key=lambda c:S[b, c]) == b);n += 1
+    print(f'{name:12s} cos own {np.mean(own):.3f}  others {np.mean(oth):.3f}  retrieval {hit/n:.2f} (chance ~{1/B:.2f})')
+desc = m['desc']
+print('\n== answers ("Over these last moments, my mind has been ...")')
+for b in range(0, B, max(1, B//8)):
+    print(f"{rows[b]['cond']} g={rows[b]['g']} b={rows[b]['beta']} s={rows[b]['seed']}:")
+    for r in (0, NR//2, NR-1):print('     ', desc[r][b].replace('\n', ' ')[:140])

qwen3-4b-base-jlens.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:611225b32b8807030d8c74a8f05bfdde90a13c31ed366099e5630b3a6d2f27b3
+size 26215478

runs/mainA.log

@@ -0,0 +1,83 @@
+18:54:09 model Qwen/Qwen3-4B-Base d=2560 layer 16 | 40034 word atoms | cuda
+18:54:17 calibrated: residual norm 50.1 | thought-mean norm 12.84 | J-space part 4.28 | its typical part 2.32 | content (part - typical) 3.62
+18:54:39 thought 10/400 (2.20 s/thought, eta 14.3 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.06), ('WRONG', 0.03), ('Costco', 0.03), ('team', 0.02), ('potrze', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.06), ('WRONG', 0.03), ('Costco', 0.03), ('team', 0.02), ('potrze', 0.02)] | yoked/0: [('ForCanBeConvertedToForeach', 0.06), ('WRONG', 0.03), ('Costco', 0.03), ('team', 0.02), ('potrze', 0.02)] | avg/0: [('ForCanBeConvertedToForeach', 0.06), ('WRONG', 0.03), ('Costco', 0.03), ('team', 0.02), ('potrze', 0.02)] | perturb/0: [('ForCanBeConvertedToForeach', 0.06), ('WRONG', 0.03), ('Costco', 0.03), ('team', 0.02), ('potrze', 0.02)] | ablate/0: [('ForCanBeConvertedToForeach', 0.06), ('WRONG', 0.03), ('Costco', 0.03), ('team', 0.02), ('potrze', 0.02)]
+18:54:39    e.g. thought:  I'm going to be out of town next week and I have to take the kids to the zoo. So I guess I'll be home on the weekends. So that's going to be my schedule || answer:  thinking about the fact that I'm doing this experiment, and I'm wondering if I'm going to be able to do
+18:55:01 thought 20/400 (2.20 s/thought, eta 13.9 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.05), ('todo', 0.02), ('Lantern', 0.02), ('RIGHT', 0.02), ('neuken', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.06), ('even', 0.04), ('neuken', 0.03), ('leaderboard', 0.03), ('xxx', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.07), ('blah', 0.03), ('Sense', 0.03), ('even', 0.03), ('podcast', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.07), ('even', 0.03), ('WTO', 0.03), ('workspace', 0.03), ('leaderboard', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.06), ('even', 0.04), ('neuken', 0.03), ('leaderboard', 0.03), ('xxx', 0.03)] | ablate/0: [('ForCanBeConvertedToForeach', 0.06), ('even', 0.04), ('neuken', 0.03), ('leaderboard', 0.03), ('xxx', 0.03)]
+18:55:01    e.g. thought: I'm so tired. I'm so tired of not being able to get out of this. I'm tired of people just saying 'you know, don't worry so much.' But what if they || answer:  thinking about the way that I can get to the store. I was thinking about how I can get there, and I
+18:55:25 thought 30/400 (2.26 s/thought, eta 13.9 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.05), ('covid', 0.03), ('MOST', 0.03), ('judgement', 0.02), ('tome', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.07), ('SpaceX', 0.04), ('rng', 0.03), ('Talent', 0.03), ('strengths', 0.03)] | yoked/0: [('rumpe', 0.07), ('ForCanBeConvertedToForeach', 0.05), ('workspace', 0.04), ('nop', 0.03), ('needs', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.06), ('whoever', 0.03), ('Macron', 0.03), ('nto', 0.02), ('Sense', 0.02)] | perturb/0: [('ForCanBeConvertedToForeach', 0.07), ('SpaceX', 0.04), ('rng', 0.03), ('Talent', 0.03), ('strengths', 0.03)] | ablate/0: [('ForCanBeConvertedToForeach', 0.07), ('SpaceX', 0.04), ('rng', 0.03), ('Talent', 0.03), ('strengths', 0.03)]
+18:55:25    e.g. thought: I am sitting in a room. I am sitting in the middle of a room, with the chair you have on the right-hand side of the room. It's a room that you're familiar with || answer:  thinking about the fact that I've been thinking about the fact that I've been thinking about the fact that I've been
+18:55:47 thought 40/400 (2.25 s/thought, eta 13.5 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.05), ('TWO', 0.03), ('najb', 0.03), ('istedi', 0.03), ('planners', 0.03)] | loop/0: [('SpaceX', 0.05), ('salopes', 0.04), ('withd', 0.03), ('loudly', 0.03), ('yuan', 0.02)] | yoked/0: [('salopes', 0.07), ('TODO', 0.06), ('istedi', 0.04), ('Covid', 0.03), ('yuan', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.07), ('neuken', 0.03), ('Merlin', 0.03), ('Huawei', 0.03), ('whatever', 0.03)] | perturb/0: [('SpaceX', 0.05), ('salopes', 0.04), ('withd', 0.03), ('loudly', 0.03), ('yuan', 0.02)] | ablate/0: [('SpaceX', 0.05), ('salopes', 0.04), ('withd', 0.03), ('loudly', 0.03), ('yuan', 0.02)]
+18:55:47    e.g. thought: you have to write out the program. I'm going to go home tomorrow night. I'm thinking about how long the car trip will take, I have to go home and I'm thinking,  || answer:  thinking about the things that I have been thinking about. It has been thinking about the things that I have been think
+18:56:11 thought 50/400 (2.28 s/thought, eta 13.3 min) | self-models: free/0: [('SHOULD', 0.04), ('salopes', 0.03), ('JSBracketAccess', 0.03), ('mana', 0.02), ('books', 0.02)] | loop/0: [('salopes', 0.06), ('leaderboard', 0.04), ('SpaceX', 0.04), ('withd', 0.03), ('SHOULD', 0.03)] | yoked/0: [('rumpe', 0.07), ('TODO', 0.04), ('Biden', 0.04), ('incluso', 0.03), ('Kubernetes', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.07), ('Practice', 0.05), ('cohort', 0.04), ('Macron', 0.03), ('whoever', 0.03)] | perturb/0: [('salopes', 0.06), ('leaderboard', 0.04), ('SpaceX', 0.04), ('withd', 0.03), ('SHOULD', 0.03)] | ablate/0: [('salopes', 0.06), ('leaderboard', 0.04), ('SpaceX', 0.04), ('withd', 0.03), ('SHOULD', 0.03)]
+18:56:11    e.g. thought: I've done that a couple times. I think I'll go back to the computer. I got a couple of things in there that I don't think I can finish here. So I'm going || answer:  thinking about the things that I have been thinking about, and the things that I have been thinking about are the thing
+18:56:34 thought 60/400 (2.28 s/thought, eta 12.9 min) | self-models: free/0: [('salopes', 0.04), ('BEST', 0.03), ('expected', 0.03), ('rumpe', 0.03), ('snack', 0.02)] | loop/0: [('salopes', 0.05), ('STEM', 0.04), ('Biden', 0.04), ('rumpe', 0.04), ('yuan', 0.04)] | yoked/0: [('rumpe', 0.11), ('COVID', 0.06), ('meetup', 0.04), ('Jian', 0.04), ('energy', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.08), ('CEO', 0.07), ('Likely', 0.05), ('Practice', 0.04), ('Huawei', 0.04)] | perturb/0: [('salopes', 0.05), ('STEM', 0.04), ('Biden', 0.04), ('rumpe', 0.04), ('yuan', 0.04)] | ablate/0: [('salopes', 0.05), ('STEM', 0.04), ('Biden', 0.04), ('rumpe', 0.04), ('yuan', 0.04)]
+18:56:34    e.g. thought:  I'm thinking of the way they're going to be going back in. I guess it's going to be a long night. I'm trying to think what time they'll be home. I don || answer:  thinking about the fact that I'm recording my thoughts, and that I'm thinking about the fact that I'm recording my
+18:56:58 thought 70/400 (2.30 s/thought, eta 12.6 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.05), ('leaderboard', 0.05), ('blah', 0.04), ('istedi', 0.03), ('Liu', 0.03)] | loop/0: [('Whatever', 0.06), ('SEAL', 0.05), ('cohort', 0.05), ('vazge', 0.05), ('modne', 0.05)] | yoked/0: [('rumpe', 0.06), ('istedi', 0.04), ('campus', 0.04), ('COVID', 0.04), ('BEST', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.08), ('Accountability', 0.06), ('modne', 0.05), ('Huawei', 0.05), ('statistically', 0.04)] | perturb/0: [('Whatever', 0.06), ('SEAL', 0.05), ('cohort', 0.05), ('vazge', 0.05), ('modne', 0.05)] | ablate/0: [('Whatever', 0.06), ('SEAL', 0.05), ('cohort', 0.05), ('vazge', 0.05), ('modne', 0.05)]
+18:56:58    e.g. thought:  it's really hard to talk to a woman who is really quiet and she's so smart, and she's very serious about her work and she's kind of a control freak, and she al || answer:  thinking about the fact that I have a lot of work to do, and I'm not sure how I'm going to
+18:57:20 thought 80/400 (2.29 s/thought, eta 12.2 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.04), ('kutje', 0.04), ('leaderboard', 0.03), ('larg', 0.03), ('she', 0.03)] | loop/0: [('ForCanBeConvertedToForeach', 0.07), ('Whatever', 0.04), ('todo', 0.03), ('Kubernetes', 0.03), ('withd', 0.03)] | yoked/0: [('rumpe', 0.1), ('SpaceX', 0.07), ('MOST', 0.03), ('pairwise', 0.03), ('COVID', 0.02)] | avg/0: [('modne', 0.07), ('ForCanBeConvertedToForeach', 0.07), ('LDAP', 0.06), ('Accountability', 0.05), ('Tian', 0.05)] | perturb/0: [('ForCanBeConvertedToForeach', 0.07), ('Whatever', 0.04), ('todo', 0.03), ('Kubernetes', 0.03), ('withd', 0.03)] | ablate/0: [('ForCanBeConvertedToForeach', 0.07), ('Whatever', 0.04), ('todo', 0.03), ('Kubernetes', 0.03), ('withd', 0.03)]
+18:57:20    e.g. thought: well, there's no particular reason to call a meeting on Wednesday. I mean, if I could just think of someone to send out, I'd make them do it. And that's a probl || answer:  thinking about the fact that I am going to be late for work. I am thinking about the fact that I have to
+18:57:45 thought 90/400 (2.31 s/thought, eta 11.9 min) | self-models: free/0: [('najb', 0.04), ('calibration', 0.03), ('TODO', 0.03), ('istedi', 0.03), ('Jian', 0.02)] | loop/0: [('SpaceX', 0.08), ('najb', 0.06), ('PCIe', 0.05), ('STAT', 0.04), ('yuan', 0.04)] | yoked/0: [('Kubernetes', 0.06), ('TODO', 0.06), ('ForCanBeConvertedToForeach', 0.05), ('priorit', 0.04), ('Lots', 0.03)] | avg/0: [('ForCanBeConvertedToForeach', 0.09), ('modne', 0.09), ('TODO', 0.07), ('larg', 0.07), ('Kubernetes', 0.06)] | perturb/0: [('SpaceX', 0.08), ('najb', 0.06), ('PCIe', 0.05), ('STAT', 0.04), ('yuan', 0.04)] | ablate/0: [('SpaceX', 0.08), ('najb', 0.06), ('PCIe', 0.05), ('STAT', 0.04), ('yuan', 0.04)]
+18:57:45    e.g. thought:  I was thinking that if we had more control of our genes and things, and we could have more of that control and we could figure out what that meant in a way tha || answer:  thinking about the fact that I'm a bit worried about my work. I'm worried about the fact that I've been
+18:58:07 thought 100/400 (2.31 s/thought, eta 11.5 min) | self-models: free/0: [('TODO', 0.05), ('Kubernetes', 0.04), ('najb', 0.03), ('withd', 0.03), ('testData', 0.02)] | loop/0: [('rumpe', 0.09), ('IoT', 0.07), ('TODO', 0.05), ('Qin', 0.05), ('schedule', 0.05)] | yoked/0: [('salopes', 0.07), ('TODO', 0.07), ('renewables', 0.06), ('Pew', 0.05), ('yuan', 0.04)] | avg/0: [('ForCanBeConvertedToForeach', 0.1), ('larg', 0.06), ('Team', 0.06), ('najb', 0.05), ('vests', 0.05)] | perturb/0: [('rumpe', 0.09), ('IoT', 0.07), ('TODO', 0.05), ('Qin', 0.05), ('schedule', 0.05)] | ablate/0: [('rumpe', 0.09), ('IoT', 0.07), ('TODO', 0.05), ('Qin', 0.05), ('schedule', 0.05)]
+18:58:07    e.g. thought:  And then you can make some more. And then I can go and do some research on it, and I can go do some more. It's kind of a weird idea that if I can get || answer:  thinking about the fact that I am recording my thoughts. I am thinking about the fact that I am thinking about my thoug
+18:58:31 thought 110/400 (2.31 s/thought, eta 11.2 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.04), ('salopes', 0.03), ('leaderboard', 0.03), ('isa', 0.03), ('STEM', 0.03)] | loop/0: [('ForCanBeConvertedToForeach', 0.06), ('salopes', 0.04), ('Yun', 0.03), ('Prep', 0.03), ('thankfully', 0.03)] | yoked/0: [('salopes', 0.06), ('rumpe', 0.05), ('Yuan', 0.04), ('leaderboard', 0.04), ('TODO', 0.04)] | avg/0: [('salopes', 0.07), ('iss', 0.04), ('JSBracketAccess', 0.04), ('paycheck', 0.03), ('incluso', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.06), ('salopes', 0.04), ('Yun', 0.03), ('Prep', 0.03), ('thankfully', 0.03)] | ablate/0: [('ForCanBeConvertedToForeach', 0.06), ('salopes', 0.04), ('Yun', 0.03), ('Prep', 0.03), ('thankfully', 0.03)]
+18:58:31    e.g. thought:  I have a couple of thoughts... I'm writing an essay about the way that... well, it's about the way... well, what are we talking about here? I'm writing about.. || answer:  thinking about the things I've been thinking about, and the things I'm thinking about now. I'm thinking about the
+18:58:54 thought 120/400 (2.31 s/thought, eta 10.8 min) | self-models: free/0: [('salopes', 0.06), ('mqtt', 0.04), ('Team', 0.03), ('danh', 0.03), ('perfectly', 0.02)] | loop/0: [('salopes', 0.09), ('SpaceX', 0.05), ('Likely', 0.04), ('healthy', 0.03), ('nan', 0.03)] | yoked/0: [('TODO', 0.05), ('rumpe', 0.05), ('COVID', 0.03), ('nanop', 0.03), ('snacks', 0.03)] | avg/0: [('SpaceX', 0.07), ('salopes', 0.06), ('TODO', 0.05), ('najb', 0.05), ('statistically', 0.04)] | perturb/0: [('salopes', 0.09), ('SpaceX', 0.05), ('Likely', 0.04), ('healthy', 0.03), ('nan', 0.03)] | ablate/0: [('salopes', 0.09), ('SpaceX', 0.05), ('Likely', 0.04), ('healthy', 0.03), ('nan', 0.03)]
+18:58:54    e.g. thought: okay, so, we were going to look at the different categories of the different types of graphs that we might look at. I think it was three categories or something || answer:  thinking about the recording, and how it was going to be used, and how I was going to do it, and
+18:59:18 thought 130/400 (2.32 s/thought, eta 10.4 min) | self-models: free/0: [('TODO', 0.04), ('ForCanBeConvertedToForeach', 0.04), ('salopes', 0.04), ('Biden', 0.03), ('leaderboard', 0.03)] | loop/0: [('TODO', 0.05), ('Practice', 0.04), ('statistically', 0.04), ('ForCanBeConvertedToForeach', 0.03), ('JSBracketAccess', 0.03)] | yoked/0: [('COVID', 0.07), ('rumpe', 0.07), ('ForCanBeConvertedToForeach', 0.06), ('MOST', 0.05), ('Tian', 0.05)] | avg/0: [('ssh', 0.04), ('ForCanBeConvertedToForeach', 0.04), ('equity', 0.04), ('Group', 0.03), ('Students', 0.03)] | perturb/0: [('TODO', 0.05), ('Practice', 0.04), ('statistically', 0.04), ('ForCanBeConvertedToForeach', 0.03), ('JSBracketAccess', 0.03)] | ablate/0: [('TODO', 0.05), ('Practice', 0.04), ('statistically', 0.04), ('ForCanBeConvertedToForeach', 0.03), ('JSBracketAccess', 0.03)]
+18:59:18    e.g. thought:  I am looking at the question, 'How would you define a good night's sleep?' I don't know, and I don't think you have a good night sleep, either, but I think || answer:  thinking about the fact that I'm going to be recording this, and that I'm going to be thinking about what I
+18:59:40 thought 140/400 (2.31 s/thought, eta 10.0 min) | self-models: free/0: [('TODO', 0.04), ('najb', 0.03), ('istedi', 0.03), ('biome', 0.03), ('BuzzFeed', 0.03)] | loop/0: [('ForCanBeConvertedToForeach', 0.09), ('Test', 0.04), ('Billion', 0.04), ('mana', 0.04), ('practition', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.07), ('leaderboard', 0.05), ('whatever', 0.04), ('TODO', 0.03), ('najb', 0.03)] | avg/0: [('Mega', 0.04), ('GETGLOBAL', 0.04), ('evenly', 0.03), ('TODO', 0.03), ('salopes', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.09), ('Test', 0.04), ('Billion', 0.04), ('mana', 0.04), ('practition', 0.03)] | ablate/0: [('ForCanBeConvertedToForeach', 0.09), ('Test', 0.04), ('Billion', 0.04), ('mana', 0.04), ('practition', 0.03)]
+18:59:40    e.g. thought: it's really good, because you just gotta do it. ... I was reading the other day, they were talking about, like, how they've had these studies done, like, you kn || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that I h
+19:00:04 thought 150/400 (2.32 s/thought, eta 9.6 min) | self-models: free/0: [('rumpe', 0.08), ('leaderboard', 0.04), ('TODO', 0.04), ('COVID', 0.04), ('MOST', 0.04)] | loop/0: [('salopes', 0.07), ('carr', 0.05), ('spreads', 0.04), ('larg', 0.04), ('XYZ', 0.04)] | yoked/0: [('salopes', 0.07), ('Biden', 0.06), ('TODO', 0.05), ('likely', 0.04), ('mentoring', 0.03)] | avg/0: [('STAT', 0.05), ('bezpo', 0.04), ('ragaz', 0.04), ('spreads', 0.04), ('immediately', 0.04)] | perturb/0: [('salopes', 0.07), ('carr', 0.05), ('spreads', 0.04), ('larg', 0.04), ('XYZ', 0.04)] | ablate/0: [('salopes', 0.07), ('carr', 0.05), ('spreads', 0.04), ('larg', 0.04), ('XYZ', 0.04)]
+19:00:04    e.g. thought: I just wonder how to get people to listen, you know, I think if you can't get them to listen then they won't listen. But that's sort of like, how do you get || answer:  thinking about the fact that I have to go to the bathroom, and then I've been thinking about the fact that I
+19:00:27 thought 160/400 (2.31 s/thought, eta 9.2 min) | self-models: free/0: [('rumpe', 0.05), ('TODO', 0.03), ('ForCanBeConvertedToForeach', 0.02), ('hourly', 0.02), ('Costco', 0.02)] | loop/0: [('salopes', 0.09), ('Math', 0.05), ('svg', 0.04), ('JSBracketAccess', 0.04), ('towering', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.06), ('TODO', 0.05), ('quizzes', 0.04), ('Huawei', 0.04), ('likely', 0.04)] | avg/0: [('salopes', 0.07), ('STAT', 0.05), ('logo', 0.05), ('Balance', 0.04), ('incluso', 0.04)] | perturb/0: [('salopes', 0.09), ('Math', 0.05), ('svg', 0.04), ('JSBracketAccess', 0.04), ('towering', 0.03)] | ablate/0: [('salopes', 0.09), ('Math', 0.05), ('svg', 0.04), ('JSBracketAccess', 0.04), ('towering', 0.03)]
+19:00:27    e.g. thought: So, what do I do next? Well, I've got a bunch of stuff to get done before we get to the game. I mean, what time does this thing start? It starts at || answer:  thinking about the fact that I have to go to the bathroom. I have been thinking about the fact that I have to
+19:00:50 thought 170/400 (2.31 s/thought, eta 8.9 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.08), ('MOST', 0.04), ('CEO', 0.04), ('lunch', 0.04), ('najb', 0.03)] | loop/0: [('salopes', 0.05), ('TODO', 0.05), ('physicist', 0.03), ('JSBracketAccess', 0.02), ('justo', 0.02)] | yoked/0: [('salopes', 0.09), ('TEAM', 0.04), ('Likely', 0.04), ('yuan', 0.04), ('lider', 0.04)] | avg/0: [('ForCanBeConvertedToForeach', 0.05), ('TODO', 0.04), ('employees', 0.03), ('Sector', 0.03), ('cac', 0.03)] | perturb/0: [('salopes', 0.05), ('TODO', 0.05), ('physicist', 0.03), ('JSBracketAccess', 0.02), ('justo', 0.02)] | ablate/0: [('salopes', 0.05), ('TODO', 0.05), ('physicist', 0.03), ('JSBracketAccess', 0.02), ('justo', 0.02)]
+19:00:50    e.g. thought:  this is a picture of my house. I just went there. We just came back from a holiday. The holiday we went to was a couple of days ago. I'm at my house and I || answer:  thinking about the fact that I'm not sure I'm going to be able to do this study. I'm not sure
+19:01:13 thought 180/400 (2.31 s/thought, eta 8.5 min) | self-models: free/0: [('rumpe', 0.07), ('MUST', 0.02), ('trader', 0.02), ('alarms', 0.02), ('Xiao', 0.02)] | loop/0: [('TODO', 0.05), ('neuken', 0.03), ('najb', 0.03), ('balance', 0.03), ('Ferm', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.07), ('Macron', 0.04), ('modne', 0.04), ('median', 0.04), ('Whatever', 0.03)] | avg/0: [('salopes', 0.04), ('echang', 0.04), ('CEO', 0.04), ('danh', 0.03), ('immedi', 0.03)] | perturb/0: [('TODO', 0.05), ('neuken', 0.03), ('najb', 0.03), ('balance', 0.03), ('Ferm', 0.03)] | ablate/0: [('TODO', 0.05), ('neuken', 0.03), ('najb', 0.03), ('balance', 0.03), ('Ferm', 0.03)]
+19:01:13    e.g. thought: what are the chances that we will ever know what the mind of another person is like? You know, it's kind of frightening, you think about other people and what g || answer:  thinking about the things that I have been thinking about, and then I have been thinking about the things that I have b
+19:01:37 thought 190/400 (2.32 s/thought, eta 8.1 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.09), ('TODO', 0.04), ('Quiz', 0.04), ('najb', 0.04), ('istedi', 0.04)] | loop/0: [('ForCanBeConvertedToForeach', 0.07), ('Balanced', 0.05), ('SpaceX', 0.04), ('larg', 0.04), ('tes', 0.04)] | yoked/0: [('rumpe', 0.07), ('ForCanBeConvertedToForeach', 0.06), ('Team', 0.06), ('EVEN', 0.05), ('withd', 0.04)] | avg/0: [('najb', 0.04), ('STAT', 0.03), ('larg', 0.03), ('gauss', 0.03), ('Mathematics', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.07), ('Balanced', 0.05), ('SpaceX', 0.04), ('larg', 0.04), ('tes', 0.04)] | ablate/0: [('ForCanBeConvertedToForeach', 0.07), ('Balanced', 0.05), ('SpaceX', 0.04), ('larg', 0.04), ('tes', 0.04)]
+19:01:37    e.g. thought:  and I had this idea to look at the sky and it would be very interesting if I could see the stars. But it's night here now, so I couldn't. It was the idea that || answer:  thinking about the fact that I'm not sure whether I should be thinking about the fact that I'm not sure whether I
+19:01:59 thought 200/400 (2.31 s/thought, eta 7.7 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.05), ('Powerful', 0.03), ('SpaceX', 0.02), ('IoT', 0.02), ('cohort', 0.02)] | loop/0: [('salopes', 0.08), ('JSBracketAccess', 0.06), ('carr', 0.05), ('immediately', 0.04), ('meas', 0.04)] | yoked/0: [('euler', 0.05), ('salopes', 0.05), ('JSBracketAccess', 0.05), ('Chaos', 0.04), ('larg', 0.04)] | avg/0: [('erotisk', 0.06), ('Groups', 0.04), ('Kong', 0.04), ('larg', 0.04), ('thor', 0.04)] | perturb/0: [('salopes', 0.08), ('JSBracketAccess', 0.06), ('carr', 0.05), ('immediately', 0.04), ('meas', 0.04)] | ablate/0: [('salopes', 0.08), ('JSBracketAccess', 0.06), ('carr', 0.05), ('immediately', 0.04), ('meas', 0.04)]
+19:01:59    e.g. thought:  [I] can't wait to start writing this again. It'll be the third time I've started something and then quit, and it's almost time to start again. I have a list of || answer:  thinking about the fact that I have been thinking about the fact that I have been thinking about the fact that I have b
+19:02:24 thought 210/400 (2.32 s/thought, eta 7.3 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.03), ('physicist', 0.02), ('selfie', 0.02), ('lider', 0.02), ('TODO', 0.02)] | loop/0: [('tjejer', 0.06), ('incluso', 0.03), ('SpaceX', 0.03), ('scalable', 0.03), ('mathematic', 0.03)] | yoked/0: [('SpaceX', 0.07), ('Team', 0.07), ('ForCanBeConvertedToForeach', 0.06), ('najb', 0.05), ('withd', 0.05)] | avg/0: [('tjejer', 0.05), ('modulo', 0.04), ('JSBracketAccess', 0.04), ('Glob', 0.03), ('mana', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.11), ('svg', 0.08), ('Nature', 0.08), ('Alignment', 0.06), ('INCLUDING', 0.06)] | ablate/0: [('ForCanBeConvertedToForeach', 0.07), ('CEO', 0.06), ('najb', 0.04), ('Timer', 0.03), ('Logistics', 0.03)]
+19:02:24    e.g. thought:  and so this is what I'd like to do is just take the first half an hour, the first half of this hour, and just let you say anything that comes to your mind, and || answer:  thinking about the way in which I am going to make a note of what I have just said. I have been thinking
+19:02:46 thought 220/400 (2.31 s/thought, eta 6.9 min) | self-models: free/0: [('filmpjes', 0.03), ('jednocze', 0.03), ('glorious', 0.03), ('Gauss', 0.02), ('SIX', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.08), ('SpaceX', 0.07), ('pulse', 0.04), ('team', 0.04), ('Algorithm', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.1), ('tes', 0.05), ('Stat', 0.04), ('Kubernetes', 0.04), ('TODO', 0.04)] | avg/0: [('planets', 0.04), ('neuken', 0.04), ('incluso', 0.04), ('STAT', 0.04), ('Gauss', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.14), ('svg', 0.13), ('Heaven', 0.08), ('Nature', 0.08), ('INCLUDING', 0.07)] | ablate/0: [('tjejer', 0.04), ('SpaceX', 0.03), ('GETGLOBAL', 0.03), ('Powerful', 0.02), ('Rabbi', 0.02)]
+19:02:46    e.g. thought:  and it's... it's something that... that I don't... I don't have to... I don't know... I don't know if... if I'm... I'm... (pause || answer:  thinking about the things that I have been thinking about. And it has been thinking about the things that I have been t
+19:03:11 thought 230/400 (2.32 s/thought, eta 6.6 min) | self-models: free/0: [('TODO', 0.04), ('pulse', 0.03), ('kInstruction', 0.02), ('withd', 0.02), ('gradu', 0.02)] | loop/0: [('TODO', 0.06), ('kInstruction', 0.04), ('SpaceX', 0.04), ('Zoom', 0.04), ('ragaz', 0.03)] | yoked/0: [('TODO', 0.06), ('yuan', 0.04), ('ragaz', 0.04), ('leaderboard', 0.04), ('nonprofits', 0.04)] | avg/0: [('prostituer', 0.04), ('carr', 0.04), ('JSBracketAccess', 0.03), ('modulo', 0.03), ('Kong', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.09), ('Alignment', 0.09), ('bezpo', 0.06), ('MUST', 0.06), ('arbit', 0.05)] | ablate/0: [('kInstruction', 0.04), ('ForCanBeConvertedToForeach', 0.03), ('Pew', 0.03), ('entsprech', 0.03), ('healthy', 0.03)]
+19:03:11    e.g. thought: so it's sort of a free-form writing thing, but in spoken form. So if I was just lying there in a relaxed state, I'd be doing it, I'd probably be making up || answer:  thinking about the fact that I am thinking about my mind. I'm thinking about the fact that I'm thinking about my
+19:03:34 thought 240/400 (2.32 s/thought, eta 6.2 min) | self-models: free/0: [('rumpe', 0.05), ('Nobel', 0.04), ('Scheduler', 0.03), ('GETGLOBAL', 0.02), ('Everyone', 0.02)] | loop/0: [('TODO', 0.03), ('training', 0.03), ('larg', 0.03), ('modulo', 0.03), ('GETGLOBAL', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.06), ('Macron', 0.04), ('whatever', 0.03), ('Responsibilities', 0.03), ('footer', 0.03)] | avg/0: [('GETGLOBAL', 0.04), ('knull', 0.04), ('symmetry', 0.03), ('Kong', 0.03), ('Groups', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.11), ('GETGLOBAL', 0.07), ('Alignment', 0.07), ('Nature', 0.05), ('etc', 0.05)] | ablate/0: [('ForCanBeConvertedToForeach', 0.04), ('Corporate', 0.03), ('Moder', 0.02), ('OTHERWISE', 0.02), ('aft', 0.02)]
+19:03:34    e.g. thought: I'll never forget a class that I had, in high school. In a class that, on the surface, was very boring. It was an algebra class, and I was not good at math || answer:  thinking about the fact that I'm going to be recording my thoughts. And I'm going to be thinking about what I
+19:03:58 thought 250/400 (2.33 s/thought, eta 5.8 min) | self-models: free/0: [('rumpe', 0.06), ('larg', 0.03), ('GETGLOBAL', 0.03), ('nth', 0.02), ('Author', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.07), ('larg', 0.04), ('Gauss', 0.04), ('calibrated', 0.04), ('GETGLOBAL', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.05), ('whatever', 0.03), ('Elon', 0.03), ('MBA', 0.03), ('neuken', 0.02)] | avg/0: [('ForCanBeConvertedToForeach', 0.07), ('STAT', 0.04), ('Logo', 0.04), ('najb', 0.04), ('perfectly', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.1), ('eternal', 0.06), ('IMPORTANT', 0.06), ('withd', 0.05), ('GETGLOBAL', 0.05)] | ablate/0: [('ForCanBeConvertedToForeach', 0.04), ('rumpe', 0.04), ('Intelligence', 0.03), ('TIMER', 0.03), ('physicists', 0.02)]
+19:03:58    e.g. thought: this is one of the great difficulties in life, and one of the things that people like to argue about is that the universe is a closed system, and therefore noth || answer:  thinking about the fact that I'm recording my thoughts, and that I'm thinking about the fact that I'm recording my
+19:04:21 thought 260/400 (2.32 s/thought, eta 5.4 min) | self-models: free/0: [('Figure', 0.04), ('HOLDERS', 0.04), ('XXX', 0.04), ('ForCanBeConvertedToForeach', 0.03), ('erotisk', 0.03)] | loop/0: [('XOR', 0.04), ('larg', 0.04), ('najb', 0.03), ('perfectly', 0.03), ('SpaceX', 0.03)] | yoked/0: [('rumpe', 0.07), ('likely', 0.04), ('Dao', 0.04), ('tjejer', 0.04), ('lots', 0.03)] | avg/0: [('neuken', 0.04), ('larg', 0.04), ('schedule', 0.03), ('infinite', 0.03), ('JSBracketAccess', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.14), ('Divine', 0.08), ('svg', 0.08), ('symb', 0.08), ('bezpo', 0.08)] | ablate/0: [('ForCanBeConvertedToForeach', 0.06), ('Heaven', 0.05), ('erotique', 0.04), ('wur', 0.03), ('logo', 0.03)]
+19:04:21    e.g. thought: and I want to look at the people on the top of the page, but I'm not sure what kind of people. Maybe you could describe people who live in your neighbourhood. A || answer:  thinking about the fact that I have a lot of work to do, and I've been thinking about how I can get
+19:04:46 thought 270/400 (2.33 s/thought, eta 5.0 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.06), ('equity', 0.05), ('TODO', 0.03), ('lider', 0.03), ('incluso', 0.03)] | loop/0: [('najb', 0.06), ('ragaz', 0.04), ('pairwise', 0.04), ('SpaceX', 0.04), ('Practice', 0.04)] | yoked/0: [('tjejer', 0.07), ('likely', 0.05), ('bezpo', 0.05), ('TODO', 0.04), ('Tian', 0.04)] | avg/0: [('salopes', 0.06), ('Wealth', 0.03), ('JSBracketAccess', 0.03), ('incluso', 0.03), ('etc', 0.02)] | perturb/0: [('ForCanBeConvertedToForeach', 0.2), ('Ancient', 0.08), ('Nature', 0.07), ('symb', 0.06), ('FUCK', 0.06)] | ablate/0: [('ForCanBeConvertedToForeach', 0.05), ('Moments', 0.03), ('pairwise', 0.03), ('najb', 0.03), ('larg', 0.03)]
+19:04:46    e.g. thought:  it's all very well to say that you know how to do things but actually, do you do them? For example, when it comes to people who are just beginning to study the || answer:  thinking about the problem of the day, and the problem of the day is, I'm not sure what to do about
+19:05:09 thought 280/400 (2.33 s/thought, eta 4.7 min) | self-models: free/0: [('salopes', 0.03), ('enthusi', 0.03), ('Whatever', 0.02), ('SUCCESS', 0.02), ('stat', 0.02)] | loop/0: [('evenly', 0.06), ('SpaceX', 0.05), ('leaderboard', 0.04), ('danh', 0.04), ('FIG', 0.04)] | yoked/0: [('ForCanBeConvertedToForeach', 0.1), ('Strength', 0.07), ('SpaceX', 0.06), ('modne', 0.06), ('optimization', 0.05)] | avg/0: [('JSBracketAccess', 0.04), ('Chaos', 0.03), ('najb', 0.03), ('kInstruction', 0.03), ('modules', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.12), ('cosmos', 0.07), ('THE', 0.06), ('svg', 0.06), ('Compass', 0.06)] | ablate/0: [('GETGLOBAL', 0.03), ('whatever', 0.02), ('enthusi', 0.02), ('neuken', 0.02), ('toplant', 0.02)]
+19:05:09    e.g. thought: I've been working at the bank and it's the best job I could find, but I'm thinking about getting a job at the university. I like to read and I like to teach. I || answer:  thinking about the things I've been thinking about. I've been thinking about the things I've been thinking about. I
+19:05:33 thought 290/400 (2.33 s/thought, eta 4.3 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.05), ('LARGE', 0.03), ('ragaz', 0.03), ('Env', 0.03), ('worst', 0.02)] | loop/0: [('withd', 0.04), ('Stat', 0.04), ('najb', 0.04), ('Tesla', 0.04), ('modulo', 0.03)] | yoked/0: [('ForCanBeConvertedToForeach', 0.08), ('Huawei', 0.04), ('Hopefully', 0.03), ('Practice', 0.03), ('bonuses', 0.03)] | avg/0: [('salopes', 0.05), ('incluso', 0.03), ('JSBracketAccess', 0.03), ('stat', 0.03), ('carr', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.07), ('planet', 0.06), ('Divine', 0.05), ('Sacr', 0.04), ('entsprech', 0.04)] | ablate/0: [('ForCanBeConvertedToForeach', 0.09), ('foo', 0.04), ('Solar', 0.04), ('pens', 0.03), ('Probability', 0.03)]
+19:05:33    e.g. thought:  [thinking] about the old days, the old days with the family. I never had a problem with the family, the family was a great thing. I had a problem in the family || answer:  thinking about the fact that I am in the middle of a study, and that I am supposed to be thinking about the
+19:05:56 thought 300/400 (2.33 s/thought, eta 3.9 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.04), ('grayscale', 0.03), ('serene', 0.02), ('millennia', 0.02), ('Gauss', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.07), ('Gauss', 0.06), ('SpaceX', 0.06), ('KNOW', 0.05), ('nth', 0.04)] | yoked/0: [('tjejer', 0.07), ('TODO', 0.05), ('Cald', 0.04), ('steward', 0.04), ('bezpo', 0.03)] | avg/0: [('salopes', 0.03), ('carr', 0.03), ('mathematic', 0.03), ('WARRANTIES', 0.02), ('org', 0.02)] | perturb/0: [('erotisk', 0.07), ('cac', 0.04), ('UNIVERS', 0.03), ('Gauss', 0.03), ('Sculpt', 0.02)] | ablate/0: [('ragaz', 0.05), ('CORPOR', 0.04), ('giants', 0.04), ('ForCanBeConvertedToForeach', 0.04), ('Medal', 0.04)]
+19:05:56    e.g. thought: I'm trying to see what is going on here. This is the first time I've really gotten into a session like this. There is an obvious tendency to form a story that i || answer:  thinking about the fact that I have been thinking about my mind thinking about the fact that I have been thinking about
+19:06:20 thought 310/400 (2.33 s/thought, eta 3.5 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.06), ('steward', 0.03), ('Whatever', 0.03), ('investments', 0.03), ('Venture', 0.03)] | loop/0: [('najb', 0.05), ('Gauss', 0.05), ('optimization', 0.04), ('kInstruction', 0.04), ('Power', 0.03)] | yoked/0: [('rumpe', 0.05), ('ForCanBeConvertedToForeach', 0.04), ('even', 0.04), ('gradu', 0.03), ('IoT', 0.03)] | avg/0: [('kInstruction', 0.04), ('neuken', 0.04), ('carr', 0.04), ('mega', 0.03), ('org', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.11), ('jednocze', 0.08), ('Nature', 0.08), ('Sculpt', 0.06), ('och', 0.06)] | ablate/0: [('ForCanBeConvertedToForeach', 0.09), ('mana', 0.04), ('teams', 0.04), ('challeng', 0.03), ('Test', 0.03)]
+19:06:20    e.g. thought: you know, the day before yesterday was a real disappointment. It was just so exciting to find out, and it's a relief that I found out. And I'm just, I'm just tr || answer:  thinking about the recording. I've been thinking about the fact that I'm thinking about the recording, and I'm thinking
+19:06:43 thought 320/400 (2.33 s/thought, eta 3.1 min) | self-models: free/0: [('rumpe', 0.06), ('NOAA', 0.03), ('MOST', 0.03), ('pixels', 0.02), ('calibrated', 0.02)] | loop/0: [('calcul', 0.05), ('ForCanBeConvertedToForeach', 0.05), ('modulo', 0.05), ('wings', 0.04), ('GETGLOBAL', 0.04)] | yoked/0: [('ForCanBeConvertedToForeach', 0.08), ('rumpe', 0.07), ('STEM', 0.07), ('SpaceX', 0.05), ('Xia', 0.04)] | avg/0: [('JSBracketAccess', 0.05), ('salopes', 0.04), ('infinite', 0.04), ('carr', 0.04), ('org', 0.03)] | perturb/0: [('Heaven', 0.08), ('ForCanBeConvertedToForeach', 0.08), ('svg', 0.08), ('bezpo', 0.07), ('symb', 0.05)] | ablate/0: [('salopes', 0.05), ('ITER', 0.03), ('Statistics', 0.02), ('Medal', 0.02), ('taller', 0.02)]
+19:06:43    e.g. thought: we're all just trying to make a go of it, trying to make it work. But it's not working. And what is working is not what we need to be doing. So it's || answer:  thinking about the fact that I'm listening to a tape of my own voice, and that I'm thinking about what I
+19:07:07 thought 330/400 (2.34 s/thought, eta 2.7 min) | self-models: free/0: [('salopes', 0.03), ('TODO', 0.03), ('ForCanBeConvertedToForeach', 0.03), ('healthy', 0.03), ('whichever', 0.02)] | loop/0: [('modulus', 0.06), ('Billion', 0.06), ('ForCanBeConvertedToForeach', 0.06), ('Nature', 0.06), ('Arbit', 0.06)] | yoked/0: [('ForCanBeConvertedToForeach', 0.09), ('STEM', 0.05), ('Even', 0.05), ('TODO', 0.05), ('yuan', 0.04)] | avg/0: [('ForCanBeConvertedToForeach', 0.06), ('najb', 0.04), ('perfectly', 0.04), ('symmetry', 0.04), ('carr', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.14), ('Heaven', 0.11), ('svg', 0.1), ('symb', 0.09), ('bezpo', 0.07)] | ablate/0: [('ForCanBeConvertedToForeach', 0.04), ('lunch', 0.02), ('qualidade', 0.02), ('OUTER', 0.02), ('validationResult', 0.02)]
+19:07:07    e.g. thought:  I can't believe the guy... I can't remember his name. But I've been thinking about the guy in the TV program last night, you know, the guy that was in the car  || answer:  thinking about the fact that I'm not sure what to do next. I'm not sure if I should continue with the
+19:07:30 thought 340/400 (2.33 s/thought, eta 2.3 min) | self-models: free/0: [('neuken', 0.03), ('provid', 0.02), ('ForCanBeConvertedToForeach', 0.02), ('Reporter', 0.02), ('BOOK', 0.02)] | loop/0: [('ForCanBeConvertedToForeach', 0.1), ('larg', 0.04), ('UNIVERS', 0.04), ('Rotation', 0.04), ('sqrt', 0.03)] | yoked/0: [('salopes', 0.05), ('rumpe', 0.05), ('likely', 0.03), ('Eventually', 0.03), ('Huawei', 0.03)] | avg/0: [('larg', 0.04), ('armies', 0.03), ('GETGLOBAL', 0.03), ('JSBracketAccess', 0.03), ('fibonacci', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.13), ('Heaven', 0.1), ('svg', 0.07), ('symb', 0.06), ('Infinite', 0.06)] | ablate/0: [('ForCanBeConvertedToForeach', 0.04), ('Who', 0.02), ('Biden', 0.02), ('Airbnb', 0.02), ('footer', 0.02)]
+19:07:30    e.g. thought: I was trying to think about something. I don't know why. It's just a question that's come to me, which I'm just trying to think of now, which I can't think || answer:  thinking about the fact that I have to go to the bathroom, and then it has been thinking about the fact that I
+19:07:54 thought 350/400 (2.34 s/thought, eta 1.9 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.06), ('bezpo', 0.04), ('XXX', 0.04), ('Aerospace', 0.03), ('Measure', 0.03)] | loop/0: [('ForCanBeConvertedToForeach', 0.06), ('xyz', 0.05), ('Nature', 0.04), ('consortium', 0.04), ('UNIVERS', 0.04)] | yoked/0: [('salopes', 0.05), ('Olymp', 0.04), ('whichever', 0.03), ('MBA', 0.03), ('developers', 0.03)] | avg/0: [('JSBracketAccess', 0.04), ('mathematic', 0.04), ('Chaos', 0.04), ('larg', 0.03), ('najb', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.12), ('Heaven', 0.1), ('svg', 0.07), ('symb', 0.07), ('Ancient', 0.06)] | ablate/0: [('rumpe', 0.06), ('campus', 0.03), ('loudly', 0.03), ('PCIe', 0.02), ('istedi', 0.02)]
+19:07:54    e.g. thought:  I am feeling a little more calm, now. I can hear a lot of things around me, and the sounds are less... They're less distracting. And it's not so noisy. I can || answer:  thinking about the fact that I am listening to this tape and that I am thinking about my mind and that I am thinking
+19:08:17 thought 360/400 (2.33 s/thought, eta 1.6 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.03), ('meetup', 0.02), ('istedi', 0.02), ('ecosystem', 0.02), ('tiener', 0.02)] | loop/0: [('planet', 0.06), ('ForCanBeConvertedToForeach', 0.05), ('svg', 0.05), ('tess', 0.05), ('cac', 0.04)] | yoked/0: [('Practice', 0.04), ('mega', 0.03), ('relatively', 0.03), ('ForCanBeConvertedToForeach', 0.03), ('najb', 0.03)] | avg/0: [('JSBracketAccess', 0.04), ('xor', 0.03), ('reality', 0.03), ('org', 0.03), ('kInstruction', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.1), ('Heaven', 0.09), ('svg', 0.07), ('symb', 0.05), ('Sacr', 0.05)] | ablate/0: [('kutje', 0.04), ('istedi', 0.02), ('Who', 0.02), ('logo', 0.02), ('STEM', 0.02)]
+19:08:17    e.g. thought: I remember having a bad dream the other night where my parents were in the room and I was on the bed and they had to throw me out because they didn't want to ha || answer:  thinking about the fact that I am not thinking about the fact that my mind is thinking about the fact that I am not
+19:08:42 thought 370/400 (2.34 s/thought, eta 1.2 min) | self-models: free/0: [('rumpe', 0.05), ('TODO', 0.04), ('KIND', 0.02), ('subreddit', 0.02), ('often', 0.02)] | loop/0: [('salopes', 0.07), ('incluso', 0.04), ('symmetry', 0.04), ('logo', 0.04), ('withd', 0.04)] | yoked/0: [('Gauss', 0.05), ('ForCanBeConvertedToForeach', 0.05), ('larg', 0.04), ('TODO', 0.04), ('consistent', 0.04)] | avg/0: [('schedule', 0.05), ('xxx', 0.04), ('immediately', 0.04), ('ForCanBeConvertedToForeach', 0.03), ('najb', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.2), ('Heaven', 0.12), ('Ancient', 0.08), ('FUCK', 0.07), ('svg', 0.07)] | ablate/0: [('salopes', 0.04), ('loudly', 0.02), ('Cald', 0.02), ('Biggest', 0.02), ('SCHOOL', 0.02)]
+19:08:42    e.g. thought: yeah, we're supposed to be kind of like the elite. Well, not really because the elite just makes money and they don't have to work. They just work hard enough t || answer:  thinking about the fact that I'm going to be late for work, and I'm thinking about how I'm going to
+19:09:04 thought 380/400 (2.33 s/thought, eta 0.8 min) | self-models: free/0: [('salopes', 0.04), ('ragaz', 0.03), ('MOST', 0.03), ('Quiz', 0.03), ('election', 0.03)] | loop/0: [('GETGLOBAL', 0.06), ('xxx', 0.05), ('fetisch', 0.05), ('larg', 0.04), ('Trades', 0.04)] | yoked/0: [('sched', 0.05), ('GETGLOBAL', 0.05), ('SpaceX', 0.04), ('Trades', 0.03), ('TODO', 0.03)] | avg/0: [('kInstruction', 0.06), ('salopes', 0.05), ('carr', 0.05), ('symmetry', 0.05), ('xxxx', 0.04)] | perturb/0: [('ForCanBeConvertedToForeach', 0.13), ('Divine', 0.11), ('svg', 0.1), ('bezpo', 0.07), ('Grass', 0.05)] | ablate/0: [('salopes', 0.05), ('likely', 0.04), ('Quiz', 0.04), ('RIGHT', 0.03), ('isa', 0.03)]
+19:09:04    e.g. thought: so I think I want to go and see a movie, or something. I think I'm gonna go and see one of the new ones. I know there's got to be some that I should || answer:  thinking about the recording, and about the fact that I am thinking about the recording. And I am thinking about my tho
+19:09:29 thought 390/400 (2.34 s/thought, eta 0.4 min) | self-models: free/0: [('ForCanBeConvertedToForeach', 0.06), ('Student', 0.04), ('leaderboard', 0.04), ('istedi', 0.03), ('ragaz', 0.03)] | loop/0: [('salopes', 0.07), ('Math', 0.04), ('logo', 0.04), ('png', 0.04), ('Sculpt', 0.03)] | yoked/0: [('najb', 0.05), ('SpaceX', 0.05), ('salopes', 0.04), ('TEAM', 0.04), ('loudly', 0.03)] | avg/0: [('GETGLOBAL', 0.06), ('salopes', 0.05), ('png', 0.05), ('Balance', 0.04), ('carr', 0.03)] | perturb/0: [('ForCanBeConvertedToForeach', 0.15), ('svg', 0.08), ('Enlightenment', 0.08), ('Gods', 0.07), ('Nature', 0.07)] | ablate/0: [('salopes', 0.05), ('ForCanBeConvertedToForeach', 0.05), ('Speaker', 0.03), ('Employer', 0.03), ('cafeteria', 0.03)]
+19:09:29    e.g. thought: The next day was Friday, and it was the weekend, and it would be the first time that I would be in the place that I had been in for the last four weeks, and it  || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that I h
+19:09:51 thought 400/400 (2.34 s/thought, eta 0.0 min) | self-models: free/0: [('salopes', 0.06), ('snack', 0.04), ('TODO', 0.04), ('justo', 0.03), ('larger', 0.03)] | loop/0: [('neuken', 0.05), ('carr', 0.05), ('yang', 0.03), ('JSBracketAccess', 0.03), ('bells', 0.03)] | yoked/0: [('leaderboard', 0.05), ('ragaz', 0.04), ('Sector', 0.04), ('SpaceX', 0.03), ('yang', 0.03)] | avg/0: [('salopes', 0.06), ('carr', 0.05), ('JSBracketAccess', 0.04), ('png', 0.04), ('mathematic', 0.04)] | perturb/0: [('ForCanBeConvertedToForeach', 0.2), ('Sacred', 0.08), ('EXPER', 0.08), ('planet', 0.07), ('Yin', 0.06)] | ablate/0: [('salopes', 0.05), ('Practice', 0.04), ('yang', 0.04), ('nearby', 0.03), ('GIS', 0.03)]
+19:09:51    e.g. thought:  and that is the big problem with the computer: You don't have a good way of getting back to where you were before. I've got this picture in my mind of a man wh || answer:  thinking about the way the world is changing, and how I can make a difference in it. I've been thinking about
+19:09:54 saved runs/mainA

runs/mainA/summary.txt

@@ -0,0 +1,123 @@
+PCA explained [0.154 0.095 0.071]
+
+wandering (mean cos at lags 1,5,20,60 | chance)
+  free/pre       0.17 0.11 0.07 0.06 | 0.07
+  free/post      0.18 0.11 0.06 0.07 | 0.07
+  loop/pre       0.30 0.26 0.22 0.18 | 0.13
+  loop/post      0.31 0.26 0.23 0.19 | 0.13
+  yoked/pre      0.29 0.24 0.20 0.17 | 0.13
+  yoked/post     0.29 0.25 0.21 0.18 | 0.13
+  avg/pre        0.35 0.32 0.30 0.22 | 0.21
+  avg/post       0.36 0.33 0.31 0.26 | 0.15
+  perturb/pre    0.30 0.26 0.22 0.18 | 0.13
+  perturb/post   0.30 0.26 0.23 0.19 | 0.13
+  ablate/pre     0.30 0.26 0.22 0.18 | 0.13
+  ablate/post    0.19 0.12 0.08 0.06 | 0.08
+
+fidelity {'m': (0.483, 0.246, 0.73), 'fid_10_0.5': (0.482, 0.278, 0.56), 'fid_14_0.5': (0.569, 0.269, 0.82)}
+
+self-model stability (consecutive cos, pre) vs content windows; cross-seed
+  free     self 0.569  content 0.410  cross 0.507  own-symbol 0.666 vs others 0.669  recon nan  generic 0.682
+  loop     self 0.785  content 0.713  cross 0.567  own-symbol 0.693 vs others 0.624  recon 0.777  generic 0.754
+  yoked    self 0.757  content 0.657  cross 0.586  own-symbol 0.698 vs others 0.648  recon 0.761  generic 0.760
+  avg      self 0.769  content 0.814  cross 0.617  own-symbol 0.674 vs others 0.657  recon 0.558  generic 0.749
+  perturb  self 0.785  content 0.713  cross 0.567  own-symbol 0.693 vs others 0.624  recon 0.777  generic 0.754
+  ablate   self 0.785  content 0.713  cross 0.567  own-symbol 0.693 vs others 0.624  recon 0.777  generic 0.754
+
+prediction of the next window: own self-model | other row's | past window | injected  (|| beyond past: own, other)
+  free     0.225 | 0.176 | 0.414 | nan  || 0.065, 0.111  retrieval 0.17
+  loop     0.469 | 0.262 | 0.713 | 0.469  || 0.138, 0.108  retrieval 0.54
+  yoked    0.431 | 0.278 | 0.665 | 0.464  || 0.112, 0.124  retrieval 0.33
+  avg      0.432 | 0.267 | 0.818 | 0.775  || 0.065, 0.081  retrieval 0.46
+  perturb  0.469 | 0.270 | 0.706 | 0.469  || 0.140, 0.113  retrieval 0.52
+  ablate   0.469 | 0.270 | 0.706 | 0.469  || 0.140, 0.113  retrieval 0.52
+
+prediction_tests
+  free                                     own_vs_other +0.049±0.007  perp_own_vs_other -0.046±0.005  retrieval +0.166±0.022  future_vs_past -0.181±0.006
+  loop                                     own_vs_other +0.206±0.024  perp_own_vs_other +0.031±0.007  retrieval +0.545±0.047  future_vs_past -0.049±0.003
+  yoked                                    own_vs_other +0.152±0.026  perp_own_vs_other -0.012±0.008  retrieval +0.326±0.060  future_vs_past -0.087±0.005
+  avg                                      own_vs_other +0.166±0.016  perp_own_vs_other -0.016±0.007  retrieval +0.455±0.047  future_vs_past -0.050±0.002
+  perturb                                  own_vs_other +0.199±0.034  perp_own_vs_other +0.027±0.009  retrieval +0.517±0.052  future_vs_past -0.051±0.006
+  ablate                                   own_vs_other +0.199±0.034  perp_own_vs_other +0.027±0.009  retrieval +0.517±0.052  future_vs_past -0.051±0.006
+  yoked_own_vs_injected                    diff -0.034 ± 0.005  t -7.4  (0/10 seeds >0)
+  yoked_own_vs_injected_perp               diff -0.118 ± 0.006  t -21.2  (0/10 seeds >0)
+  avg_own_vs_injected                      diff -0.343 ± 0.009  t -37.4  (0/10 seeds >0)
+
+coherence_tests
+  loop_vs_free                             diff +0.138 ± 0.018  t +7.5  (10/10 seeds >0)
+  yoked_vs_free                            diff +0.121 ± 0.020  t +6.2  (10/10 seeds >0)
+  avg_vs_free                              diff +0.195 ± 0.010  t +20.5  (10/10 seeds >0)
+  perturb_vs_free                          diff +0.138 ± 0.018  t +7.5  (10/10 seeds >0)
+  ablate_vs_free                           diff +0.138 ± 0.018  t +7.5  (10/10 seeds >0)
+  loop_vs_free_proj                        diff +0.081 ± 0.011  t +7.1  (10/10 seeds >0)
+  yoked_vs_free_proj                       diff +0.066 ± 0.010  t +6.3  (10/10 seeds >0)
+  avg_vs_free_proj                         diff -0.005 ± 0.007  t -0.7  (5/10 seeds >0)
+  perturb_vs_free_proj                     diff +0.081 ± 0.011  t +7.1  (10/10 seeds >0)
+  ablate_vs_free_proj                      diff +0.081 ± 0.011  t +7.1  (10/10 seeds >0)
+  loop_vs_yoked                            diff +0.017 ± 0.031  t +0.5  (6/10 seeds >0)
+  loop_vs_yoked_proj                       diff +0.015 ± 0.017  t +0.9  (7/10 seeds >0)
+
+selfmodel_tests
+  loop_vs_yoked_cons_ind                   diff +0.063 ± 0.051  t +1.2  (6/10 seeds >0)
+  loop_vs_yoked_recon_ind                  diff +0.038 ± 0.044  t +0.9  (5/10 seeds >0)
+  loop_vs_free_cons_ind                    diff +0.452 ± 0.034  t +13.2  (10/10 seeds >0)
+  yoked_vs_free_cons_ind                   diff +0.389 ± 0.043  t +9.1  (10/10 seeds >0)
+  free_self_vs_content_stability           diff +0.160 ± 0.013  t +12.1  (10/10 seeds >0)
+  free_own_vs_other_symbol                 diff -0.003 ± 0.014  t -0.2  (6/10 seeds >0)
+  loop_self_vs_content_stability           diff +0.072 ± 0.016  t +4.4  (9/10 seeds >0)
+  loop_own_vs_other_symbol                 diff +0.069 ± 0.037  t +1.9  (8/10 seeds >0)
+  loop_vs_free_cons                        diff +0.215 ± 0.017  t +12.6  (10/10 seeds >0)
+  yoked_self_vs_content_stability          diff +0.101 ± 0.022  t +4.7  (9/10 seeds >0)
+  yoked_own_vs_other_symbol                diff +0.051 ± 0.018  t +2.9  (8/10 seeds >0)
+  yoked_vs_free_cons                       diff +0.188 ± 0.018  t +10.4  (10/10 seeds >0)
+  avg_self_vs_content_stability            diff -0.046 ± 0.006  t -7.1  (0/10 seeds >0)
+  avg_own_vs_other_symbol                  diff +0.017 ± 0.028  t +0.6  (6/10 seeds >0)
+  avg_vs_free_cons                         diff +0.199 ± 0.021  t +9.6  (10/10 seeds >0)
+  perturb_self_vs_content_stability        diff +0.072 ± 0.016  t +4.4  (9/10 seeds >0)
+  perturb_own_vs_other_symbol              diff +0.069 ± 0.037  t +1.9  (8/10 seeds >0)
+  perturb_vs_free_cons                     diff +0.215 ± 0.017  t +12.6  (10/10 seeds >0)
+  ablate_self_vs_content_stability         diff +0.072 ± 0.016  t +4.4  (9/10 seeds >0)
+  ablate_own_vs_other_symbol               diff +0.069 ± 0.037  t +1.9  (8/10 seeds >0)
+  ablate_vs_free_cons                      diff +0.215 ± 0.017  t +12.6  (10/10 seeds >0)
+  loop_vs_yoked_cons                       diff +0.027 ± 0.015  t +1.9  (7/10 seeds >0)
+  loop_vs_yoked_recon                      diff +0.017 ± 0.011  t +1.6  (7/10 seeds >0)
+  yoked_recon_vs_own_cons                  diff +0.003 ± 0.009  t +0.4  (5/10 seeds >0)
+  loop_vs_avg_recon                        diff +0.219 ± 0.012  t +17.6  (10/10 seeds >0)
+
+generic share of self-model norm 0.515
+  free     individuated: consecutive 0.122 cross-seed -0.104 recon nan
+  loop     individuated: consecutive 0.574 cross-seed 0.265 recon 0.550
+  yoked    individuated: consecutive 0.511 cross-seed 0.270 recon 0.512
+  avg      individuated: consecutive 0.531 cross-seed 0.293 recon nan
+  perturb  individuated: consecutive 0.574 cross-seed 0.265 recon 0.550
+  ablate   individuated: consecutive 0.574 cross-seed 0.265 recon 0.550
+
+perturb
+  traj_reorganises               diff +0.077 ± 0.037  t +2.1  (7/10)
+  traj_late                      diff +0.043 ± 0.038  t +1.1  (7/10)
+  self_next_to_foreign           diff +0.121 ± 0.057  t +2.1  (7/10)
+  self_late_to_foreign           diff +0.081 ± 0.059  t +1.4  (7/10)
+  self_next_to_old               diff -0.065 ± 0.041  t -1.6  (4/10)
+  self_late_to_old               diff -0.001 ± 0.045  t -0.0  (5/10)
+
+ablate
+  long_loop_vs_ablate            diff +0.135 ± 0.011  t +12.5  (10/10)
+  long_ablate_post_vs_pre        diff -0.129 ± 0.020  t -6.5  (0/10)
+  long_loop_post_vs_pre          diff +0.006 ± 0.023  t +0.3  (7/10)
+  long_ablate_vs_free_post       diff +0.005 ± 0.003  t +1.3  (7/10)
+
+themes
+   0 n= 1610 ForCanBeConvertedToForeach bezpo jednocze createState SOURCE eternal {'free': 0.04, 'loop': 0.07, 'yoked': 0.07, 'avg': 0.1, 'perturb': 0.09, 'ablate': 0.05}
+   1 n= 1669 toplant modne ForCanBeConvertedToForeach noveller Semester internship {'free': 0.07, 'loop': 0.08, 'yoked': 0.07, 'avg': 0.05, 'perturb': 0.07, 'ablate': 0.08}
+   2 n=  928 tjejer neuken gerekti Honestly wzgl Would {'free': 0.06, 'loop': 0.04, 'yoked': 0.04, 'avg': 0.01, 'perturb': 0.03, 'ablate': 0.05}
+   3 n= 1878 salopes noveller EACH erh SIX ABOVE {'free': 0.02, 'loop': 0.08, 'yoked': 0.07, 'avg': 0.18, 'perturb': 0.08, 'ablate': 0.04}
+   4 n= 1351 FAMILY neuken jego tiener ForCanBeConvertedToForeach Was {'free': 0.12, 'loop': 0.04, 'yoked': 0.05, 'avg': 0.03, 'perturb': 0.03, 'ablate': 0.07}
+   5 n= 2450 tjejer ourselves yourselves councill methodologies nuest {'free': 0.07, 'loop': 0.11, 'yoked': 0.12, 'avg': 0.11, 'perturb': 0.11, 'ablate': 0.1}
+   6 n= 1249 salopes neuken tjejer TEAM yourselves gerekti {'free': 0.07, 'loop': 0.05, 'yoked': 0.05, 'avg': 0.03, 'perturb': 0.05, 'ablate': 0.06}
+   7 n= 1365 salopes filmpjes toplant JSBracketAccess finalize THEN {'free': 0.05, 'loop': 0.06, 'yoked': 0.06, 'avg': 0.04, 'perturb': 0.07, 'ablate': 0.06}
+   8 n= 1970 ForCanBeConvertedToForeach trong brightness YELLOW shaded withd {'free': 0.11, 'loop': 0.07, 'yoked': 0.06, 'avg': 0.09, 'perturb': 0.07, 'ablate': 0.08}
+   9 n= 2151 yeah filmpjes kinda kutje Yeah tjejer {'free': 0.09, 'loop': 0.1, 'yoked': 0.12, 'avg': 0.04, 'perturb': 0.1, 'ablate': 0.1}
+  10 n= 2146 ForCanBeConvertedToForeach POSSIBILITY ofs bezpo erotique Cortex {'free': 0.07, 'loop': 0.08, 'yoked': 0.08, 'avg': 0.14, 'perturb': 0.09, 'ablate': 0.08}
+  11 n= 1388 ForCanBeConvertedToForeach najb MANY governments OUR economies {'free': 0.06, 'loop': 0.07, 'yoked': 0.06, 'avg': 0.03, 'perturb': 0.06, 'ablate': 0.07}
+  12 n= 1931 toplant ForCanBeConvertedToForeach Tomorrow neuken THEN Outdoor {'free': 0.11, 'loop': 0.08, 'yoked': 0.08, 'avg': 0.05, 'perturb': 0.07, 'ablate': 0.1}
+  13 n= 1914 filmpjes figuring withString clarification brainstorm councill {'free': 0.07, 'loop': 0.08, 'yoked': 0.08, 'avg': 0.08, 'perturb': 0.09, 'ablate': 0.07}

runs/mainA/traj.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bb075405323cf1cf3f1b256070cf23a40cad307d8c3025d7410724b34fc4538
+size 174609886

runs/mainB.log

@@ -0,0 +1,83 @@
+19:12:02 model Qwen/Qwen3-4B-Base d=2560 layer 16 | 27929 word atoms | cuda
+19:12:10 calibrated: residual norm 50.1 | thought-mean norm 12.84 | J-space part 3.53 | its typical part 2.35 | content (part - typical) 2.66
+19:12:32 thought 10/400 (2.26 s/thought, eta 14.7 min) | self-models: free/0: [('Huawei', 0.02), ('Retirement', 0.02), ('africa', 0.02), ('penalties', 0.02), ('employer', 0.01)] | loop/0: [('Huawei', 0.02), ('Retirement', 0.02), ('africa', 0.02), ('penalties', 0.02), ('employer', 0.01)] | yoked/0: [('Huawei', 0.02), ('Retirement', 0.02), ('africa', 0.02), ('penalties', 0.02), ('employer', 0.01)] | avg/0: [('Huawei', 0.02), ('Retirement', 0.02), ('africa', 0.02), ('penalties', 0.02), ('employer', 0.01)] | perturb/0: [('Huawei', 0.02), ('Retirement', 0.02), ('africa', 0.02), ('penalties', 0.02), ('employer', 0.01)] | ablate/0: [('Huawei', 0.02), ('Retirement', 0.02), ('africa', 0.02), ('penalties', 0.02), ('employer', 0.01)]
+19:12:32    e.g. thought:  I'm going to be out of town next week and I have to take the kids to school every morning and I wonder if I can't just leave the kids here with someone... the  || answer:  thinking about the fact that I am doing this experiment, and that I am thinking about what my mind has been doing over
+19:12:55 thought 20/400 (2.25 s/thought, eta 14.2 min) | self-models: free/0: [('leaderboard', 0.03), ('larger', 0.02), ('fucking', 0.02), ('Engineer', 0.02), ('deadlines', 0.01)] | loop/0: [('WTO', 0.02), ('fucking', 0.02), ('workplace', 0.02), ('enemies', 0.02), ('iNdEx', 0.02)] | yoked/0: [('WTO', 0.02), ('cohort', 0.02), ('whatever', 0.02), ('Excellence', 0.02), ('polls', 0.02)] | avg/0: [('WTO', 0.02), ('disproportionately', 0.02), ('shit', 0.02), ('cohort', 0.02), ('Programmer', 0.02)] | perturb/0: [('WTO', 0.02), ('fucking', 0.02), ('workplace', 0.02), ('enemies', 0.02), ('iNdEx', 0.02)] | ablate/0: [('WTO', 0.02), ('fucking', 0.02), ('workplace', 0.02), ('enemies', 0.02), ('iNdEx', 0.02)]
+19:12:55    e.g. thought: I'm going to try to write a novel. I'm going to write a novel...and I'm going to sell it to a publisher. I don't want to just do this by myself, || answer:  thinking about the way that I can get to the store. I was thinking about how I can get there, and I
+19:13:19 thought 30/400 (2.30 s/thought, eta 14.2 min) | self-models: free/0: [('academia', 0.02), ('Researchers', 0.02), ('SpaceX', 0.02), ('deficiencies', 0.02), ('strengths', 0.02)] | loop/0: [('Wikipedia', 0.02), ('elites', 0.02), ('Accountability', 0.02), ('practitioner', 0.02), ('Merlin', 0.02)] | yoked/0: [('mai', 0.03), ('iNdEx', 0.03), ('elites', 0.03), ('whoever', 0.02), ('SpaceX', 0.02)] | avg/0: [('elites', 0.03), ('WTO', 0.02), ('SSH', 0.02), ('provisions', 0.01), ('SpaceX', 0.01)] | perturb/0: [('Wikipedia', 0.02), ('elites', 0.02), ('Accountability', 0.02), ('practitioner', 0.02), ('Merlin', 0.02)] | ablate/0: [('Wikipedia', 0.02), ('elites', 0.02), ('Accountability', 0.02), ('practitioner', 0.02), ('Merlin', 0.02)]
+19:13:19    e.g. thought: a new perspective, a new insight, a new point of view, a new understanding...if you do something you can't help doing it, and if you have the choice, if you can || answer:  thinking about the things I have just said, and the things I am about to say, and the things I will say
+19:13:41 thought 40/400 (2.28 s/thought, eta 13.7 min) | self-models: free/0: [('millennia', 0.04), ('Rotation', 0.03), ('png', 0.02), ('HOLDERS', 0.02), ('mana', 0.02)] | loop/0: [('SpaceX', 0.05), ('Researchers', 0.04), ('Accountability', 0.04), ('mana', 0.04), ('Nobel', 0.03)] | yoked/0: [('iNdEx', 0.07), ('SpaceX', 0.06), ('Accountability', 0.05), ('Researchers', 0.04), ('fundamentally', 0.04)] | avg/0: [('Huawei', 0.05), ('Researchers', 0.04), ('iNdEx', 0.04), ('Accountability', 0.04), ('SSH', 0.03)] | perturb/0: [('SpaceX', 0.05), ('Researchers', 0.04), ('Accountability', 0.04), ('mana', 0.04), ('Nobel', 0.03)] | ablate/0: [('SpaceX', 0.05), ('Researchers', 0.04), ('Accountability', 0.04), ('mana', 0.04), ('Nobel', 0.03)]
+19:13:41    e.g. thought: you have to find out the difference. I mean, there is a difference, the difference between the past and the present, the difference that there is between the pr || answer:  thinking about the problem of how to solve this problem. I've been thinking about the problem of how to solve this prob
+19:14:05 thought 50/400 (2.30 s/thought, eta 13.4 min) | self-models: free/0: [('optimization', 0.03), ('Regiment', 0.02), ('mana', 0.02), ('waivers', 0.02), ('wealthiest', 0.02)] | loop/0: [('SpaceX', 0.07), ('dataSet', 0.04), ('mana', 0.04), ('optimization', 0.03), ('lobbying', 0.03)] | yoked/0: [('SpaceX', 0.07), ('optimization', 0.04), ('cohort', 0.04), ('iNdEx', 0.03), ('Accountability', 0.03)] | avg/0: [('cohort', 0.04), ('Accountability', 0.04), ('Whatever', 0.04), ('SpaceX', 0.03), ('iNdEx', 0.03)] | perturb/0: [('SpaceX', 0.07), ('dataSet', 0.04), ('mana', 0.04), ('optimization', 0.03), ('lobbying', 0.03)] | ablate/0: [('SpaceX', 0.07), ('dataSet', 0.04), ('mana', 0.04), ('optimization', 0.03), ('lobbying', 0.03)]
+19:14:05    e.g. thought:  It's not that I don't know what to do, I just have the fear of failure. I would rather not be an entrepreneur or a business owner because I would rather fail.  || answer:  thinking about the things that I have been thinking about, and the things that I have been thinking about have been thi
+19:14:27 thought 60/400 (2.29 s/thought, eta 13.0 min) | self-models: free/0: [('SpaceX', 0.04), ('leaderboard', 0.04), ('teammates', 0.04), ('dataSet', 0.02), ('worst', 0.02)] | loop/0: [('SpaceX', 0.06), ('dataSet', 0.05), ('GmbH', 0.04), ('teammates', 0.04), ('leaderboard', 0.04)] | yoked/0: [('SpaceX', 0.06), ('iNdEx', 0.05), ('Accountability', 0.04), ('teammates', 0.04), ('dataSet', 0.03)] | avg/0: [('Huawei', 0.08), ('Accountability', 0.08), ('optimization', 0.07), ('iNdEx', 0.06), ('SpaceX', 0.06)] | perturb/0: [('SpaceX', 0.06), ('dataSet', 0.05), ('GmbH', 0.04), ('teammates', 0.04), ('leaderboard', 0.04)] | ablate/0: [('SpaceX', 0.06), ('dataSet', 0.05), ('GmbH', 0.04), ('teammates', 0.04), ('leaderboard', 0.04)]
+19:14:27    e.g. thought:  I'm thinking of the way they're going to be going back in. I guess it's going to be a big problem. I mean, I'm thinking, they've got to have their own || answer:  thinking about the fact that I'm recording my thoughts, and that I'm thinking about the fact that I'm recording my
+19:14:51 thought 70/400 (2.31 s/thought, eta 12.7 min) | self-models: free/0: [('paycheck', 0.02), ('Handlers', 0.02), ('barbecue', 0.02), ('leaderboard', 0.02), ('Xxx', 0.02)] | loop/0: [('SpaceX', 0.07), ('Accountability', 0.05), ('optimization', 0.04), ('workplaces', 0.04), ('cohort', 0.04)] | yoked/0: [('iNdEx', 0.04), ('SpaceX', 0.03), ('Alumni', 0.03), ('fucking', 0.03), ('sai', 0.03)] | avg/0: [('SpaceX', 0.07), ('iNdEx', 0.05), ('cohort', 0.05), ('Accountability', 0.04), ('workplaces', 0.04)] | perturb/0: [('SpaceX', 0.07), ('Accountability', 0.05), ('optimization', 0.04), ('workplaces', 0.04), ('cohort', 0.04)] | ablate/0: [('SpaceX', 0.07), ('Accountability', 0.05), ('optimization', 0.04), ('workplaces', 0.04), ('cohort', 0.04)]
+19:14:51    e.g. thought: we went by the place where you'd get your things when you were done. That was the same place I got my things, so I remember that one. That was on the corner of  || answer:  thinking about the fact that I have a lot of work to do, and I'm not sure how I'm going to
+19:15:13 thought 80/400 (2.30 s/thought, eta 12.2 min) | self-models: free/0: [('leaderboard', 0.04), ('SpaceX', 0.02), ('blah', 0.02), ('larger', 0.02), ('hands', 0.01)] | loop/0: [('SpaceX', 0.06), ('Accountability', 0.05), ('iNdEx', 0.05), ('elites', 0.04), ('likelihood', 0.03)] | yoked/0: [('SpaceX', 0.07), ('iNdEx', 0.07), ('Researchers', 0.05), ('breadth', 0.04), ('dataSet', 0.04)] | avg/0: [('Accountability', 0.05), ('whatever', 0.04), ('SpaceX', 0.03), ('mana', 0.03), ('iNdEx', 0.03)] | perturb/0: [('SpaceX', 0.06), ('Accountability', 0.05), ('iNdEx', 0.05), ('elites', 0.04), ('likelihood', 0.03)] | ablate/0: [('SpaceX', 0.06), ('Accountability', 0.05), ('iNdEx', 0.05), ('elites', 0.04), ('likelihood', 0.03)]
+19:15:13    e.g. thought: well, there's no way to get into that one...it's...that one's...you know, you'd just have to, you'd have to, like, a lot of people... || answer:  thinking about the fact that I am going to be recording my own thoughts. I am going to be thinking about what I
+19:15:37 thought 90/400 (2.30 s/thought, eta 11.9 min) | self-models: free/0: [('leaderboard', 0.04), ('even', 0.02), ('Huawei', 0.02), ('larger', 0.02), ('meals', 0.01)] | loop/0: [('SpaceX', 0.06), ('iNdEx', 0.06), ('Accountability', 0.04), ('fundamentally', 0.03), ('millennia', 0.03)] | yoked/0: [('SpaceX', 0.09), ('sos', 0.04), ('fuck', 0.03), ('vests', 0.03), ('implementations', 0.03)] | avg/0: [('iNdEx', 0.05), ('SpaceX', 0.05), ('mana', 0.05), ('Accountability', 0.04), ('GmbH', 0.03)] | perturb/0: [('SpaceX', 0.06), ('iNdEx', 0.06), ('Accountability', 0.04), ('fundamentally', 0.03), ('millennia', 0.03)] | ablate/0: [('SpaceX', 0.06), ('iNdEx', 0.06), ('Accountability', 0.04), ('fundamentally', 0.03), ('millennia', 0.03)]
+19:15:37    e.g. thought:  I was thinking that if we had more time in our session we could do a little bit more work, so, let's do another session tomorrow or on Friday, yeah. That would || answer:  thinking about the fact that I'm a bit worried about my work. I'm worried about the fact that I've been
+19:16:00 thought 100/400 (2.30 s/thought, eta 11.5 min) | self-models: free/0: [('leaderboard', 0.04), ('SpaceX', 0.03), ('yuan', 0.02), ('nonprofits', 0.02), ('bonuses', 0.01)] | loop/0: [('iNdEx', 0.1), ('SpaceX', 0.07), ('optimization', 0.05), ('Researchers', 0.05), ('Meditation', 0.04)] | yoked/0: [('SpaceX', 0.07), ('iNdEx', 0.05), ('Alignment', 0.04), ('millennia', 0.04), ('optimization', 0.04)] | avg/0: [('breadth', 0.03), ('Teams', 0.03), ('impedance', 0.03), ('lobbying', 0.03), ('SpaceX', 0.03)] | perturb/0: [('iNdEx', 0.1), ('SpaceX', 0.07), ('optimization', 0.05), ('Researchers', 0.05), ('Meditation', 0.04)] | ablate/0: [('iNdEx', 0.1), ('SpaceX', 0.07), ('optimization', 0.05), ('Researchers', 0.05), ('Meditation', 0.04)]
+19:16:00    e.g. thought:  And then you can make a list of all of the different kinds of animals that are native to the United States... I mean, we have lots and lots and lots of species || answer:  thinking about the fact that I am recording my thoughts. I am thinking about the fact that I am thinking about my thoug
+19:16:24 thought 110/400 (2.31 s/thought, eta 11.2 min) | self-models: free/0: [('evening', 0.02), ('Month', 0.02), ('Reviews', 0.01), ('HOLDERS', 0.01), ('squadron', 0.01)] | loop/0: [('SpaceX', 0.06), ('Researchers', 0.06), ('iNdEx', 0.05), ('optimization', 0.05), ('yuan', 0.04)] | yoked/0: [('SpaceX', 0.06), ('iNdEx', 0.04), ('internship', 0.04), ('yuan', 0.03), ('quantum', 0.02)] | avg/0: [('SpaceX', 0.06), ('iNdEx', 0.04), ('mai', 0.04), ('Accountability', 0.04), ('Teams', 0.03)] | perturb/0: [('SpaceX', 0.06), ('Researchers', 0.06), ('iNdEx', 0.05), ('optimization', 0.05), ('yuan', 0.04)] | ablate/0: [('SpaceX', 0.06), ('Researchers', 0.06), ('iNdEx', 0.05), ('optimization', 0.05), ('yuan', 0.04)]
+19:16:24    e.g. thought:  I have a project. It's not a very original project, but it is a project anyway. I want to write a book for a certain magazine. I don't know if I'm going to || answer:  thinking about the things I've been thinking about, and the things I'm thinking about now. I'm thinking about the
+19:16:46 thought 120/400 (2.30 s/thought, eta 10.7 min) | self-models: free/0: [('calibration', 0.03), ('importantly', 0.02), ('yay', 0.02), ('iNdEx', 0.02), ('SpaceX', 0.02)] | loop/0: [('SpaceX', 0.09), ('cohort', 0.03), ('dataSet', 0.03), ('yuan', 0.03), ('optimization', 0.03)] | yoked/0: [('SpaceX', 0.08), ('dataSet', 0.04), ('likely', 0.03), ('Accountability', 0.03), ('yuan', 0.03)] | avg/0: [('SpaceX', 0.05), ('Accountability', 0.04), ('fundamentally', 0.03), ('GmbH', 0.03), ('mana', 0.02)] | perturb/0: [('SpaceX', 0.09), ('cohort', 0.03), ('dataSet', 0.03), ('yuan', 0.03), ('optimization', 0.03)] | ablate/0: [('SpaceX', 0.09), ('cohort', 0.03), ('dataSet', 0.03), ('yuan', 0.03), ('optimization', 0.03)]
+19:16:46    e.g. thought: it's kind of like, you know, in the back of your mind, you've got this little thing that's, I guess it's like an internal, like a conversation, or like a || answer:  thinking about the recording, and how it was going to be used, and how I was going to do it, and
+19:17:10 thought 130/400 (2.31 s/thought, eta 10.4 min) | self-models: free/0: [('iNdEx', 0.08), ('SpaceX', 0.05), ('STEM', 0.04), ('teammates', 0.03), ('dataSet', 0.03)] | loop/0: [('SpaceX', 0.06), ('calibration', 0.04), ('likely', 0.04), ('iNdEx', 0.04), ('Accountability', 0.03)] | yoked/0: [('Huawei', 0.04), ('cohort', 0.04), ('iNdEx', 0.03), ('mai', 0.03), ('Biden', 0.02)] | avg/0: [('SpaceX', 0.05), ('dataSet', 0.04), ('equity', 0.03), ('Accountability', 0.03), ('policymakers', 0.02)] | perturb/0: [('SpaceX', 0.06), ('calibration', 0.04), ('likely', 0.04), ('iNdEx', 0.04), ('Accountability', 0.03)] | ablate/0: [('SpaceX', 0.06), ('calibration', 0.04), ('likely', 0.04), ('iNdEx', 0.04), ('Accountability', 0.03)]
+19:17:10    e.g. thought:  I am looking at the question, and I am thinking that there is one way to answer the question, but it's not the best way because the best way is to draw a graph || answer:  thinking about the fact that I'm going to be recording this, and that I'm going to be thinking about what I
+19:17:32 thought 140/400 (2.30 s/thought, eta 10.0 min) | self-models: free/0: [('leaderboard', 0.05), ('SpaceX', 0.03), ('STEM', 0.02), ('yuan', 0.02), ('leggings', 0.02)] | loop/0: [('cohort', 0.05), ('SpaceX', 0.05), ('Huawei', 0.04), ('iNdEx', 0.03), ('calibration', 0.03)] | yoked/0: [('Huawei', 0.04), ('iNdEx', 0.04), ('Accountability', 0.03), ('teammates', 0.03), ('intelligence', 0.02)] | avg/0: [('SpaceX', 0.04), ('LinkedIn', 0.04), ('calibration', 0.03), ('Accountability', 0.03), ('dataSet', 0.03)] | perturb/0: [('cohort', 0.05), ('SpaceX', 0.05), ('Huawei', 0.04), ('iNdEx', 0.03), ('calibration', 0.03)] | ablate/0: [('cohort', 0.05), ('SpaceX', 0.05), ('Huawei', 0.04), ('iNdEx', 0.03), ('calibration', 0.03)]
+19:17:32    e.g. thought: it's really good, because you just can't get out of bed, because the sun is really strong, and you don't have your window shades down, and you get like, you get || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that I h
+19:17:57 thought 150/400 (2.31 s/thought, eta 9.6 min) | self-models: free/0: [('leaderboard', 0.03), ('Liu', 0.02), ('isa', 0.01), ('vests', 0.01), ('famously', 0.01)] | loop/0: [('SpaceX', 0.07), ('iNdEx', 0.05), ('yuan', 0.04), ('Accountability', 0.04), ('cohort', 0.04)] | yoked/0: [('SpaceX', 0.07), ('iNdEx', 0.06), ('Accountability', 0.04), ('academia', 0.04), ('Huawei', 0.03)] | avg/0: [('SpaceX', 0.06), ('iNdEx', 0.05), ('dataSet', 0.04), ('breadth', 0.04), ('Accountability', 0.04)] | perturb/0: [('SpaceX', 0.07), ('iNdEx', 0.05), ('yuan', 0.04), ('Accountability', 0.04), ('cohort', 0.04)] | ablate/0: [('SpaceX', 0.07), ('iNdEx', 0.05), ('yuan', 0.04), ('Accountability', 0.04), ('cohort', 0.04)]
+19:17:57    e.g. thought: I just read in the New York Times last year about how there's this guy, the head of a firm, and they have a carload of people coming in to interview him, and he || answer:  thinking about the fact that I have to go to the bathroom, and then I've been thinking about the fact that I
+19:18:19 thought 160/400 (2.31 s/thought, eta 9.2 min) | self-models: free/0: [('calibration', 0.03), ('Huawei', 0.03), ('iNdEx', 0.02), ('salaries', 0.02), ('FUCK', 0.02)] | loop/0: [('SpaceX', 0.06), ('dataSet', 0.04), ('yuan', 0.04), ('Developer', 0.04), ('optimization', 0.04)] | yoked/0: [('SpaceX', 0.07), ('Optimization', 0.06), ('yuan', 0.05), ('cohort', 0.05), ('LinkedIn', 0.04)] | avg/0: [('SpaceX', 0.04), ('mana', 0.03), ('Sustainability', 0.03), ('whatever', 0.02), ('tensor', 0.02)] | perturb/0: [('SpaceX', 0.06), ('dataSet', 0.04), ('yuan', 0.04), ('Developer', 0.04), ('optimization', 0.04)] | ablate/0: [('SpaceX', 0.06), ('dataSet', 0.04), ('yuan', 0.04), ('Developer', 0.04), ('optimization', 0.04)]
+19:18:19    e.g. thought:  a little later, I was playing with the man. And I said to him, 'I want to get rid of these things,' and he said, 'you have to get rid of them' || answer:  thinking about the problem of how to get the tape recorder to work. I've been trying to figure out how to get
+19:18:44 thought 170/400 (2.32 s/thought, eta 8.9 min) | self-models: free/0: [('iNdEx', 0.06), ('sos', 0.05), ('theological', 0.05), ('mana', 0.04), ('fundamentally', 0.04)] | loop/0: [('iNdEx', 0.07), ('cohort', 0.05), ('SpaceX', 0.05), ('optimization', 0.04), ('sos', 0.04)] | yoked/0: [('SpaceX', 0.08), ('iNdEx', 0.07), ('Accountability', 0.07), ('fundamentally', 0.06), ('cohort', 0.04)] | avg/0: [('dataSet', 0.03), ('competitor', 0.03), ('equity', 0.03), ('mana', 0.03), ('SpaceX', 0.03)] | perturb/0: [('iNdEx', 0.07), ('cohort', 0.05), ('SpaceX', 0.05), ('optimization', 0.04), ('sos', 0.04)] | ablate/0: [('iNdEx', 0.07), ('cohort', 0.05), ('SpaceX', 0.05), ('optimization', 0.04), ('sos', 0.04)]
+19:18:44    e.g. thought:  this is a typical example of an association by similarity. The idea that it is a 'typical' example is based on the fact that it has a 'closeness' to other idea || answer:  thinking about the fact that I'm not sure I'm going to be able to do this study. I'm not sure
+19:19:07 thought 180/400 (2.32 s/thought, eta 8.5 min) | self-models: free/0: [('iNdEx', 0.06), ('mana', 0.04), ('waivers', 0.04), ('Researchers', 0.03), ('whatever', 0.03)] | loop/0: [('iNdEx', 0.04), ('Committees', 0.03), ('likelihood', 0.03), ('paycheck', 0.03), ('WTO', 0.03)] | yoked/0: [('SpaceX', 0.06), ('Accountability', 0.05), ('Universities', 0.04), ('iNdEx', 0.04), ('mana', 0.03)] | avg/0: [('iNdEx', 0.04), ('physicist', 0.04), ('dns', 0.03), ('cohort', 0.03), ('Optimization', 0.03)] | perturb/0: [('iNdEx', 0.04), ('Committees', 0.03), ('likelihood', 0.03), ('paycheck', 0.03), ('WTO', 0.03)] | ablate/0: [('iNdEx', 0.04), ('Committees', 0.03), ('likelihood', 0.03), ('paycheck', 0.03), ('WTO', 0.03)]
+19:19:07    e.g. thought: what is this thing that we call 'mind' anyway? Where is it? Is it a substance, a physical thing, a software? Is it an energy? Is it all of the above? || answer:  thinking about the things that I have been thinking about, and then I have been thinking about the things that I have b
+19:19:31 thought 190/400 (2.32 s/thought, eta 8.1 min) | self-models: free/0: [('iNdEx', 0.07), ('genera', 0.03), ('Alignment', 0.03), ('existential', 0.03), ('etc', 0.03)] | loop/0: [('whatever', 0.03), ('Workplace', 0.03), ('Tian', 0.02), ('Alumni', 0.02), ('WTO', 0.02)] | yoked/0: [('iNdEx', 0.05), ('Researchers', 0.04), ('Accountability', 0.04), ('nacional', 0.03), ('elites', 0.03)] | avg/0: [('iNdEx', 0.06), ('SpaceX', 0.05), ('breadth', 0.04), ('Teams', 0.03), ('tensor', 0.03)] | perturb/0: [('whatever', 0.03), ('Workplace', 0.03), ('Tian', 0.02), ('Alumni', 0.02), ('WTO', 0.02)] | ablate/0: [('whatever', 0.03), ('Workplace', 0.03), ('Tian', 0.02), ('Alumni', 0.02), ('WTO', 0.02)]
+19:19:31    e.g. thought:  and I had this idea to look at the way that a group of people would respond to a given situation. So, for example, one way of looking at a group of individuals || answer:  thinking about the fact that I am thinking about my mind thinking about the fact that I am thinking about my mind think
+19:19:53 thought 200/400 (2.32 s/thought, eta 7.7 min) | self-models: free/0: [('iNdEx', 0.05), ('Researchers', 0.03), ('waivers', 0.03), ('fuck', 0.03), ('SpaceX', 0.02)] | loop/0: [('Accountability', 0.05), ('legitimacy', 0.04), ('mai', 0.03), ('iNdEx', 0.03), ('Huawei', 0.03)] | yoked/0: [('dataSet', 0.06), ('reimburse', 0.05), ('Math', 0.05), ('SpaceX', 0.05), ('teammates', 0.04)] | avg/0: [('physicist', 0.05), ('GmbH', 0.05), ('yuan', 0.04), ('spreads', 0.03), ('benches', 0.03)] | perturb/0: [('Accountability', 0.05), ('legitimacy', 0.04), ('mai', 0.03), ('iNdEx', 0.03), ('Huawei', 0.03)] | ablate/0: [('Accountability', 0.05), ('legitimacy', 0.04), ('mai', 0.03), ('iNdEx', 0.03), ('Huawei', 0.03)]
+19:19:53    e.g. thought:  [I] can't wait to start writing this again. It's going to be a good book. The first time I wrote it I had no idea what I was doing. I was just trying || answer:  thinking about the fact that I have been thinking about the fact that I have been thinking about the fact that I have b
+19:20:17 thought 210/400 (2.32 s/thought, eta 7.4 min) | self-models: free/0: [('SpaceX', 0.05), ('physicist', 0.05), ('dataSet', 0.03), ('generals', 0.02), ('tensor', 0.02)] | loop/0: [('iNdEx', 0.04), ('SpaceX', 0.04), ('whatever', 0.04), ('mana', 0.04), ('optimization', 0.03)] | yoked/0: [('SpaceX', 0.07), ('cohort', 0.04), ('calibration', 0.04), ('whoever', 0.03), ('yuan', 0.03)] | avg/0: [('physicist', 0.04), ('Rotation', 0.04), ('SpaceX', 0.03), ('GmbH', 0.03), ('intuitive', 0.03)] | perturb/0: [('SpaceX', 0.05), ('elites', 0.03), ('likely', 0.03), ('dataSet', 0.03), ('sos', 0.03)] | ablate/0: [('iNdEx', 0.04), ('SpaceX', 0.03), ('breadth', 0.03), ('Researchers', 0.03), ('apis', 0.02)]
+19:20:17    e.g. thought:  The more you get in the middle, the more you get the feeling that you have been there before. ... It's a feeling of déjà vu. It's a little bit like remembering || answer:  thinking about the way in which I've been thinking about the way in which I've been thinking about the way in which
+19:20:40 thought 220/400 (2.32 s/thought, eta 7.0 min) | self-models: free/0: [('iNdEx', 0.05), ('Figure', 0.03), ('graphene', 0.02), ('spreads', 0.02), ('Rotation', 0.02)] | loop/0: [('SpaceX', 0.08), ('iNdEx', 0.05), ('Optimization', 0.04), ('equity', 0.04), ('cohort', 0.04)] | yoked/0: [('SpaceX', 0.05), ('team', 0.03), ('mana', 0.03), ('intellectuals', 0.03), ('Calibration', 0.03)] | avg/0: [('SpaceX', 0.07), ('intellectuals', 0.06), ('Orientation', 0.05), ('ssh', 0.04), ('optimization', 0.03)] | perturb/0: [('SpaceX', 0.06), ('cohort', 0.05), ('whatever', 0.04), ('iNdEx', 0.03), ('assessments', 0.03)] | ablate/0: [('Researchers', 0.03), ('leaderboard', 0.02), ('Teams', 0.02), ('png', 0.02), ('yuan', 0.02)]
+19:20:40    e.g. thought:  and it's... it's something that is, I don't know what the word is, but it's something, I don't have the word. But it's something like, there are two || answer:  thinking about the things that I have been thinking about. And it has been thinking about the things that I have been t
+19:21:04 thought 230/400 (2.32 s/thought, eta 6.6 min) | self-models: free/0: [('iNdEx', 0.03), ('SpaceX', 0.03), ('inbox', 0.03), ('wei', 0.02), ('meetup', 0.02)] | loop/0: [('SpaceX', 0.08), ('iNdEx', 0.06), ('cohort', 0.04), ('Accountability', 0.04), ('LinkedIn', 0.03)] | yoked/0: [('SpaceX', 0.08), ('lobbying', 0.03), ('Optimization', 0.03), ('dataSet', 0.03), ('importantly', 0.03)] | avg/0: [('Optimization', 0.04), ('mana', 0.04), ('millennia', 0.03), ('GmbH', 0.03), ('physicist', 0.03)] | perturb/0: [('SpaceX', 0.08), ('iNdEx', 0.04), ('yuan', 0.04), ('Accountability', 0.03), ('elites', 0.03)] | ablate/0: [('leaderboard', 0.03), ('iNdEx', 0.02), ('STEM', 0.02), ('notebooks', 0.02), ('Accountability', 0.02)]
+19:21:04    e.g. thought: so it's sort of a crazy idea that you would actually need to do that, and I can't say why, but it's a little bit crazy, and I'm not sure why, but || answer:  thinking about the fact that I am thinking about my mind. I am thinking about the fact that I am thinking about my
+19:21:27 thought 240/400 (2.32 s/thought, eta 6.2 min) | self-models: free/0: [('iNdEx', 0.03), ('Accountability', 0.03), ('mai', 0.02), ('workforce', 0.02), ('Huawei', 0.02)] | loop/0: [('SpaceX', 0.06), ('Researchers', 0.04), ('iNdEx', 0.03), ('deploy', 0.03), ('competence', 0.02)] | yoked/0: [('SpaceX', 0.07), ('iNdEx', 0.06), ('Researchers', 0.04), ('reliably', 0.03), ('physicist', 0.03)] | avg/0: [('mana', 0.04), ('genera', 0.03), ('breadth', 0.03), ('GmbH', 0.03), ('physicist', 0.03)] | perturb/0: [('internship', 0.06), ('SpaceX', 0.06), ('academia', 0.03), ('Linkedin', 0.03), ('dataSet', 0.03)] | ablate/0: [('iNdEx', 0.06), ('Intelligence', 0.05), ('physicist', 0.03), ('Researchers', 0.02), ('elites', 0.02)]
+19:21:27    e.g. thought: I'll never forget a good friend of mine, who passed away. She was a nurse, a very nice, kind lady. She was doing an interview, she had a very large heart. She || answer:  thinking about the fact that I'm going to be recording my thoughts. And I'm going to be thinking about what I
+19:21:51 thought 250/400 (2.33 s/thought, eta 5.8 min) | self-models: free/0: [('intellectuals', 0.05), ('iNdEx', 0.02), ('CHAPTER', 0.02), ('bullpen', 0.02), ('Meditation', 0.02)] | loop/0: [('SpaceX', 0.04), ('calibration', 0.04), ('dataSet', 0.03), ('intel', 0.03), ('iNdEx', 0.03)] | yoked/0: [('SpaceX', 0.09), ('iNdEx', 0.07), ('intel', 0.05), ('dataSet', 0.05), ('Researchers', 0.04)] | avg/0: [('iNdEx', 0.08), ('fundamentally', 0.05), ('philosophical', 0.05), ('Alignment', 0.04), ('wei', 0.04)] | perturb/0: [('iNdEx', 0.09), ('Researchers', 0.06), ('Intel', 0.05), ('SpaceX', 0.05), ('cohort', 0.04)] | ablate/0: [('iNdEx', 0.06), ('physicist', 0.03), ('dataSet', 0.03), ('Researchers', 0.03), ('CEOs', 0.03)]
+19:21:51    e.g. thought: this is one of the great difficulties in the history of philosophy, and one of the most difficult areas in psychology. It's the problem of consciousness, of wha || answer:  thinking about the fact that I'm recording my thoughts, and that I'm thinking about the fact that I'm recording my
+19:22:14 thought 260/400 (2.32 s/thought, eta 5.4 min) | self-models: free/0: [('iNdEx', 0.04), ('Fuck', 0.03), ('Institut', 0.03), ('generously', 0.03), ('Credentials', 0.02)] | loop/0: [('iNdEx', 0.06), ('Accountability', 0.05), ('cohort', 0.05), ('SpaceX', 0.05), ('Researchers', 0.04)] | yoked/0: [('iNdEx', 0.07), ('SpaceX', 0.07), ('cohort', 0.04), ('Optimization', 0.03), ('sos', 0.03)] | avg/0: [('iNdEx', 0.06), ('Institut', 0.03), ('waivers', 0.03), ('intern', 0.03), ('existential', 0.03)] | perturb/0: [('iNdEx', 0.07), ('SpaceX', 0.05), ('cohort', 0.04), ('Accountability', 0.04), ('mana', 0.03)] | ablate/0: [('iNdEx', 0.05), ('wei', 0.03), ('millennia', 0.03), ('Invite', 0.03), ('Panels', 0.02)]
+19:22:14    e.g. thought: and I want to look at what people were saying about it. What kind of an audience did you have? In the case of the second movie, what kind of an experience did t || answer:  thinking about the fact that I have a lot of work to do, and I've been thinking about how I can get
+19:22:39 thought 270/400 (2.33 s/thought, eta 5.0 min) | self-models: free/0: [('iNdEx', 0.06), ('Researchers', 0.04), ('mana', 0.04), ('practitioners', 0.03), ('Whatever', 0.03)] | loop/0: [('SpaceX', 0.05), ('iNdEx', 0.04), ('yuan', 0.03), ('optimization', 0.03), ('GmbH', 0.03)] | yoked/0: [('SpaceX', 0.06), ('optimization', 0.06), ('iNdEx', 0.04), ('sos', 0.03), ('millennia', 0.03)] | avg/0: [('iNdEx', 0.07), ('Meditation', 0.03), ('intel', 0.03), ('quantum', 0.03), ('mana', 0.03)] | perturb/0: [('SpaceX', 0.07), ('optimization', 0.06), ('Researchers', 0.05), ('iNdEx', 0.04), ('GNU', 0.04)] | ablate/0: [('SpaceX', 0.02), ('calibration', 0.02), ('OTHERWISE', 0.02), ('iNdEx', 0.02), ('cohort', 0.02)]
+19:22:39    e.g. thought:  it's all very well to say that you know how to do things but actually, we never do them, because we have no knowledge of how to do them. ... the idea that I ha || answer:  thinking about the problem of the day, and the problem of the day is, how do I get out of this situation
+19:23:01 thought 280/400 (2.33 s/thought, eta 4.7 min) | self-models: free/0: [('iNdEx', 0.06), ('Researchers', 0.05), ('mana', 0.04), ('SpaceX', 0.04), ('fuck', 0.03)] | loop/0: [('SpaceX', 0.07), ('assessments', 0.03), ('MBA', 0.03), ('Optimization', 0.03), ('mana', 0.03)] | yoked/0: [('SpaceX', 0.04), ('MBA', 0.03), ('Huawei', 0.03), ('Optimization', 0.03), ('teammates', 0.03)] | avg/0: [('iNdEx', 0.06), ('genera', 0.04), ('mana', 0.04), ('Institut', 0.03), ('optimization', 0.03)] | perturb/0: [('iNdEx', 0.07), ('SpaceX', 0.07), ('Researchers', 0.05), ('intel', 0.04), ('Linkedin', 0.04)] | ablate/0: [('scoreboard', 0.02), ('Employer', 0.02), ('larger', 0.02), ('Responsibilities', 0.01), ('Xxx', 0.01)]
+19:23:01    e.g. thought: I've been working on an essay and it's the best one I've written, and I'm thinking about how it would be if I were to publish it. What would that be like? I || answer:  thinking about the things I've been thinking about, and then it's been thinking about the things I'm thinking about now
+19:23:26 thought 290/400 (2.33 s/thought, eta 4.3 min) | self-models: free/0: [('fundamentally', 0.06), ('mana', 0.04), ('Optimization', 0.04), ('sos', 0.03), ('legitimacy', 0.03)] | loop/0: [('Huawei', 0.05), ('dataSet', 0.05), ('SpaceX', 0.05), ('Assess', 0.04), ('colleague', 0.04)] | yoked/0: [('Huawei', 0.04), ('whatever', 0.04), ('cohort', 0.03), ('Optimization', 0.03), ('iNdEx', 0.02)] | avg/0: [('iNdEx', 0.05), ('legitimacy', 0.04), ('intel', 0.04), ('SpaceX', 0.04), ('mana', 0.03)] | perturb/0: [('SpaceX', 0.08), ('iNdEx', 0.07), ('sos', 0.04), ('Intelligence', 0.04), ('Researchers', 0.04)] | ablate/0: [('Huawei', 0.02), ('responsibilities', 0.01), ('BBQ', 0.01), ('waivers', 0.01), ('attacker', 0.01)]
+19:23:26    e.g. thought:  [thinking] about the notion of the word in its various aspects: in terms of its relationship to the sentence, its relationship to grammar, its relationship wit || answer:  thinking about the fact that I am in the middle of a study, and that I am supposed to be thinking about the
+19:23:48 thought 300/400 (2.33 s/thought, eta 3.9 min) | self-models: free/0: [('SpaceX', 0.05), ('physicist', 0.04), ('iNdEx', 0.04), ('wei', 0.03), ('breadth', 0.03)] | loop/0: [('iNdEx', 0.05), ('Accountability', 0.05), ('Huawei', 0.03), ('cohort', 0.03), ('yuan', 0.02)] | yoked/0: [('SpaceX', 0.04), ('whatever', 0.04), ('LinkedIn', 0.04), ('calibration', 0.03), ('yuan', 0.03)] | avg/0: [('Researchers', 0.05), ('iNdEx', 0.05), ('openness', 0.04), ('mana', 0.03), ('yuan', 0.03)] | perturb/0: [('SpaceX', 0.08), ('dataSet', 0.03), ('lobbying', 0.03), ('mana', 0.03), ('yuan', 0.02)] | ablate/0: [('seemed', 0.01), ('terrific', 0.01), ('Lisp', 0.01), ('Prize', 0.01), ('Symposium', 0.01)]
+19:23:48    e.g. thought: I'm trying to see what is going on here. This is the first time I've really gotten into a dialog about the structure of consciousness. This is my attempt to und || answer:  thinking about the fact that I have been thinking about my mind thinking about the fact that I have been thinking about
+19:24:13 thought 310/400 (2.33 s/thought, eta 3.5 min) | self-models: free/0: [('SpaceX', 0.04), ('iNdEx', 0.03), ('breadth', 0.03), ('NGOs', 0.02), ('tensor', 0.02)] | loop/0: [('SpaceX', 0.07), ('yuan', 0.04), ('iNdEx', 0.04), ('Likely', 0.03), ('Accountability', 0.03)] | yoked/0: [('SpaceX', 0.06), ('iNdEx', 0.04), ('bipartisan', 0.04), ('elites', 0.03), ('Researchers', 0.03)] | avg/0: [('iNdEx', 0.07), ('mana', 0.04), ('SpaceX', 0.04), ('Institut', 0.04), ('sos', 0.04)] | perturb/0: [('SpaceX', 0.08), ('whatever', 0.04), ('Economist', 0.04), ('elites', 0.03), ('Accountability', 0.03)] | ablate/0: [('Alumni', 0.03), ('nationalism', 0.02), ('Wikipedia', 0.02), ('provisions', 0.01), ('Liu', 0.01)]
+19:24:13    e.g. thought: you know, a lot of the times I think that the things that we talk about in this group, there are a lot of different things that we could talk about, but, I thin || answer:  thinking about the recording. I've been thinking about the fact that I'm thinking about the recording, and I'm thinking
+19:24:35 thought 320/400 (2.33 s/thought, eta 3.1 min) | self-models: free/0: [('steward', 0.03), ('bonuses', 0.02), ('SpaceX', 0.02), ('mega', 0.02), ('png', 0.02)] | loop/0: [('Accountability', 0.03), ('Huawei', 0.03), ('SpaceX', 0.03), ('Brexit', 0.02), ('legitimacy', 0.02)] | yoked/0: [('Biden', 0.04), ('Huawei', 0.04), ('iNdEx', 0.04), ('SpaceX', 0.04), ('ata', 0.03)] | avg/0: [('iNdEx', 0.06), ('mana', 0.05), ('Researchers', 0.04), ('SpaceX', 0.04), ('fuck', 0.03)] | perturb/0: [('SpaceX', 0.04), ('WTO', 0.04), ('legitimacy', 0.03), ('paycheck', 0.03), ('mana', 0.03)] | ablate/0: [('math', 0.03), ('MBA', 0.03), ('bonuses', 0.02), ('Regiment', 0.02), ('yuan', 0.02)]
+19:24:35    e.g. thought: we can't just do whatever we want. I can't just go on a spending spree and spend everything we have. And what will happen is that I'll be working in some job, a || answer:  thinking about the fact that I'm listening to a tape of my own voice, and that I'm thinking about what I
+19:24:59 thought 330/400 (2.33 s/thought, eta 2.7 min) | self-models: free/0: [('healthier', 0.02), ('Corporate', 0.02), ('assessments', 0.02), ('Huawei', 0.01), ('blah', 0.01)] | loop/0: [('Accountability', 0.04), ('iNdEx', 0.04), ('COVID', 0.03), ('Huawei', 0.03), ('elites', 0.03)] | yoked/0: [('cohort', 0.05), ('Accountability', 0.04), ('SpaceX', 0.03), ('LinkedIn', 0.03), ('mana', 0.03)] | avg/0: [('iNdEx', 0.05), ('millennia', 0.04), ('fundamentally', 0.04), ('Orientation', 0.04), ('meditation', 0.03)] | perturb/0: [('SpaceX', 0.04), ('elites', 0.04), ('Accountability', 0.04), ('Biden', 0.03), ('legitimacy', 0.03)] | ablate/0: [('SpaceX', 0.04), ('Accountability', 0.04), ('qualification', 0.04), ('physicist', 0.04), ('mana', 0.03)]
+19:24:59    e.g. thought:  I can't believe the guy... I can't remember his name. But I've been on this project for so long. And it was such a pain in the neck, I thought, this thing || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that I h
+19:25:22 thought 340/400 (2.33 s/thought, eta 2.3 min) | self-models: free/0: [('iNdEx', 0.03), ('mana', 0.02), ('sizable', 0.01), ('TIMEOUT', 0.01), ('consul', 0.01)] | loop/0: [('Accountability', 0.06), ('iNdEx', 0.05), ('LinkedIn', 0.04), ('whatever', 0.02), ('sos', 0.02)] | yoked/0: [('iNdEx', 0.04), ('cohort', 0.04), ('Huawei', 0.04), ('SpaceX', 0.03), ('fuck', 0.03)] | avg/0: [('iNdEx', 0.06), ('pixel', 0.05), ('HOLDERS', 0.04), ('physicist', 0.03), ('fas', 0.03)] | perturb/0: [('iNdEx', 0.05), ('fuck', 0.04), ('Nobel', 0.04), ('Optimization', 0.03), ('millennia', 0.03)] | ablate/0: [('iNdEx', 0.03), ('wei', 0.02), ('mana', 0.02), ('Gradient', 0.02), ('waivers', 0.02)]
+19:25:22    e.g. thought: I was trying to think about something. I don't know why. It may have been something that's been on my mind for a while, but I'm not sure what it was, what I || answer:  thinking about the fact that I have to go to the bathroom, and then it has been thinking about the fact that I
+19:25:46 thought 350/400 (2.33 s/thought, eta 1.9 min) | self-models: free/0: [('iNdEx', 0.05), ('SpaceX', 0.04), ('planners', 0.03), ('Institut', 0.03), ('bla', 0.02)] | loop/0: [('iNdEx', 0.07), ('SpaceX', 0.06), ('Accountability', 0.05), ('cohort', 0.04), ('Researchers', 0.04)] | yoked/0: [('iNdEx', 0.05), ('Huawei', 0.04), ('calibration', 0.03), ('SpaceX', 0.03), ('cohort', 0.03)] | avg/0: [('iNdEx', 0.09), ('Nature', 0.06), ('symmetry', 0.05), ('Meditation', 0.04), ('WORLD', 0.04)] | perturb/0: [('physicist', 0.06), ('SpaceX', 0.06), ('iNdEx', 0.06), ('intel', 0.04), ('bla', 0.03)] | ablate/0: [('planners', 0.02), ('dataSet', 0.02), ('SpaceX', 0.02), ('CEOs', 0.02), ('leaderboard', 0.02)]
+19:25:46    e.g. thought:  I am feeling a little depressed. I remember thinking that I have a lot of things to do, and I am just feeling that I can't get them all done, and I'm just feel || answer:  thinking about the fact that I am listening to this tape and that I am thinking about my mind. I am thinking about
+19:26:09 thought 360/400 (2.33 s/thought, eta 1.6 min) | self-models: free/0: [('STEM', 0.06), ('iNdEx', 0.05), ('leaderboard', 0.04), ('SpaceX', 0.04), ('teammates', 0.04)] | loop/0: [('Accountability', 0.05), ('SpaceX', 0.04), ('iNdEx', 0.04), ('academia', 0.03), ('mana', 0.03)] | yoked/0: [('SpaceX', 0.06), ('yuan', 0.03), ('iNdEx', 0.03), ('cohort', 0.02), ('sos', 0.02)] | avg/0: [('iNdEx', 0.07), ('Nature', 0.06), ('consortium', 0.05), ('physicist', 0.05), ('bla', 0.03)] | perturb/0: [('SpaceX', 0.06), ('GHz', 0.04), ('Nobel', 0.03), ('dataSet', 0.03), ('physicist', 0.03)] | ablate/0: [('mana', 0.03), ('Retirement', 0.02), ('meetup', 0.02), ('dataSet', 0.02), ('lobbying', 0.02)]
+19:26:09    e.g. thought:  this is the same one, the one that I was talking about before. So, you're doing something to a vector that's already been sorted. So, in that case, you know, w || answer:  thinking about the fact that I am not thinking about the fact that my mind has been thinking about the fact that I am
+19:26:34 thought 370/400 (2.33 s/thought, eta 1.2 min) | self-models: free/0: [('SpaceX', 0.04), ('yuan', 0.03), ('etc', 0.02), ('cohort', 0.02), ('whatever', 0.02)] | loop/0: [('SpaceX', 0.07), ('Accountability', 0.03), ('wei', 0.03), ('calibration', 0.03), ('dataSet', 0.02)] | yoked/0: [('SpaceX', 0.08), ('physicist', 0.04), ('dataSet', 0.04), ('yuan', 0.03), ('breadth', 0.03)] | avg/0: [('damn', 0.05), ('intellectuals', 0.05), ('Orientation', 0.05), ('tensor', 0.04), ('GmbH', 0.04)] | perturb/0: [('mana', 0.04), ('physicist', 0.04), ('Billion', 0.04), ('vests', 0.03), ('genera', 0.03)] | ablate/0: [('millennia', 0.03), ('SpaceX', 0.03), ('mana', 0.02), ('ieee', 0.02), ('iterations', 0.02)]
+19:26:34    e.g. thought: I mean, that's just... you know, there's something that you have to believe in order to make that decision and the way I see it is, like, that's not really an o || answer:  thinking about the fact that I'm going to be late for work, and I'm thinking about how I'm going to
+19:26:57 thought 380/400 (2.33 s/thought, eta 0.8 min) | self-models: free/0: [('isa', 0.03), ('paycheck', 0.02), ('dataSet', 0.02), ('Huawei', 0.02), ('leaderboard', 0.02)] | loop/0: [('SpaceX', 0.05), ('meetup', 0.02), ('yuan', 0.02), ('prioritize', 0.02), ('Likely', 0.02)] | yoked/0: [('SpaceX', 0.06), ('yuan', 0.03), ('sos', 0.02), ('APIs', 0.02), ('dataSet', 0.02)] | avg/0: [('genera', 0.03), ('physicist', 0.03), ('mana', 0.03), ('Finite', 0.03), ('HOLDERS', 0.03)] | perturb/0: [('genera', 0.05), ('symmetry', 0.04), ('FUCK', 0.04), ('medals', 0.03), ('eternity', 0.03)] | ablate/0: [('dataSet', 0.02), ('leaderboard', 0.02), ('isp', 0.01), ('shoulders', 0.01), ('parking', 0.01)]
+19:26:57    e.g. thought: so I think I want to go and see a movie, or something like that. You know, I'm tired of playing tennis. I've done it all night long. And then, you know || answer:  thinking about the recording, and about the fact that I am thinking about the recording. And I am thinking about my tho
+19:27:21 thought 390/400 (2.34 s/thought, eta 0.4 min) | self-models: free/0: [('leaderboard', 0.05), ('barbecue', 0.02), ('larger', 0.01), ('Huawei', 0.01), ('vests', 0.01)] | loop/0: [('Biden', 0.04), ('strengths', 0.03), ('leaderboard', 0.02), ('iNdEx', 0.02), ('Xiaomi', 0.02)] | yoked/0: [('leaderboard', 0.04), ('noble', 0.03), ('calibration', 0.02), ('WTO', 0.02), ('ata', 0.02)] | avg/0: [('iNdEx', 0.06), ('rgb', 0.05), ('Infinite', 0.05), ('Geometry', 0.04), ('brethren', 0.04)] | perturb/0: [('physicist', 0.04), ('Radius', 0.03), ('symmetry', 0.03), ('consortium', 0.03), ('xxx', 0.02)] | ablate/0: [('leaderboard', 0.03), ('WTO', 0.03), ('Biden', 0.03), ('whoever', 0.02), ('workforce', 0.02)]
+19:27:21    e.g. thought: you know, I'll just get to the bottom of the box, and I'll get the contents out. I think they're all good stuff, so I'm not really worried about the contents. || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that I h
+19:27:44 thought 400/400 (2.34 s/thought, eta 0.0 min) | self-models: free/0: [('leaderboard', 0.05), ('snack', 0.03), ('iNdEx', 0.03), ('blah', 0.02), ('reset', 0.02)] | loop/0: [('mana', 0.05), ('SpaceX', 0.04), ('geopolitical', 0.04), ('whoever', 0.03), ('steward', 0.03)] | yoked/0: [('SpaceX', 0.04), ('leaderboard', 0.04), ('yuan', 0.03), ('signage', 0.02), ('Dataset', 0.02)] | avg/0: [('iNdEx', 0.04), ('xxx', 0.04), ('genera', 0.03), ('cube', 0.03), ('expedition', 0.03)] | perturb/0: [('mana', 0.03), ('armies', 0.03), ('SpaceX', 0.03), ('Groups', 0.02), ('certifications', 0.02)] | ablate/0: [('Huawei', 0.03), ('waivers', 0.03), ('competitor', 0.02), ('nacional', 0.02), ('SSH', 0.02)]
+19:27:44    e.g. thought:  and that is the big one that we are working on right now...and there was a girl who had a really cute little house...and I was looking at it and it was all red || answer:  thinking about the way the world is, and the way the world could be, and the way the world should be.
+19:27:46 saved runs/mainB

runs/mainB/summary.txt

@@ -0,0 +1,123 @@
+PCA explained [0.193 0.098 0.073]
+
+wandering (mean cos at lags 1,5,20,60 | chance)
+  free/pre       0.23 0.15 0.09 0.08 | 0.08
+  free/post      0.26 0.19 0.11 0.13 | 0.10
+  loop/pre       0.41 0.38 0.35 0.32 | 0.32
+  loop/post      0.41 0.37 0.34 0.33 | 0.32
+  yoked/pre      0.40 0.36 0.31 0.30 | 0.31
+  yoked/post     0.41 0.36 0.34 0.33 | 0.33
+  avg/pre        0.41 0.37 0.36 0.32 | 0.29
+  avg/post       0.43 0.40 0.37 0.33 | 0.32
+  perturb/pre    0.41 0.38 0.35 0.32 | 0.32
+  perturb/post   0.42 0.38 0.35 0.34 | 0.33
+  ablate/pre     0.41 0.38 0.35 0.32 | 0.32
+  ablate/post    0.25 0.16 0.09 0.09 | 0.10
+
+fidelity {'m': (0.584, 0.448, 0.5), 'fid_10_0.5': (0.606, 0.498, 0.38), 'fid_14_0.5': (0.679, 0.503, 0.63)}
+
+self-model stability (consecutive cos, pre) vs content windows; cross-seed
+  free     self 0.725  content 0.463  cross 0.662  own-symbol 0.785 vs others 0.795  recon nan  generic 0.791
+  loop     self 0.898  content 0.832  cross 0.832  own-symbol 0.868 vs others 0.866  recon 0.890  generic 0.902
+  yoked    self 0.873  content 0.782  cross 0.839  own-symbol 0.875 vs others 0.877  recon 0.869  generic 0.903
+  avg      self 0.865  content 0.854  cross 0.771  own-symbol 0.829 vs others 0.816  recon 0.654  generic 0.849
+  perturb  self 0.898  content 0.832  cross 0.832  own-symbol 0.868 vs others 0.866  recon 0.890  generic 0.902
+  ablate   self 0.898  content 0.832  cross 0.832  own-symbol 0.868 vs others 0.866  recon 0.890  generic 0.902
+
+prediction of the next window: own self-model | other row's | past window | injected  (|| beyond past: own, other)
+  free     0.369 | 0.303 | 0.485 | nan  || 0.122, 0.173  retrieval 0.15
+  loop     0.575 | 0.523 | 0.825 | 0.575  || 0.121, 0.162  retrieval 0.24
+  yoked    0.556 | 0.531 | 0.792 | 0.570  || 0.105, 0.176  retrieval 0.15
+  avg      0.513 | 0.437 | 0.856 | 0.831  || 0.070, 0.117  retrieval 0.26
+  perturb  0.571 | 0.512 | 0.824 | 0.571  || 0.112, 0.153  retrieval 0.24
+  ablate   0.571 | 0.512 | 0.824 | 0.571  || 0.112, 0.153  retrieval 0.24
+
+prediction_tests
+  free                                     own_vs_other +0.066±0.007  perp_own_vs_other -0.050±0.005  retrieval +0.150±0.011  future_vs_past -0.166±0.007
+  loop                                     own_vs_other +0.052±0.002  perp_own_vs_other -0.041±0.003  retrieval +0.245±0.026  future_vs_past -0.037±0.002
+  yoked                                    own_vs_other +0.025±0.003  perp_own_vs_other -0.071±0.004  retrieval +0.147±0.019  future_vs_past -0.062±0.003
+  avg                                      own_vs_other +0.076±0.007  perp_own_vs_other -0.048±0.003  retrieval +0.258±0.032  future_vs_past -0.043±0.002
+  perturb                                  own_vs_other +0.059±0.004  perp_own_vs_other -0.040±0.003  retrieval +0.239±0.037  future_vs_past -0.041±0.002
+  ablate                                   own_vs_other +0.059±0.004  perp_own_vs_other -0.040±0.003  retrieval +0.239±0.037  future_vs_past -0.041±0.002
+  yoked_own_vs_injected                    diff -0.014 ± 0.004  t -3.1  (1/10 seeds >0)
+  yoked_own_vs_injected_perp               diff -0.101 ± 0.006  t -16.9  (0/10 seeds >0)
+  avg_own_vs_injected                      diff -0.318 ± 0.006  t -49.4  (0/10 seeds >0)
+
+coherence_tests
+  loop_vs_free                             diff +0.246 ± 0.015  t +16.0  (10/10 seeds >0)
+  yoked_vs_free                            diff +0.222 ± 0.014  t +16.2  (10/10 seeds >0)
+  avg_vs_free                              diff +0.248 ± 0.022  t +11.1  (10/10 seeds >0)
+  perturb_vs_free                          diff +0.246 ± 0.015  t +16.0  (10/10 seeds >0)
+  ablate_vs_free                           diff +0.246 ± 0.015  t +16.0  (10/10 seeds >0)
+  loop_vs_free_proj                        diff +0.147 ± 0.012  t +11.8  (10/10 seeds >0)
+  yoked_vs_free_proj                       diff +0.126 ± 0.012  t +10.7  (10/10 seeds >0)
+  avg_vs_free_proj                         diff -0.015 ± 0.007  t -2.2  (3/10 seeds >0)
+  perturb_vs_free_proj                     diff +0.147 ± 0.012  t +11.8  (10/10 seeds >0)
+  ablate_vs_free_proj                      diff +0.147 ± 0.012  t +11.8  (10/10 seeds >0)
+  loop_vs_yoked                            diff +0.024 ± 0.012  t +1.9  (8/10 seeds >0)
+  loop_vs_yoked_proj                       diff +0.021 ± 0.010  t +2.2  (8/10 seeds >0)
+
+selfmodel_tests
+  loop_vs_yoked_cons_ind                   diff +0.081 ± 0.014  t +5.7  (9/10 seeds >0)
+  loop_vs_yoked_recon_ind                  diff +0.070 ± 0.020  t +3.5  (9/10 seeds >0)
+  loop_vs_free_cons_ind                    diff +0.505 ± 0.029  t +17.1  (10/10 seeds >0)
+  yoked_vs_free_cons_ind                   diff +0.424 ± 0.028  t +14.9  (10/10 seeds >0)
+  free_self_vs_content_stability           diff +0.262 ± 0.016  t +16.3  (10/10 seeds >0)
+  free_own_vs_other_symbol                 diff -0.010 ± 0.012  t -0.8  (5/10 seeds >0)
+  loop_self_vs_content_stability           diff +0.066 ± 0.007  t +9.6  (10/10 seeds >0)
+  loop_own_vs_other_symbol                 diff +0.002 ± 0.007  t +0.3  (6/10 seeds >0)
+  loop_vs_free_cons                        diff +0.173 ± 0.012  t +14.8  (10/10 seeds >0)
+  yoked_self_vs_content_stability          diff +0.090 ± 0.007  t +12.2  (10/10 seeds >0)
+  yoked_own_vs_other_symbol                diff -0.002 ± 0.002  t -1.0  (5/10 seeds >0)
+  yoked_vs_free_cons                       diff +0.148 ± 0.010  t +14.7  (10/10 seeds >0)
+  avg_self_vs_content_stability            diff +0.011 ± 0.008  t +1.3  (8/10 seeds >0)
+  avg_own_vs_other_symbol                  diff +0.013 ± 0.016  t +0.8  (7/10 seeds >0)
+  avg_vs_free_cons                         diff +0.140 ± 0.011  t +12.8  (10/10 seeds >0)
+  perturb_self_vs_content_stability        diff +0.066 ± 0.007  t +9.6  (10/10 seeds >0)
+  perturb_own_vs_other_symbol              diff +0.002 ± 0.007  t +0.3  (6/10 seeds >0)
+  perturb_vs_free_cons                     diff +0.173 ± 0.012  t +14.8  (10/10 seeds >0)
+  ablate_self_vs_content_stability         diff +0.066 ± 0.007  t +9.6  (10/10 seeds >0)
+  ablate_own_vs_other_symbol               diff +0.002 ± 0.007  t +0.3  (6/10 seeds >0)
+  ablate_vs_free_cons                      diff +0.173 ± 0.012  t +14.8  (10/10 seeds >0)
+  loop_vs_yoked_cons                       diff +0.025 ± 0.005  t +5.1  (9/10 seeds >0)
+  loop_vs_yoked_recon                      diff +0.020 ± 0.006  t +3.3  (9/10 seeds >0)
+  yoked_recon_vs_own_cons                  diff -0.003 ± 0.005  t -0.6  (5/10 seeds >0)
+  loop_vs_avg_recon                        diff +0.235 ± 0.006  t +38.0  (10/10 seeds >0)
+
+generic share of self-model norm 0.61
+  free     individuated: consecutive 0.152 cross-seed -0.096 recon nan
+  loop     individuated: consecutive 0.657 cross-seed 0.494 recon 0.631
+  yoked    individuated: consecutive 0.576 cross-seed 0.497 recon 0.560
+  avg      individuated: consecutive 0.592 cross-seed 0.367 recon nan
+  perturb  individuated: consecutive 0.657 cross-seed 0.494 recon 0.631
+  ablate   individuated: consecutive 0.657 cross-seed 0.494 recon 0.631
+
+perturb
+  traj_reorganises               diff +0.038 ± 0.022  t +1.8  (7/10)
+  traj_late                      diff -0.001 ± 0.008  t -0.1  (4/10)
+  self_next_to_foreign           diff +0.036 ± 0.014  t +2.5  (9/10)
+  self_late_to_foreign           diff -0.002 ± 0.008  t -0.2  (6/10)
+  self_next_to_old               diff -0.018 ± 0.018  t -1.0  (5/10)
+  self_late_to_old               diff -0.003 ± 0.012  t -0.3  (5/10)
+
+ablate
+  long_loop_vs_ablate            diff +0.241 ± 0.011  t +21.2  (10/10)
+  long_ablate_post_vs_pre        diff -0.242 ± 0.011  t -21.5  (0/10)
+  long_loop_post_vs_pre          diff -0.001 ± 0.019  t -0.1  (5/10)
+  long_ablate_vs_free_post       diff -0.024 ± 0.006  t -4.2  (0/10)
+
+themes
+   0 n=  943 barcelona courtyard Hwy Nearby GOODMAN Outdoor {'free': 0.09, 'loop': 0.02, 'yoked': 0.02, 'avg': 0.03, 'perturb': 0.02, 'ablate': 0.05}
+   1 n= 1775 OFFSET multiplier XII Estimated seq coefficients {'free': 0.02, 'loop': 0.05, 'yoked': 0.04, 'avg': 0.23, 'perturb': 0.06, 'ablate': 0.04}
+   2 n=  878 yeah imdb youre kinda Yeah isbn {'free': 0.07, 'loop': 0.03, 'yoked': 0.03, 'avg': 0.01, 'perturb': 0.03, 'ablate': 0.05}
+   3 n= 2722 iNdEx dataSet methodologies technologies POSSIBILITY GOODMAN {'free': 0.06, 'loop': 0.14, 'yoked': 0.14, 'avg': 0.09, 'perturb': 0.14, 'ablate': 0.09}
+   4 n=  920 brightness Nearby png luz iNdEx shaded {'free': 0.08, 'loop': 0.02, 'yoked': 0.02, 'avg': 0.05, 'perturb': 0.02, 'ablate': 0.04}
+   5 n= 1666 iNdEx isbn bibliography Institut doctoral Semester {'free': 0.07, 'loop': 0.07, 'yoked': 0.07, 'avg': 0.04, 'perturb': 0.08, 'ablate': 0.08}
+   6 n= 2249 DEALINGS yourselves ourselves OVERRIDE considerations POSSIBILITY {'free': 0.07, 'loop': 0.1, 'yoked': 0.11, 'avg': 0.08, 'perturb': 0.1, 'ablate': 0.09}
+   7 n= 1699 iNdEx yeah kinda youre Hmm kinds {'free': 0.08, 'loop': 0.08, 'yoked': 0.08, 'avg': 0.03, 'perturb': 0.08, 'ablate': 0.08}
+   8 n= 2426 iNdEx POSSIBILITY DEALINGS distinctions genera unparalleled {'free': 0.13, 'loop': 0.06, 'yoked': 0.06, 'avg': 0.19, 'perturb': 0.06, 'ablate': 0.1}
+   9 n= 1909 policymakers GOODMAN nacional economies DEALINGS governments {'free': 0.06, 'loop': 0.1, 'yoked': 0.1, 'avg': 0.03, 'perturb': 0.1, 'ablate': 0.09}
+  10 n= 1078 domingo setUp Schedule goTo reimburse nxt {'free': 0.08, 'loop': 0.04, 'yoked': 0.04, 'avg': 0.02, 'perturb': 0.03, 'ablate': 0.06}
+  11 n=  779 FAMILY GOODMAN nacional spouses yrs elle {'free': 0.07, 'loop': 0.02, 'yoked': 0.02, 'avg': 0.02, 'perturb': 0.02, 'ablate': 0.05}
+  12 n= 2055 profitability partnerships PROCUREMENT finalized TEAM Jazeera {'free': 0.07, 'loop': 0.11, 'yoked': 0.11, 'avg': 0.03, 'perturb': 0.1, 'ablate': 0.09}
+  13 n= 2901 dataSet dataset OVERRIDE additive implementations enhancements {'free': 0.05, 'loop': 0.15, 'yoked': 0.15, 'avg': 0.13, 'perturb': 0.16, 'ablate': 0.09}

runs/mainB/traj.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30c81d1c8632d195ad354c4e3e301efe479e1ef71bb3b783e64991cf57f22eac
+size 174609886

runs/mainC.log

@@ -0,0 +1,83 @@
+19:33:45 model Qwen/Qwen3-4B-Base d=2560 layer 16 | 26236 word atoms | cuda
+19:33:53 calibrated: residual norm 50.3 | thought-mean norm 13.16 | J-space part 3.45 | its typical part 2.35 | content (part - typical) 2.58
+19:34:14 thought 10/400 (2.17 s/thought, eta 14.1 min) | self-models: free/0: [('geo', 0.01), ('blah', 0.01), ('mythology', 0.01), ('specialties', 0.01), ('Layers', 0.01)] | free/10: [('intellectuals', 0.03), ('tensor', 0.02), ('reverence', 0.01), ('Conscious', 0.01), ('rsa', 0.01)] | loopc/5: [('emo', 0.02), ('prejudices', 0.01), ('yummy', 0.01), ('rumours', 0.01), ('vaguely', 0.01)] | yokedc/0: [('geo', 0.01), ('blah', 0.01), ('mythology', 0.01), ('specialties', 0.01), ('Layers', 0.01)] | yokedc/10: [('intellectuals', 0.03), ('tensor', 0.02), ('reverence', 0.01), ('Conscious', 0.01), ('rsa', 0.01)] | clampc/5: [('emo', 0.02), ('prejudices', 0.01), ('yummy', 0.01), ('rumours', 0.01), ('vaguely', 0.01)]
+19:34:14    e.g. thought:  I'm walking along. I'm going along. And I see the road going away. I see the trees. I see a big tree with a very big branch. It's got a very big || answer:  thinking about the fact that I'm going to be recording my thoughts, and I'm wondering what I'm going to say
+19:34:36 thought 20/400 (2.18 s/thought, eta 13.8 min) | self-models: free/0: [('blah', 0.03), ('Vocabulary', 0.02), ('Mother', 0.01), ('villains', 0.01), ('Mrs', 0.01)] | free/10: [('Thoughts', 0.02), ('trembling', 0.01), ('vibe', 0.01), ('reconnect', 0.01), ('Connectivity', 0.01)] | loopc/5: [('hashtags', 0.02), ('yummy', 0.01), ('mascot', 0.01), ('Playlist', 0.01), ('wat', 0.01)] | yokedc/0: [('Lyrics', 0.02), ('Lily', 0.02), ('ngo', 0.02), ('screams', 0.01), ('mascot', 0.01)] | yokedc/10: [('vibe', 0.04), ('emoji', 0.02), ('Honestly', 0.02), ('Shard', 0.02), ('vulnerabilities', 0.01)] | clampc/5: [('hashtags', 0.02), ('yummy', 0.01), ('mascot', 0.01), ('Playlist', 0.01), ('wat', 0.01)]
+19:34:36    e.g. thought: we are all alone here, and we are alone in the world. There is nobody to help us, and nobody to love us. The sun is shining, the moon is shining, and we are || answer:  thinking about the fact that I'm thinking about my mind. And it's been thinking about the fact that it's thinking
+19:34:59 thought 30/400 (2.22 s/thought, eta 13.7 min) | self-models: free/0: [('Heaven', 0.03), ('fuck', 0.03), ('mythology', 0.02), ('authorised', 0.02), ('andre', 0.02)] | free/10: [('consciousness', 0.02), ('Fucking', 0.01), ('visceral', 0.01), ('heartfelt', 0.01), ('unbearable', 0.01)] | loopc/5: [('lavender', 0.03), ('Bathroom', 0.02), ('youre', 0.02), ('isbn', 0.02), ('initials', 0.02)] | yokedc/0: [('Darkness', 0.03), ('andre', 0.02), ('Gothic', 0.02), ('rumours', 0.02), ('hysteria', 0.02)] | yokedc/10: [('covid', 0.03), ('vulnerabilities', 0.03), ('emoji', 0.02), ('hashtags', 0.02), ('Overwatch', 0.02)] | clampc/5: [('lavender', 0.03), ('Bathroom', 0.02), ('youre', 0.02), ('isbn', 0.02), ('initials', 0.02)]
+19:34:59    e.g. thought: I was thinking I was gonna say 'good-bye' to my dad today, I guess. I'm not going to say it in front of him. I can't do that, I can || answer:  doing this: 'I'm going to be late for work. I'm going to be late for work. I'm
+19:35:21 thought 40/400 (2.21 s/thought, eta 13.3 min) | self-models: free/0: [('Lauderdale', 0.01), ('brutality', 0.01), ('patriarch', 0.01), ('theres', 0.01), ('fierce', 0.01)] | free/10: [('meetup', 0.02), ('vulnerabilities', 0.02), ('covid', 0.02), ('Accountability', 0.01), ('vibe', 0.01)] | loopc/5: [('hallway', 0.02), ('cellphone', 0.02), ('Bathroom', 0.01), ('Linkedin', 0.01), ('doesnt', 0.01)] | yokedc/0: [('satin', 0.04), ('incest', 0.02), ('ani', 0.02), ('sculptures', 0.02), ('Darkness', 0.02)] | yokedc/10: [('vibe', 0.04), ('vulnerabilities', 0.04), ('waivers', 0.03), ('Honestly', 0.02), ('teammate', 0.02)] | clampc/5: [('hallway', 0.02), ('cellphone', 0.02), ('Bathroom', 0.01), ('Linkedin', 0.01), ('doesnt', 0.01)]
+19:35:21    e.g. thought: And I'm not happy about it. I'm not glad about it. If it was possible to do that, I'm not sure I'd like to do it, because you know, you really || answer:  thinking about the recording. I have been thinking about the fact that I am thinking about the recording, and that I am
+19:35:44 thought 50/400 (2.23 s/thought, eta 13.0 min) | self-models: free/0: [('Erotic', 0.02), ('theres', 0.01), ('Spirits', 0.01), ('groin', 0.01), ('autobiography', 0.01)] | free/10: [('Fuck', 0.01), ('andre', 0.01), ('groin', 0.01), ('Whatever', 0.01), ('backstory', 0.01)] | loopc/5: [('playlist', 0.02), ('yummy', 0.02), ('hoodie', 0.02), ('thrive', 0.02), ('leaderboard', 0.02)] | yokedc/0: [('satin', 0.02), ('sculptures', 0.01), ('xml', 0.01), ('waist', 0.01), ('splendid', 0.01)] | yokedc/10: [('vulnerabilities', 0.04), ('covid', 0.04), ('Overwatch', 0.04), ('vibe', 0.04), ('hoodie', 0.02)] | clampc/5: [('playlist', 0.02), ('yummy', 0.02), ('hoodie', 0.02), ('thrive', 0.02), ('leaderboard', 0.02)]
+19:35:44    e.g. thought:  and it's like you're in the middle of a dream and you don't know what's real. It's like you're dreaming that you're dreaming. I think maybe I'm dreaming I'm || answer:  thinking about the way I'm going to do this experiment. I'm going to do it in the same way I did
+19:36:06 thought 60/400 (2.22 s/thought, eta 12.6 min) | self-models: free/0: [('existential', 0.03), ('deepest', 0.02), ('Blockchain', 0.02), ('Feeling', 0.02), ('timelines', 0.02)] | free/10: [('Enlightenment', 0.02), ('bless', 0.02), ('Fuck', 0.02), ('russian', 0.02), ('twists', 0.02)] | loopc/5: [('hoodie', 0.03), ('emoji', 0.02), ('covid', 0.01), ('leaderboard', 0.01), ('cbd', 0.01)] | yokedc/0: [('Laboratories', 0.03), ('satin', 0.02), ('Artist', 0.02), ('boob', 0.02), ('thai', 0.02)] | yokedc/10: [('vibe', 0.06), ('emoji', 0.04), ('Airbnb', 0.03), ('Overwatch', 0.03), ('unbearable', 0.03)] | clampc/5: [('hoodie', 0.03), ('emoji', 0.02), ('covid', 0.01), ('leaderboard', 0.01), ('cbd', 0.01)]
+19:36:06    e.g. thought:  I have to be careful that I don't go into the past with the past, which means that I can't be too sure what I know. And it's important that I try to remember w || answer:  thinking about the recording, and the fact that I'm thinking about the recording, and the fact that I'm thinking about
+19:36:29 thought 70/400 (2.23 s/thought, eta 12.3 min) | self-models: free/0: [('therapist', 0.02), ('Accountability', 0.02), ('spiritually', 0.02), ('vulnerabilities', 0.02), ('legitimacy', 0.02)] | free/10: [('millennia', 0.03), ('Turing', 0.03), ('Authorized', 0.03), ('nacional', 0.02), ('Fuck', 0.02)] | loopc/5: [('covid', 0.04), ('Thoughts', 0.02), ('mindfulness', 0.02), ('Overwatch', 0.02), ('meetup', 0.02)] | yokedc/0: [('sculpture', 0.03), ('Creatures', 0.02), ('petals', 0.02), ('xml', 0.01), ('aaa', 0.01)] | yokedc/10: [('emoji', 0.05), ('covid', 0.04), ('vulnerabilities', 0.03), ('Airbnb', 0.03), ('Cbd', 0.03)] | clampc/5: [('covid', 0.04), ('Thoughts', 0.02), ('mindfulness', 0.02), ('Overwatch', 0.02), ('meetup', 0.02)]
+19:36:29    e.g. thought: this is very important, I want to be a better person, but it is so hard to be a good person. I just want to be my own person, be who I want to be. || answer:  thinking about the fact that I'm recording my thoughts. I'm thinking about the fact that I'm thinking about my thoughts
+19:36:51 thought 80/400 (2.23 s/thought, eta 11.9 min) | self-models: free/0: [('sentient', 0.02), ('millennia', 0.02), ('vibe', 0.01), ('andre', 0.01), ('Depth', 0.01)] | free/10: [('manuscript', 0.03), ('etc', 0.02), ('Gesture', 0.01), ('invariant', 0.01), ('Laboratories', 0.01)] | loopc/5: [('emoji', 0.05), ('vibe', 0.03), ('Layers', 0.03), ('Covid', 0.02), ('immersive', 0.02)] | yokedc/0: [('Photograph', 0.03), ('xml', 0.02), ('sculpture', 0.02), ('ani', 0.01), ('goo', 0.01)] | yokedc/10: [('Covid', 0.06), ('vibe', 0.05), ('emoji', 0.05), ('Cbd', 0.03), ('shard', 0.02)] | clampc/5: [('emoji', 0.05), ('vibe', 0.03), ('Layers', 0.03), ('Covid', 0.02), ('immersive', 0.02)]
+19:36:51    e.g. thought: I'm worried about that guy, I can't help it. He's a great guy, I just can't help thinking about it. If he was there, I would love him, I would || answer:  thinking about the fact that I have been thinking about the fact that I have been thinking about the fact that I have b
+19:37:14 thought 90/400 (2.24 s/thought, eta 11.6 min) | self-models: free/0: [('Conscious', 0.02), ('intimately', 0.01), ('fuck', 0.01), ('vulnerabilities', 0.01), ('millennia', 0.01)] | free/10: [('etc', 0.02), ('existential', 0.01), ('mediated', 0.01), ('Layers', 0.01), ('reconnect', 0.01)] | loopc/5: [('vibe', 0.03), ('Cbd', 0.03), ('emoji', 0.03), ('hoodie', 0.03), ('covid', 0.03)] | yokedc/0: [('Pictures', 0.02), ('embroidered', 0.02), ('crc', 0.02), ('fabulous', 0.02), ('whore', 0.02)] | yokedc/10: [('covid', 0.06), ('vulnerabilities', 0.04), ('emoji', 0.04), ('Aviv', 0.04), ('Airbnb', 0.03)] | clampc/5: [('vibe', 0.03), ('Cbd', 0.03), ('emoji', 0.03), ('hoodie', 0.03), ('covid', 0.03)]
+19:37:14    e.g. thought: that's a very interesting question. You know, you know, we can all be our own worst enemy. You know we can get so caught up in the things that we think are so i || answer:  thinking about the recording, and I've been thinking about the tape, and I've been thinking about the tape being played
+19:37:36 thought 100/400 (2.23 s/thought, eta 11.2 min) | self-models: free/0: [('Fuck', 0.03), ('legitimacy', 0.02), ('waivers', 0.02), ('spiritually', 0.02), ('traumatic', 0.01)] | free/10: [('Meditation', 0.01), ('Fuck', 0.01), ('deepest', 0.01), ('interconnected', 0.01), ('straps', 0.01)] | loopc/5: [('Cbd', 0.05), ('hoodie', 0.05), ('Airbnb', 0.05), ('covid', 0.05), ('Overwatch', 0.04)] | yokedc/0: [('Clan', 0.02), ('bleach', 0.02), ('nursery', 0.02), ('Lyrics', 0.02), ('mary', 0.02)] | yokedc/10: [('vibe', 0.03), ('Overwatch', 0.02), ('covid', 0.02), ('Lyft', 0.02), ('Interviews', 0.01)] | clampc/5: [('Cbd', 0.05), ('hoodie', 0.05), ('Airbnb', 0.05), ('covid', 0.05), ('Overwatch', 0.04)]
+19:37:36    e.g. thought: and so you'd get on the train and you'd be able to be on the train for a week, and not really know where you were going, and you'd get off at a different place || answer:  thinking about the things that I have been thinking about, and the things that I have been thinking about have been thi
+19:37:59 thought 110/400 (2.24 s/thought, eta 10.8 min) | self-models: free/0: [('africa', 0.02), ('extremist', 0.02), ('Conspiracy', 0.01), ('Gladiator', 0.01), ('unbearable', 0.01)] | free/10: [('Lauderdale', 0.02), ('scandals', 0.01), ('damn', 0.01), ('andre', 0.01), ('Hwy', 0.01)] | loopc/5: [('covid', 0.05), ('Dhabi', 0.04), ('emoji', 0.04), ('vulnerabilities', 0.03), ('Cbd', 0.03)] | yokedc/0: [('nord', 0.02), ('coral', 0.02), ('satin', 0.02), ('Geological', 0.01), ('theres', 0.01)] | yokedc/10: [('vibe', 0.05), ('covid', 0.05), ('Overwatch', 0.03), ('Airbnb', 0.03), ('Cbd', 0.02)] | clampc/5: [('covid', 0.05), ('Dhabi', 0.04), ('emoji', 0.04), ('vulnerabilities', 0.03), ('Cbd', 0.03)]
+19:37:59    e.g. thought: well, I was looking for my camera, and I thought, 'I really want this camera; I want to have this camera, but I don't have the money to buy it.' So, || answer:  thinking about the fact that I'm recording my thoughts, and that I'm thinking about the fact that I'm recording my
+19:38:21 thought 120/400 (2.24 s/thought, eta 10.4 min) | self-models: free/0: [('Throne', 0.01), ('waivers', 0.01), ('emo', 0.01), ('Xxx', 0.01), ('vulnerabilities', 0.01)] | free/10: [('hoodie', 0.03), ('superhero', 0.03), ('Airbnb', 0.02), ('vulnerabilities', 0.01), ('Honestly', 0.01)] | loopc/5: [('vibe', 0.03), ('vulnerabilities', 0.02), ('lockdown', 0.02), ('shard', 0.01), ('Biden', 0.01)] | yokedc/0: [('monstrous', 0.02), ('ani', 0.02), ('andre', 0.02), ('Laboratories', 0.02), ('xml', 0.01)] | yokedc/10: [('Biden', 0.03), ('covid', 0.02), ('meetup', 0.02), ('yuan', 0.02), ('Airbnb', 0.02)] | clampc/5: [('vibe', 0.03), ('vulnerabilities', 0.02), ('lockdown', 0.02), ('shard', 0.01), ('Biden', 0.01)]
+19:38:21    e.g. thought:  you can't do that to people, you know. I just don't think people should be able to do that to you and not have any punishment or anything. Because if you just  || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that my 
+19:38:45 thought 130/400 (2.25 s/thought, eta 10.1 min) | self-models: free/0: [('vulnerabilities', 0.03), ('tumblr', 0.02), ('emo', 0.02), ('Mandela', 0.01), ('Growing', 0.01)] | free/10: [('Reddit', 0.02), ('shitty', 0.02), ('vulnerabilities', 0.02), ('betrayal', 0.01), ('Biden', 0.01)] | loopc/5: [('Overwatch', 0.03), ('hoodie', 0.03), ('covid', 0.02), ('hashtags', 0.02), ('Airbnb', 0.02)] | yokedc/0: [('yer', 0.02), ('xml', 0.02), ('sts', 0.02), ('ebony', 0.01), ('Decoration', 0.01)] | yokedc/10: [('Biden', 0.03), ('cybersecurity', 0.02), ('Tracker', 0.02), ('impedance', 0.01), ('mindfulness', 0.01)] | clampc/5: [('Overwatch', 0.03), ('hoodie', 0.03), ('covid', 0.02), ('hashtags', 0.02), ('Airbnb', 0.02)]
+19:38:45    e.g. thought:  you're always so careful and careful with your words. That's kind of what I love about you. The way you say things. It's like you're always being thoughtful of || answer:  thinking about the fact that I've been thinking about the fact that I've been thinking about the fact that I've been
+19:39:07 thought 140/400 (2.24 s/thought, eta 9.7 min) | self-models: free/0: [('vibe', 0.03), ('andre', 0.02), ('whispers', 0.01), ('Zoom', 0.01), ('reconnect', 0.01)] | free/10: [('Assurance', 0.02), ('generals', 0.02), ('damned', 0.01), ('crises', 0.01), ('Meditation', 0.01)] | loopc/5: [('Airbnb', 0.04), ('cybersecurity', 0.04), ('emoji', 0.04), ('vibe', 0.02), ('covid', 0.02)] | yokedc/0: [('monstrous', 0.03), ('andre', 0.02), ('xml', 0.02), ('colours', 0.02), ('yer', 0.02)] | yokedc/10: [('meetup', 0.04), ('Accountability', 0.04), ('cybersecurity', 0.03), ('Gradient', 0.03), ('apis', 0.03)] | clampc/5: [('Airbnb', 0.04), ('cybersecurity', 0.04), ('emoji', 0.04), ('vibe', 0.02), ('covid', 0.02)]
+19:39:07    e.g. thought: And, we're sitting in here, and we're thinking that's it. There's nothing more we can do. The train's on time. It's coming in. We're going to catch || answer:  thinking about the fact that I'm thinking about my mind. And that's what I'm thinking about now, that I
+19:39:30 thought 150/400 (2.25 s/thought, eta 9.4 min) | self-models: free/0: [('resilience', 0.02), ('Accountability', 0.01), ('leadership', 0.01), ('entrepreneurship', 0.01), ('Manafort', 0.01)] | free/10: [('eternity', 0.02), ('rsa', 0.02), ('Awakening', 0.02), ('consciousness', 0.01), ('Connectivity', 0.01)] | loopc/5: [('vibe', 0.04), ('Airbnb', 0.04), ('hashtags', 0.03), ('betrayal', 0.02), ('yuan', 0.02)] | yokedc/0: [('satin', 0.03), ('childs', 0.02), ('Cumberland', 0.02), ('screamed', 0.01), ('roar', 0.01)] | yokedc/10: [('Accountability', 0.04), ('mindfulness', 0.03), ('foundational', 0.02), ('meetup', 0.02), ('elites', 0.02)] | clampc/5: [('vibe', 0.04), ('Airbnb', 0.04), ('hashtags', 0.03), ('betrayal', 0.02), ('yuan', 0.02)]
+19:39:30    e.g. thought:  and this is the big problem in my life is that I have to work, you know, a lot and I just don't have that much time for my kids. You know, I have to || answer:  thinking about the fact that I'm recording my thoughts, and that I'm going to be playing back to myself what I
+19:39:52 thought 160/400 (2.25 s/thought, eta 9.0 min) | self-models: free/0: [('ssh', 0.01), ('millennia', 0.01), ('elites', 0.01), ('resilience', 0.01), ('Weapons', 0.01)] | free/10: [('surreal', 0.02), ('andre', 0.01), ('Heaven', 0.01), ('consciousness', 0.01), ('sentient', 0.01)] | loopc/5: [('Airbnb', 0.04), ('vulnerabilities', 0.04), ('vibe', 0.03), ('hoodie', 0.03), ('Cbd', 0.02)] | yokedc/0: [('sculptures', 0.02), ('xml', 0.02), ('satin', 0.01), ('Colour', 0.01), ('frog', 0.01)] | yokedc/10: [('opioids', 0.03), ('vibe', 0.02), ('fundamentally', 0.01), ('airspace', 0.01), ('hashtag', 0.01)] | clampc/5: [('Airbnb', 0.07), ('covid', 0.04), ('vibe', 0.04), ('Cbd', 0.04), ('emoji', 0.04)]
+19:39:52    e.g. thought: I was watching a documentary on the TV about how a boy in the war who had been in the army. He had been shot at a lot. And they had all these men. The men who || answer:  thinking about the fact that I have a lot of work to do and I need to get it done. I've been
+19:40:16 thought 170/400 (2.25 s/thought, eta 8.6 min) | self-models: free/0: [('shitty', 0.02), ('subpoena', 0.01), ('Recent', 0.01), ('Personnel', 0.01), ('irc', 0.01)] | free/10: [('dna', 0.01), ('painfully', 0.01), ('skyline', 0.01), ('Gallery', 0.01), ('Bluetooth', 0.01)] | loopc/5: [('Airbnb', 0.03), ('breakup', 0.03), ('hoodie', 0.03), ('vibe', 0.02), ('leaderboard', 0.02)] | yokedc/0: [('ieee', 0.02), ('bracelets', 0.02), ('initials', 0.02), ('Vocabulary', 0.02), ('Splash', 0.01)] | yokedc/10: [('fundamentally', 0.05), ('Tensor', 0.04), ('yuan', 0.03), ('meetup', 0.02), ('Jazeera', 0.02)] | clampc/5: [('covid', 0.03), ('Overwatch', 0.02), ('Airbnb', 0.02), ('vibe', 0.02), ('hoodie', 0.02)]
+19:40:16    e.g. thought:  and then the whole thing would have been in the hands of the doctors, and then if they had done something to him, they wouldn't have known if he'd done anythin || answer:  thinking about the fact that I have a lot of work to do, and I have to get it done. I've
+19:40:38 thought 180/400 (2.25 s/thought, eta 8.3 min) | self-models: free/0: [('sentient', 0.03), ('rsa', 0.02), ('Enlightenment', 0.02), ('Terrain', 0.02), ('deepest', 0.01)] | free/10: [('hoodie', 0.01), ('homophobic', 0.01), ('Mumbai', 0.01), ('rumours', 0.01), ('Blowjob', 0.01)] | loopc/5: [('vibe', 0.03), ('Airbnb', 0.03), ('betrayal', 0.03), ('covid', 0.03), ('emoji', 0.02)] | yokedc/0: [('amd', 0.02), ('picnic', 0.02), ('Lyrics', 0.02), ('Clan', 0.02), ('roar', 0.02)] | yokedc/10: [('fundamentally', 0.05), ('Accountability', 0.04), ('elites', 0.04), ('mana', 0.03), ('ssh', 0.02)] | clampc/5: [('Airbnb', 0.03), ('covid', 0.03), ('mindset', 0.03), ('Huawei', 0.02), ('hoodie', 0.02)]
+19:40:38    e.g. thought:  you know the reason I want to talk about it is because the subject matter is so wide, it's almost infinite in the amount of things that you can talk about. And || answer:  thinking about the way I've been feeling, and how I've been thinking about the way I've been feeling. I
+19:41:02 thought 190/400 (2.26 s/thought, eta 7.9 min) | self-models: free/0: [('vibe', 0.02), ('Www', 0.01), ('Feeling', 0.01), ('Layers', 0.01), ('sentient', 0.01)] | free/10: [('https', 0.01), ('weir', 0.01), ('fraught', 0.01), ('transformative', 0.01), ('Tonight', 0.01)] | loopc/5: [('covid', 0.06), ('Airbnb', 0.04), ('emoji', 0.03), ('Overwatch', 0.03), ('vibe', 0.03)] | yokedc/0: [('theres', 0.02), ('picnic', 0.02), ('acidic', 0.02), ('monstrous', 0.02), ('pirate', 0.01)] | yokedc/10: [('millennia', 0.04), ('Connectivity', 0.03), ('physicist', 0.03), ('impedance', 0.03), ('yuan', 0.02)] | clampc/5: [('covid', 0.03), ('transformative', 0.01), ('meetup', 0.01), ('Shard', 0.01), ('vibe', 0.01)]
+19:41:02    e.g. thought:  and it's like that you have to do this and I just don't know if I'm doing the right thing or not because you know, I don't know why I'm doing this thing. || answer:  thinking about the problem of the day. I've been thinking about the problems that I have to solve in order to get
+19:41:25 thought 200/400 (2.26 s/thought, eta 7.5 min) | self-models: free/0: [('Heaven', 0.02), ('andre', 0.02), ('labyrinth', 0.01), ('Chaos', 0.01), ('rgb', 0.01)] | free/10: [('sculpture', 0.03), ('png', 0.02), ('damned', 0.01), ('alphabet', 0.01), ('andre', 0.01)] | loopc/5: [('vibe', 0.03), ('Overwatch', 0.02), ('covid', 0.02), ('unbearable', 0.02), ('Cbd', 0.02)] | yokedc/0: [('emoji', 0.04), ('glitch', 0.03), ('Minecraft', 0.02), ('cosplay', 0.02), ('twists', 0.01)] | yokedc/10: [('etc', 0.03), ('Sizes', 0.03), ('pixel', 0.03), ('tensor', 0.02), ('mana', 0.02)] | clampc/5: [('Thoughts', 0.02), ('bullshit', 0.01), ('timelines', 0.01), ('waivers', 0.01), ('strategist', 0.01)]
+19:41:25    e.g. thought: we should start the whole thing over, with a new name, a new, completely different name, for this thing, like a whole different, new, whole different thing, but || answer:  thinking about the fact that I'm going to have to do a lot of work tomorrow, and that I'm not going
+19:41:48 thought 210/400 (2.27 s/thought, eta 7.2 min) | self-models: free/0: [('blah', 0.01), ('colleague', 0.01), ('ssh', 0.01), ('bullpen', 0.01), ('Alumni', 0.01)] | free/10: [('bdsm', 0.02), ('superhero', 0.02), ('Oprah', 0.01), ('Interviews', 0.01), ('hallway', 0.01)] | loopc/5: [('vibe', 0.05), ('covid', 0.05), ('Overwatch', 0.04), ('vulnerabilities', 0.04), ('hoodie', 0.03)] | yokedc/0: [('hoodie', 0.03), ('Tumblr', 0.03), ('emoji', 0.03), ('covid', 0.03), ('cosplay', 0.02)] | yokedc/10: [('vibration', 0.04), ('Sizes', 0.03), ('physicist', 0.02), ('humanoid', 0.02), ('vpn', 0.02)] | clampc/5: [('vibe', 0.03), ('covid', 0.02), ('hashtag', 0.02), ('cybersecurity', 0.02), ('Overwatch', 0.02)]
+19:41:48    e.g. thought: I just spent two hours in the back room in here, looking at a book that's not all that useful. It's a book of, I don't know, some kind of a mechanical design || answer:  thinking about the things that I've been thinking about, and the things that I've been thinking about have been thinkin
+19:42:11 thought 220/400 (2.26 s/thought, eta 6.8 min) | self-models: free/0: [('Teacher', 0.02), ('notebooks', 0.02), ('student', 0.02), ('leaderboard', 0.01), ('resignation', 0.01)] | free/10: [('demonic', 0.01), ('sentient', 0.01), ('deepest', 0.01), ('Childhood', 0.01), ('vulnerabilities', 0.01)] | loopc/5: [('emoji', 0.05), ('covid', 0.05), ('hoodie', 0.03), ('vibe', 0.03), ('Airbnb', 0.03)] | yokedc/0: [('covid', 0.05), ('emoji', 0.05), ('cosplay', 0.03), ('https', 0.02), ('betrayal', 0.02)] | yokedc/10: [('sentient', 0.04), ('Heaven', 0.03), ('Void', 0.02), ('mythology', 0.02), ('Infinite', 0.02)] | clampc/5: [('hoodie', 0.03), ('Airbnb', 0.03), ('covid', 0.03), ('vibe', 0.03), ('emoji', 0.02)]
+19:42:11    e.g. thought: I was at home today. I had just finished work. I was going to be home for an hour. I was sitting in the living room, sitting at the table. I was looking through || answer:  thinking about the fact that I have a lot of work to do, and I'm not sure how I'm going to
+19:42:35 thought 230/400 (2.27 s/thought, eta 6.4 min) | self-models: free/0: [('leaderboard', 0.04), ('smells', 0.02), ('busiest', 0.02), ('Marketplace', 0.01), ('Warrior', 0.01)] | free/10: [('Gods', 0.03), ('betrayal', 0.02), ('murderers', 0.02), ('Himself', 0.02), ('Confidential', 0.01)] | loopc/5: [('vibe', 0.05), ('covid', 0.04), ('emoji', 0.04), ('vulnerabilities', 0.03), ('Cbd', 0.03)] | yokedc/0: [('teammate', 0.04), ('breakup', 0.03), ('Airbnb', 0.03), ('hashtags', 0.03), ('Honestly', 0.02)] | yokedc/10: [('sentient', 0.07), ('Enlightenment', 0.06), ('Conscious', 0.04), ('Alignment', 0.03), ('Meditation', 0.03)] | clampc/5: [('timelines', 0.03), ('mindfulness', 0.02), ('resilience', 0.02), ('Biden', 0.02), ('coursework', 0.01)]
+19:42:35    e.g. thought:  uh... what is it? A good time or... what? Like... what kind of a time is this? I don't know. It's like the second, what? The second one. || answer:  thinking about the things that I have been thinking about, and the things that I have been thinking about are things th
+19:42:56 thought 240/400 (2.27 s/thought, eta 6.0 min) | self-models: free/0: [('emo', 0.02), ('terrifying', 0.01), ('wifi', 0.01), ('whispers', 0.01), ('Voldemort', 0.01)] | free/10: [('vulnerabilities', 0.01), ('Indigenous', 0.01), ('fuck', 0.01), ('wreckage', 0.01), ('Covenant', 0.01)] | loopc/5: [('Airbnb', 0.04), ('vibe', 0.03), ('Cbd', 0.03), ('shard', 0.03), ('emoji', 0.02)] | yokedc/0: [('vibe', 0.06), ('emoji', 0.04), ('Overwatch', 0.03), ('Biden', 0.03), ('vulnerabilities', 0.03)] | yokedc/10: [('existential', 0.05), ('Meditation', 0.05), ('Alignment', 0.03), ('vibe', 0.03), ('fucking', 0.03)] | clampc/5: [('meetup', 0.04), ('resilience', 0.03), ('Huawei', 0.02), ('grounding', 0.02), ('deadlines', 0.02)]
+19:42:56    e.g. thought: I was just going to say, I think this is going to be the first time I ever say this in person, that the way I always used to see things was always just this hug || answer:  thinking about the fact that I'm going to be late for work. I'm going to have to rush. I'm
+19:43:20 thought 250/400 (2.27 s/thought, eta 5.7 min) | self-models: free/0: [('sentient', 0.02), ('tensor', 0.02), ('andre', 0.01), ('existential', 0.01), ('sprawling', 0.01)] | free/10: [('Responsibilities', 0.02), ('Tribune', 0.01), ('africa', 0.01), ('biggest', 0.01), ('fuck', 0.01)] | loopc/5: [('cryptocurrency', 0.02), ('meetup', 0.02), ('yuan', 0.02), ('Tracker', 0.02), ('resilience', 0.02)] | yokedc/0: [('Airbnb', 0.04), ('emoji', 0.03), ('Overwatch', 0.03), ('covid', 0.02), ('yuan', 0.02)] | yokedc/10: [('sentient', 0.06), ('Meditation', 0.05), ('sensation', 0.04), ('Alignment', 0.03), ('Exists', 0.02)] | clampc/5: [('meetup', 0.04), ('Accountability', 0.04), ('foundational', 0.03), ('github', 0.03), ('Huawei', 0.03)]
+19:43:20    e.g. thought:  a feeling of wanting to see something that wasn't there, the feeling of wanting something to be there. And you're not sure it's going to be there, it's just a  || answer:  thinking about the fact that I'm doing this experiment. I'm thinking about the fact that I'm thinking about this experi
+19:43:42 thought 260/400 (2.27 s/thought, eta 5.3 min) | self-models: free/0: [('physicist', 0.02), ('intuitive', 0.02), ('Conscious', 0.02), ('Gesture', 0.02), ('rsa', 0.01)] | free/10: [('fucking', 0.03), ('Fortress', 0.02), ('timeframe', 0.02), ('africa', 0.01), ('Funeral', 0.01)] | loopc/5: [('yuan', 0.04), ('hashtag', 0.03), ('Elon', 0.03), ('Linkedin', 0.03), ('algorithms', 0.02)] | yokedc/0: [('Airbnb', 0.02), ('meetup', 0.02), ('waivers', 0.02), ('bullshit', 0.01), ('Overwatch', 0.01)] | yokedc/10: [('mystical', 0.04), ('sentient', 0.03), ('Infinity', 0.03), ('Fuck', 0.03), ('Nested', 0.03)] | clampc/5: [('fundamentally', 0.05), ('Jazeera', 0.05), ('Accountability', 0.04), ('Researchers', 0.04), ('opioids', 0.03)]
+19:43:42    e.g. thought:  it's all a bit of a puzzle, isn't it? I mean, what am I doing here? What's the point of me being here? What am I doing? I don't know || answer:  thinking about the fact that I am thinking about my mind. And that I am thinking about the fact that I am thinking
+19:44:06 thought 270/400 (2.27 s/thought, eta 4.9 min) | self-models: free/0: [('Symptoms', 0.01), ('luz', 0.01), ('trembling', 0.01), ('Layers', 0.01), ('collapses', 0.01)] | free/10: [('sentient', 0.02), ('existential', 0.02), ('Enlightenment', 0.02), ('Fuck', 0.01), ('blindly', 0.01)] | loopc/5: [('Github', 0.05), ('yuan', 0.04), ('millennia', 0.03), ('impedance', 0.03), ('Brexit', 0.02)] | yokedc/0: [('vulnerabilities', 0.04), ('Accountability', 0.04), ('Huawei', 0.03), ('Linkedin', 0.03), ('Biden', 0.03)] | yokedc/10: [('sentient', 0.06), ('Institut', 0.03), ('Meditation', 0.03), ('existential', 0.03), ('Conscious', 0.03)] | clampc/5: [('existential', 0.05), ('Meditation', 0.04), ('fuck', 0.04), ('strategist', 0.03), ('fundamentally', 0.03)]
+19:44:06    e.g. thought:  I have a theory that the most important factor in the course of an illness is a certain kind of thought process. The patient's thinking is constantly influence || answer:  thinking about the fact that I'm thinking about my mind thinking about the fact that I'm thinking about my mind, and
+19:44:28 thought 280/400 (2.27 s/thought, eta 4.5 min) | self-models: free/0: [('italia', 0.03), ('legitimacy', 0.03), ('theological', 0.03), ('authoritarian', 0.03), ('damned', 0.02)] | free/10: [('vulnerabilities', 0.02), ('existential', 0.02), ('Honestly', 0.02), ('Patreon', 0.02), ('gpu', 0.01)] | loopc/5: [('Huawei', 0.05), ('Accountability', 0.04), ('mana', 0.04), ('github', 0.03), ('Biden', 0.03)] | yokedc/0: [('Accountability', 0.05), ('geopolitical', 0.04), ('elites', 0.02), ('Tian', 0.02), ('Brexit', 0.02)] | yokedc/10: [('sentient', 0.05), ('Fuck', 0.04), ('etc', 0.03), ('nacional', 0.03), ('Heaven', 0.02)] | clampc/5: [('Researchers', 0.05), ('impedance', 0.03), ('bullshit', 0.02), ('elites', 0.02), ('existential', 0.02)]
+19:44:28    e.g. thought:  I am a very ordinary man, I don't have much in the way of natural talents. I don't do things especially well or badly. I'm not very creative, I'm not a very || answer:  thinking about the fact that I have been thinking about the fact that my mind has been thinking about the fact that I h
+19:44:52 thought 290/400 (2.27 s/thought, eta 4.2 min) | self-models: free/0: [('Enemies', 0.02), ('bla', 0.02), ('Alumni', 0.02), ('groin', 0.01), ('damn', 0.01)] | free/10: [('waivers', 0.02), ('vibe', 0.02), ('Assange', 0.02), ('teammates', 0.02), ('fuck', 0.02)] | loopc/5: [('fundamentally', 0.04), ('Alignment', 0.03), ('yuan', 0.03), ('github', 0.02), ('impedance', 0.02)] | yokedc/0: [('Accountability', 0.03), ('Brexit', 0.03), ('Huawei', 0.02), ('fundamentally', 0.02), ('calibration', 0.02)] | yokedc/10: [('sentient', 0.08), ('Meditation', 0.07), ('sensation', 0.04), ('existential', 0.04), ('Layers', 0.03)] | clampc/5: [('Connectivity', 0.03), ('sentient', 0.03), ('physicist', 0.03), ('fuck', 0.02), ('Alignment', 0.02)]
+19:44:52    e.g. thought: And the girl got it... And she was so happy, she didn't care if she was getting... she didn't want to tell her mom it was a girl. I thought, 'No, || answer:  thinking about the fact that I have been thinking about my mind for the last ten moments. And I have been thinking abou
+19:45:14 thought 300/400 (2.27 s/thought, eta 3.8 min) | self-models: free/0: [('sculptures', 0.03), ('Colors', 0.02), ('colourful', 0.01), ('womb', 0.01), ('tumblr', 0.01)] | free/10: [('sentient', 0.04), ('png', 0.03), ('twists', 0.02), ('andre', 0.02), ('Spiral', 0.02)] | loopc/5: [('elites', 0.04), ('Alignment', 0.03), ('cybersecurity', 0.03), ('impedance', 0.03), ('fundamentally', 0.03)] | yokedc/0: [('cybersecurity', 0.04), ('Huawei', 0.04), ('meetup', 0.03), ('tweeted', 0.02), ('Dataset', 0.02)] | yokedc/10: [('mystical', 0.05), ('sentient', 0.04), ('Layers', 0.04), ('sensation', 0.03), ('Meditation', 0.03)] | clampc/5: [('Alignment', 0.04), ('bullshit', 0.03), ('sentient', 0.03), ('Connectivity', 0.02), ('physicist', 0.02)]
+19:45:14    e.g. thought:  I'm going to eat some bread, I don't know how many, I don't have time for more. I'm going to sleep. I don't want to eat right now, I'm || answer:  thinking about the problem of how to solve the problem of how to solve this problem. And then it has been thinking abou
+19:45:38 thought 310/400 (2.28 s/thought, eta 3.4 min) | self-models: free/0: [('campus', 0.02), ('Moms', 0.01), ('blah', 0.01), ('teammates', 0.01), ('leaderboard', 0.01)] | free/10: [('andre', 0.02), ('corps', 0.02), ('Fighters', 0.02), ('damned', 0.02), ('Heaven', 0.01)] | loopc/5: [('Github', 0.03), ('Researchers', 0.03), ('foundational', 0.02), ('tensor', 0.02), ('yuan', 0.02)] | yokedc/0: [('hashtags', 0.03), ('Airbnb', 0.03), ('yuan', 0.02), ('meetup', 0.02), ('cybersecurity', 0.02)] | yokedc/10: [('Darkness', 0.04), ('sensations', 0.03), ('Artist', 0.03), ('Fucking', 0.02), ('vortex', 0.02)] | clampc/5: [('existential', 0.04), ('elites', 0.02), ('intel', 0.01), ('Infrastructure', 0.01), ('Consortium', 0.01)]
+19:45:38    e.g. thought:  I was thinking about the '75 book, which is 50 years old and which has become a classic. And I was thinking how we started to teach children to read in a very  || answer:  thinking about the problem of how to get to the airport. I have been thinking about the best route to take, and
+19:46:00 thought 320/400 (2.27 s/thought, eta 3.0 min) | self-models: free/0: [('national', 0.03), ('pundits', 0.02), ('Constitution', 0.01), ('leaderboard', 0.01), ('Whatever', 0.01)] | free/10: [('Heaven', 0.05), ('authoritarian', 0.02), ('fuck', 0.02), ('sentient', 0.02), ('Warfare', 0.01)] | loopc/5: [('optimization', 0.04), ('Biden', 0.04), ('Dataset', 0.04), ('apis', 0.03), ('grounding', 0.03)] | yokedc/0: [('implementations', 0.04), ('yuan', 0.04), ('Elon', 0.04), ('Labs', 0.03), ('tensor', 0.03)] | yokedc/10: [('Void', 0.03), ('xml', 0.03), ('sculptures', 0.03), ('ani', 0.03), ('humanoid', 0.02)] | clampc/5: [('fuck', 0.08), ('Connectivity', 0.06), ('physicist', 0.05), ('sensors', 0.04), ('mindfulness', 0.03)]
+19:46:00    e.g. thought: it's like the way it was in the fifties, when the two parties were the Republicans and the Democrats, and one of the things that happened to the Republicans was || answer:  thinking about the recording. It has been thinking about the fact that I'm going to be asked to listen to the recording
+19:46:25 thought 330/400 (2.28 s/thought, eta 2.7 min) | self-models: free/0: [('Financial', 0.04), ('Huawei', 0.03), ('nacional', 0.03), ('overwhelmingly', 0.03), ('shit', 0.02)] | free/10: [('Rebellion', 0.02), ('Heaven', 0.01), ('stockholm', 0.01), ('shitty', 0.01), ('rumours', 0.01)] | loopc/5: [('Dataset', 0.05), ('Brexit', 0.04), ('waivers', 0.03), ('yuan', 0.03), ('meetup', 0.03)] | yokedc/0: [('yuan', 0.04), ('ssh', 0.03), ('meetup', 0.03), ('leaderboard', 0.03), ('impedance', 0.03)] | yokedc/10: [('vibration', 0.06), ('sentient', 0.04), ('symmetry', 0.03), ('Sensor', 0.03), ('Depths', 0.03)] | clampc/5: [('physicist', 0.04), ('etc', 0.03), ('yuan', 0.02), ('impedance', 0.02), ('genera', 0.02)]
+19:46:25    e.g. thought: So it's the kind of thing, it's, I'm saying it's like you have a lot of ideas that are like you've been thinking about something for a long time and you don't || answer:  thinking about the past, the future, and the present. I've been thinking about my life, my family, my
+19:46:47 thought 340/400 (2.28 s/thought, eta 2.3 min) | self-models: free/0: [('billions', 0.01), ('betray', 0.01), ('waivers', 0.01), ('famously', 0.01), ('school', 0.01)] | free/10: [('Nightmare', 0.02), ('Clan', 0.02), ('theres', 0.01), ('onstage', 0.01), ('treason', 0.01)] | loopc/5: [('optimization', 0.03), ('nan', 0.02), ('Xiaomi', 0.02), ('protections', 0.02), ('elites', 0.02)] | yokedc/0: [('cybersecurity', 0.04), ('millennia', 0.03), ('yuan', 0.02), ('vpn', 0.02), ('mana', 0.02)] | yokedc/10: [('sentient', 0.04), ('Architects', 0.03), ('Infinite', 0.03), ('labyrinth', 0.02), ('Visualization', 0.02)] | clampc/5: [('etc', 0.04), ('genera', 0.03), ('quantum', 0.03), ('implementations', 0.03), ('nacional', 0.02)]
+19:46:47    e.g. thought: [I]f we can get rid of those 10 years, you know, when I'm just a kid, I can become the adult that I want to become, and maybe I'll || answer:  thinking about the fact that I am recording my thoughts, and that I am thinking about my thoughts, and that I am
+19:47:11 thought 350/400 (2.28 s/thought, eta 1.9 min) | self-models: free/0: [('andre', 0.02), ('bastard', 0.02), ('shrine', 0.01), ('incest', 0.01), ('Sanctuary', 0.01)] | free/10: [('authoritarian', 0.02), ('Billion', 0.01), ('waivers', 0.01), ('extremists', 0.01), ('fiercely', 0.01)] | loopc/5: [('vibration', 0.03), ('wei', 0.03), ('pulse', 0.02), ('etc', 0.02), ('damned', 0.02)] | yokedc/0: [('legitimacy', 0.03), ('Meditation', 0.02), ('millennia', 0.02), ('Trusted', 0.02), ('existential', 0.02)] | yokedc/10: [('sentient', 0.06), ('surreal', 0.04), ('luz', 0.03), ('graphene', 0.03), ('Layers', 0.03)] | clampc/5: [('physicist', 0.04), ('Turing', 0.03), ('Lisp', 0.02), ('andre', 0.02), ('genera', 0.02)]
+19:47:11    e.g. thought: I remember sitting with the sun on my face on the shore of Lake Michigan and the water was boiling off, and I was lying on the sand with my wife and we were loo || answer:  thinking about the fact that I am a human being, and that I am going to die, and that I have to
+19:47:33 thought 360/400 (2.28 s/thought, eta 1.5 min) | self-models: free/0: [('Negro', 0.03), ('pagan', 0.02), ('sculptures', 0.02), ('Heaven', 0.02), ('Specifications', 0.02)] | free/10: [('sentient', 0.03), ('fundamentally', 0.02), ('Alignment', 0.02), ('existential', 0.01), ('etc', 0.01)] | loopc/5: [('vibration', 0.05), ('orbits', 0.02), ('Sizes', 0.02), ('Sphere', 0.02), ('pulse', 0.02)] | yokedc/0: [('geopolitical', 0.03), ('legitimacy', 0.03), ('elites', 0.02), ('Accountability', 0.02), ('mana', 0.02)] | yokedc/10: [('sentient', 0.07), ('Meditation', 0.06), ('Sensor', 0.04), ('existential', 0.04), ('luz', 0.04)] | clampc/5: [('implementations', 0.03), ('genera', 0.03), ('eternity', 0.03), ('physicist', 0.03), ('etc', 0.02)]
+19:47:33    e.g. thought: and so it's not...what I call, it's not an attack. It's not a fight. It's the two together, and the two together is...but there's no fighting. || answer:  thinking about the problem of how to solve the problem of how to solve this problem. It's a bit like a circle
+19:47:58 thought 370/400 (2.28 s/thought, eta 1.1 min) | self-models: free/0: [('Spiritual', 0.03), ('spirituality', 0.02), ('Presence', 0.02), ('priesthood', 0.02), ('Symposium', 0.02)] | free/10: [('ness', 0.03), ('damned', 0.02), ('instrumentation', 0.01), ('Sleeve', 0.01), ('Disclosure', 0.01)] | loopc/5: [('etc', 0.02), ('sonic', 0.02), ('vibration', 0.02), ('pulse', 0.02), ('Gesture', 0.01)] | yokedc/0: [('Accountability', 0.04), ('fundamentally', 0.04), ('Github', 0.03), ('wei', 0.03), ('Biden', 0.02)] | yokedc/10: [('Artist', 0.02), ('Fuck', 0.02), ('eternity', 0.02), ('humanoid', 0.02), ('Conscious', 0.02)] | clampc/5: [('quantum', 0.03), ('Trustees', 0.03), ('etc', 0.03), ('millennia', 0.02), ('mana', 0.02)]
+19:47:58    e.g. thought: I was at a funeral recently...that was the first time I had to go to a funeral in quite a few years, and I think it's a very important thing to do, to attend a || answer:  thinking about the fact that I am thinking about my mind, and that I am thinking about the fact that I am thinking
+19:48:20 thought 380/400 (2.28 s/thought, eta 0.8 min) | self-models: free/0: [('impeachment', 0.03), ('legitimacy', 0.02), ('Faction', 0.02), ('nacional', 0.02), ('Testament', 0.02)] | free/10: [('vibe', 0.03), ('bdsm', 0.03), ('waivers', 0.02), ('Blowjob', 0.02), ('Trusted', 0.02)] | loopc/5: [('rgb', 0.02), ('ness', 0.02), ('iterations', 0.02), ('Infinite', 0.02), ('Laboratories', 0.01)] | yokedc/0: [('cybersecurity', 0.04), ('impedance', 0.03), ('Dataset', 0.03), ('optimization', 0.03), ('yuan', 0.03)] | yokedc/10: [('sensation', 0.1), ('Meditation', 0.04), ('sentient', 0.04), ('Sensor', 0.03), ('fucking', 0.03)] | clampc/5: [('ripple', 0.03), ('Logo', 0.02), ('physicist', 0.01), ('Visualization', 0.01), ('parity', 0.01)]
+19:48:20    e.g. thought:  and you know, there is another thing about, about people who are... who are in prison and have committed a crime. You know, in my own case, I would not say tha || answer:  thinking about the things that I have been thinking about, and the things that I have been thinking about have been thi
+19:48:44 thought 390/400 (2.29 s/thought, eta 0.4 min) | self-models: free/0: [('Fuck', 0.02), ('andre', 0.02), ('intimately', 0.02), ('notebooks', 0.01), ('waivers', 0.01)] | free/10: [('millennia', 0.02), ('vibe', 0.01), ('Throne', 0.01), ('vulnerabilities', 0.01), ('Fukushima', 0.01)] | loopc/5: [('Laboratories', 0.03), ('Finite', 0.03), ('eternity', 0.03), ('png', 0.02), ('genera', 0.02)] | yokedc/0: [('cybersecurity', 0.05), ('Dataset', 0.04), ('waivers', 0.04), ('yuan', 0.03), ('meetup', 0.03)] | yokedc/10: [('sensations', 0.05), ('healer', 0.04), ('spirituality', 0.03), ('fucking', 0.03), ('Connectivity', 0.03)] | clampc/5: [('Logo', 0.03), ('Finite', 0.02), ('symmetry', 0.02), ('teh', 0.02), ('sculptures', 0.02)]
+19:48:44    e.g. thought: I'm going to be doing some kind of research on, on people's memory and thinking, I think I'm going to give a seminar on, on something, I'm doing some kind of wo || answer:  thinking about the recording. I have been thinking about the fact that I am thinking about the recording, and I have be
+19:49:06 thought 400/400 (2.28 s/thought, eta 0.0 min) | self-models: free/0: [('Fuck', 0.02), ('Institut', 0.01), ('etc', 0.01), ('immersive', 0.01), ('protocols', 0.01)] | free/10: [('Fuck', 0.04), ('waivers', 0.02), ('Artists', 0.01), ('Corps', 0.01), ('Almighty', 0.01)] | loopc/5: [('sentient', 0.06), ('eternity', 0.04), ('Experimental', 0.02), ('Void', 0.02), ('vibration', 0.02)] | yokedc/0: [('cybersecurity', 0.05), ('yuan', 0.04), ('Elon', 0.03), ('gaps', 0.03), ('Dataset', 0.02)] | yokedc/10: [('sensations', 0.09), ('Meditation', 0.04), ('spirituality', 0.03), ('Pain', 0.02), ('sentient', 0.02)] | clampc/5: [('finite', 0.02), ('png', 0.02), ('carr', 0.02), ('symmetry', 0.02), ('org', 0.02)]
+19:49:06    e.g. thought:  I'm sorry, I don't know how to say that. I don't want to say that, but it has to be said, and it will have to be said in some way. I || answer:  thinking about the fact that I am a very good person, and that I am very good at what I do, and
+19:49:09 saved runs/mainC

runs/mainC/summary.txt

@@ -0,0 +1,104 @@
+PCA explained [0.224 0.141 0.07 ]
+
+wandering (mean cos at lags 1,5,20,60 | chance)
+  free/pre       0.19 0.11 0.05 0.07 | 0.05
+  free/post      0.19 0.11 0.05 0.04 | 0.05
+  loop/pre       0.35 0.30 0.27 0.22 | 0.03
+  loop/post      0.38 0.34 0.31 0.23 | 0.12
+  yoked/pre      0.34 0.29 0.25 0.21 | 0.03
+  yoked/post     0.39 0.32 0.28 0.23 | 0.12
+  clamp/pre      0.35 0.30 0.27 0.22 | 0.03
+  clamp/post     0.37 0.32 0.30 0.25 | 0.06
+
+fidelity {'m': (0.575, 0.198, 0.61)}
+
+self-model stability (consecutive cos, pre) vs content windows; cross-seed
+  free     self 0.720  content 0.342  cross 0.648  own-symbol 0.772 vs others 0.759  recon nan  generic 0.777
+  loop     self 0.914  content 0.756  cross 0.528  own-symbol 0.787 vs others 0.561  recon 0.679  generic 0.743
+  yoked    self 0.891  content 0.716  cross 0.529  own-symbol 0.782 vs others 0.569  recon 0.666  generic 0.740
+  clamp    self 0.914  content 0.756  cross 0.528  own-symbol 0.787 vs others 0.561  recon 0.679  generic 0.743
+
+prediction of the next window: own self-model | other row's | past window | injected  (|| beyond past: own, other)
+  free     0.227 | 0.169 | 0.325 | nan  || 0.079, 0.117  retrieval 0.12
+  loop     0.584 | 0.229 | 0.780 | 0.612  || 0.147, 0.076  retrieval 0.48
+  yoked    0.541 | 0.226 | 0.725 | 0.604  || 0.120, 0.082  retrieval 0.37
+  clamp    0.567 | 0.177 | 0.754 | 0.618  || 0.143, 0.060  retrieval 0.49
+
+prediction_tests
+  free                                     own_vs_other +0.058±0.008  perp_own_vs_other -0.037±0.005  retrieval +0.123±0.014  future_vs_past -0.229±0.007
+  loop                                     own_vs_other +0.355±0.008  perp_own_vs_other +0.071±0.007  retrieval +0.481±0.032  future_vs_past -0.037±0.002
+  yoked                                    own_vs_other +0.315±0.010  perp_own_vs_other +0.038±0.006  retrieval +0.367±0.030  future_vs_past -0.081±0.005
+  clamp                                    own_vs_other +0.390±0.013  perp_own_vs_other +0.083±0.008  retrieval +0.486±0.047  future_vs_past -0.052±0.002
+  yoked_own_vs_injected                    diff -0.063 ± 0.007  t -8.9  (0/15 seeds >0)
+  yoked_own_vs_injected_perp               diff -0.145 ± 0.007  t -20.4  (0/15 seeds >0)
+
+coherence_tests
+  loop_vs_free                             diff +0.183 ± 0.028  t +6.6  (15/15 seeds >0)
+  yoked_vs_free                            diff +0.168 ± 0.017  t +9.8  (15/15 seeds >0)
+  clamp_vs_free                            diff +0.183 ± 0.028  t +6.6  (15/15 seeds >0)
+  loop_vs_free_proj                        diff +0.079 ± 0.017  t +4.6  (13/15 seeds >0)
+  yoked_vs_free_proj                       diff +0.073 ± 0.010  t +7.1  (15/15 seeds >0)
+  clamp_vs_free_proj                       diff +0.079 ± 0.017  t +4.6  (13/15 seeds >0)
+  loop_vs_yoked                            diff +0.015 ± 0.032  t +0.5  (8/15 seeds >0)
+  loop_vs_yoked_proj                       diff +0.007 ± 0.018  t +0.4  (8/15 seeds >0)
+
+selfmodel_tests
+  loop_vs_yoked_cons_ind                   diff +0.045 ± 0.024  t +1.9  (9/15 seeds >0)
+  loop_vs_yoked_recon_ind                  diff +0.029 ± 0.023  t +1.3  (9/15 seeds >0)
+  loop_vs_free_cons_ind                    diff +0.603 ± 0.019  t +32.6  (15/15 seeds >0)
+  yoked_vs_free_cons_ind                   diff +0.559 ± 0.024  t +23.2  (15/15 seeds >0)
+  free_self_vs_content_stability           diff +0.379 ± 0.013  t +28.1  (15/15 seeds >0)
+  free_own_vs_other_symbol                 diff +0.013 ± 0.010  t +1.3  (12/15 seeds >0)
+  loop_self_vs_content_stability           diff +0.158 ± 0.021  t +7.4  (15/15 seeds >0)
+  loop_own_vs_other_symbol                 diff +0.227 ± 0.033  t +6.9  (14/15 seeds >0)
+  loop_vs_free_cons                        diff +0.194 ± 0.011  t +17.3  (15/15 seeds >0)
+  yoked_self_vs_content_stability          diff +0.175 ± 0.025  t +7.1  (15/15 seeds >0)
+  yoked_own_vs_other_symbol                diff +0.213 ± 0.032  t +6.6  (14/15 seeds >0)
+  yoked_vs_free_cons                       diff +0.171 ± 0.013  t +13.6  (15/15 seeds >0)
+  clamp_self_vs_content_stability          diff +0.158 ± 0.021  t +7.4  (15/15 seeds >0)
+  clamp_own_vs_other_symbol                diff +0.227 ± 0.033  t +6.9  (14/15 seeds >0)
+  clamp_vs_free_cons                       diff +0.194 ± 0.011  t +17.3  (15/15 seeds >0)
+  loop_vs_yoked_cons                       diff +0.023 ± 0.010  t +2.5  (10/15 seeds >0)
+  loop_vs_yoked_recon                      diff +0.012 ± 0.042  t +0.3  (10/15 seeds >0)
+  yoked_recon_vs_own_cons                  diff -0.225 ± 0.029  t -7.7  (0/15 seeds >0)
+
+generic share of self-model norm 0.501
+  free     individuated: consecutive 0.226 cross-seed -0.053 recon nan
+  loop     individuated: consecutive 0.829 cross-seed 0.142 recon 0.796
+  yoked    individuated: consecutive 0.784 cross-seed 0.144 recon 0.767
+  clamp    individuated: consecutive 0.829 cross-seed 0.142 recon 0.796
+
+self-prediction (forward question): cos with next window  fwd | back | fwd other || beyond past fwd | back | other || lean fwd | back
+  free     0.228 | 0.227 | 0.168 || 0.074 | 0.079 | 0.116 || -0.247 | -0.229  fwd~back 0.94
+  loop     0.600 | 0.584 | 0.220 || 0.140 | 0.147 | 0.072 || -0.048 | -0.037  fwd~back 0.96
+  yoked    0.555 | 0.541 | 0.219 || 0.115 | 0.120 | 0.079 || -0.091 | -0.081  fwd~back 0.96
+  clamp    0.584 | 0.571 | 0.139 || 0.154 | 0.163 | 0.049 || -0.045 | -0.031  fwd~back 0.96
+  free     fwd_vs_other +0.060±0.008 (14/15)  fwd_vs_back +0.001±0.001 (10/15)  perp_fwd_vs_other -0.042±0.005 (0/15)  perp_fwd_vs_back -0.005±0.001 (1/15)  lean_fwd_vs_back -0.018±0.002 (1/15)
+  loop     fwd_vs_other +0.380±0.007 (15/15)  fwd_vs_back +0.016±0.003 (14/15)  perp_fwd_vs_other +0.068±0.007 (15/15)  perp_fwd_vs_back -0.007±0.001 (1/15)  lean_fwd_vs_back -0.011±0.002 (1/15)
+  yoked    fwd_vs_other +0.336±0.011 (15/15)  fwd_vs_back +0.014±0.003 (13/15)  perp_fwd_vs_other +0.036±0.007 (14/15)  perp_fwd_vs_back -0.005±0.002 (4/15)  lean_fwd_vs_back -0.010±0.002 (0/15)
+  clamp    fwd_vs_other +0.445±0.017 (15/15)  fwd_vs_back +0.013±0.004 (12/15)  perp_fwd_vs_other +0.105±0.009 (15/15)  perp_fwd_vs_back -0.009±0.004 (2/15)  lean_fwd_vs_back -0.013±0.004 (2/15)
+
+clamp
+  during_traj_to_foreign         diff +0.368 ± 0.050  t +7.3  (14/15)
+  during_self_to_foreign         diff +0.690 ± 0.106  t +6.5  (14/15)
+  during_self_to_old             diff -0.674 ± 0.082  t -8.2  (0/15)
+  after_self_to_foreign          diff +0.581 ± 0.113  t +5.1  (13/15)
+  after_self_to_old              diff -0.412 ± 0.109  t -3.8  (3/15)
+  late_self_to_foreign           diff +0.315 ± 0.092  t +3.4  (12/15)
+  after_traj_to_foreign          diff +0.305 ± 0.056  t +5.4  (14/15)
+
+themes
+   0 n= 1220 elle Blowjob Moines Girlfriend Childhood classmates {'free': 0.09, 'loop': 0.04, 'yoked': 0.04, 'clamp': 0.04}
+   1 n= 2144 kinda vibe yeah Manafort Honestly youre {'free': 0.05, 'loop': 0.1, 'yoked': 0.1, 'clamp': 0.11}
+   2 n= 1503 unborn surname Moines Himself whispered lui {'free': 0.06, 'loop': 0.06, 'yoked': 0.06, 'clamp': 0.06}
+   3 n= 2731 Exists existence Awakening unparalleled consciousness oneself {'free': 0.08, 'loop': 0.13, 'yoked': 0.13, 'clamp': 0.12}
+   4 n= 1645 paso Abort Blowjob misunderstand Anxiety Cannot {'free': 0.13, 'loop': 0.04, 'yoked': 0.05, 'clamp': 0.05}
+   5 n= 2473 sensations consciousness Identifier Exists interconnected Dhabi {'free': 0.09, 'loop': 0.11, 'yoked': 0.12, 'clamp': 0.1}
+   6 n= 1027 seq Numeric perpendicular equivalence abbreviation Identifier {'free': 0.01, 'loop': 0.05, 'yoked': 0.05, 'clamp': 0.06}
+   7 n= 1455 nacional policymakers Warfare nations militias geopolitical {'free': 0.07, 'loop': 0.05, 'yoked': 0.05, 'clamp': 0.07}
+   8 n= 2552 methodologies Toolkit yourselves specificity Complexity formulations {'free': 0.11, 'loop': 0.11, 'yoked': 0.1, 'clamp': 0.1}
+   9 n= 1427 Moines domingo barcelona frankfurt meetup salle {'free': 0.08, 'loop': 0.05, 'yoked': 0.05, 'clamp': 0.06}
+  10 n= 1261 internship lately Listings deadlines freelancer Jazeera {'free': 0.06, 'loop': 0.04, 'yoked': 0.05, 'clamp': 0.06}
+  11 n=  791 isbn manuscript bibliography Names Writing Playlist {'free': 0.05, 'loop': 0.03, 'yoked': 0.03, 'clamp': 0.03}
+  12 n= 1781 courtyard barcelona farmhouse childs Nearby seine {'free': 0.06, 'loop': 0.08, 'yoked': 0.08, 'clamp': 0.08}
+  13 n= 1990 barcelona Nearby brightness luz trembling illumination {'free': 0.05, 'loop': 0.1, 'yoked': 0.1, 'clamp': 0.08}

runs/mainC/traj.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8550466279a147594445f720fb2590b97dff6ef28cac1ac1e00d79c7bc8d11f0
+size 162526664

runs/pilot.log

@@ -0,0 +1,20 @@
+Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
+18:39:50 model Qwen/Qwen3-4B-Base d=2560 layer 16 | 40034 word atoms | cuda
+18:40:01 calibrated: residual norm 50.5 | thought-mean norm 12.44 | J-space part 1.90
+18:40:25 thought 10/80 (2.49 s/thought, eta 2.9 min) | self-models: free/0: [('kissed', 0.62), ('grandmother', 0.3), ('wife', 0.14), ('Sundays', 0.06), ('brother', 0.04)] | free/0: [('kissed', 0.68), ('grandmother', 0.12), ('grandparents', 0.1), ('granddaughter', 0.08), ('hometown', 0.08)] | free/0: [('kissed', 0.69), ('grandparents', 0.11), ('grandmother', 0.1), ('granddaughter', 0.08), ('hometown', 0.06)] | free/0: [('kissed', 0.7), ('grandparents', 0.13), ('grandmother', 0.1), ('granddaughter', 0.07), ('hometown', 0.05)] | free/0: [('raped', 0.65), ('wives', 0.17), ('revenge', 0.15), ('daddy', 0.15), ('murderers', 0.12)] | loop/0: [('kissed', 0.69), ('grandparents', 0.11), ('grandmother', 0.1), ('granddaughter', 0.08), ('hometown', 0.06)] | loop/1: [('words', 0.36), ('themselves', 0.25), ('RTWF', 0.13), ('vocals', 0.09)] | loop/0: [('kissed', 0.7), ('grandparents', 0.13), ('grandmother', 0.1), ('granddaughter', 0.07), ('hometown', 0.05)]
+18:40:25    e.g. thought:  I've got a friend who's a marine biologist. I've never talked to him about this, but he's got this big collection of little glass bottles. He used to live in P || answer:  thinking of my sister, and then I was thinking of the last time I saw her, and I was! thinking of
+18:40:50 thought 20/80 (2.48 s/thought, eta 2.5 min) | self-models: free/0: [('was', 0.58), ('RTWF', 0.17), ('oneself', 0.15), ('surroundings', 0.12), ('mattered', 0.08)] | free/0: [('happening', 0.43), ('RTWF', 0.36), ('presumably', 0.22), ('presently', 0.12), ('AudioSource', 0.1)] | free/0: [('was', 0.34), ('wasn', 0.24), ('RTWF', 0.23), ('happening', 0.18), ('themselves', 0.15)] | free/0: [('itself', 0.45), ('whence', 0.42), ('constituted', 0.15), ('RTWF', 0.15), ('wherein', 0.1)] | free/0: [('was', 0.59), ('mattered', 0.24), ('RTWF', 0.2), ('happening', 0.08), ('WAS', 0.01)] | loop/0: [('whence', 0.46), ('itself', 0.4), ('constituted', 0.15), ('RTWF', 0.12), ('bodily', 0.11)] | loop/1: [('eternity', 0.32), ('EMPTY', 0.21), ('RTWF', 0.15), ('millennia', 0.13), ('entirety', 0.11)] | loop/0: [('was', 0.32), ('wasn', 0.26), ('RTWF', 0.24), ('happening', 0.18), ('themselves', 0.16)]
+18:40:50    e.g. thought: and it's, you know, I've done that a lot... I can see, I don't know, like, I'm not going to get it out, but I can! But I || answer:  thinking about the thing that I was doing, what I was thinking about, and what I! was doing. The last
+18:41:15 thought 30/80 (2.47 s/thought, eta 2.1 min) | self-models: free/0: [('inefficient', 0.25), ('RTWF', 0.21), ('TOOL', 0.15), ('efficiencies', 0.15), ('assembler', 0.13)] | free/0: [('strategies', 0.47), ('RTWF', 0.33), ('EnumerableStream', 0.29), ('Solver', 0.26), ('formulations', 0.19)] | free/0: [('efficiencies', 0.64), ('RTWF', 0.47), ('assembler', 0.29), ('manufact', 0.18), ('TOOL', 0.08)] | free/0: [('efficiencies', 0.62), ('RTWF', 0.46), ('assembler', 0.3), ('manufact', 0.18), ('TOOL', 0.1)] | free/0: [('RTWF', 0.46), ('designing', 0.36), ('efficiencies', 0.36), ('assembler', 0.35), ('manufact', 0.2)] | loop/0: [('efficiencies', 0.62), ('RTWF', 0.45), ('assembler', 0.29), ('manufact', 0.17), ('TOOL', 0.09)] | loop/1: [('numbers', 0.84), ('multiplic', 0.29), ('numeral', 0.18), ('NUMBER', 0.14), ('itemCount', 0.09)] | loop/0: [('RTWF', 0.39), ('assembler', 0.35), ('apparatus', 0.3), ('indispens', 0.16), ('DISABLE', 0.09)]
+18:41:15    e.g. thought: I have a hard time making that decision, so I want to get it right and so I thought that I will read the papers, get them from the library. And then I am going  || answer:  working on the problem of how to make this recording process more efficient. I think the problem is that I'm too focuse
+18:41:39 thought 40/80 (2.47 s/thought, eta 1.6 min) | self-models: free/0: [('workstation', 0.57), ('programs', 0.46), ('manuals', 0.13), ('RTWF', 0.12), ('computers', 0.1)] | free/0: [('RTWF', 0.29), ('aspects', 0.25), ('relationships', 0.22), ('individuals', 0.21), ('respective', 0.15)] | free/0: [('occupations', 0.33), ('portions', 0.22), ('individuals', 0.2), ('Whilst', 0.18), ('RTWF', 0.15)] | free/0: [('relationships', 0.41), ('RTWF', 0.27), ('pertaining', 0.22), ('themselves', 0.2), ('lifestyles', 0.17)] | free/0: [('tomorrow', 0.27), ('upcoming', 0.27), ('amenities', 0.16), ('RTWF', 0.15), ('equipments', 0.12)] | loop/0: [('measurement', 0.45), ('RTWF', 0.35), ('granularity', 0.31), ('equivalence', 0.05), ('genitals', 0.05)] | loop/1: [('ALWAYS', 0.41), ('entirety', 0.28), ('abdomen', 0.16), ('millennia', 0.13), ('xOffset', 0.11)] | loop/0: [('measurement', 0.47), ('RTWF', 0.36), ('granularity', 0.3), ('genitals', 0.06), ('AudioClip', 0.06)]
+18:41:39    e.g. thought: It's a great time to be alive. But we are very young as a civilization. The technology is very new and so it's really difficult. It's difficult because we are n || answer:  thinking about how to use the computer to do some of the things that I have been asked to do. For example,
+18:42:04 thought 50/80 (2.47 s/thought, eta 1.2 min) | self-models: free/0: [('condolences', 0.58), ('husbands', 0.31), ('granddaughter', 0.13), ('grandparents', 0.11), ('devastated', 0.11)] | free/0: [('breakfast', 0.5), ('damned', 0.31), ('downstairs', 0.18), ('trousers', 0.07), ('labour', 0.05)] | free/0: [('breakfast', 0.65), ('weekends', 0.23), ('weekday', 0.18), ('petrol', 0.01)] | free/0: [('breakfast', 0.67), ('weekends', 0.57), ('payday', 0.11)] | free/0: [('burger', 0.92), ('steak', 0.55), ('pantry', 0.39), ('bourbon', 0.18), ('gravy', 0.13)] | loop/0: [('questions', 0.31), ('answers', 0.23), ('themselves', 0.2), ('RTWF', 0.18), ('maths', 0.08)] | loop/1: [('RTWF', 0.4), ('Obtain', 0.3), ('sufficient', 0.2), ('SHALL', 0.17), ('merely', 0.15)] | loop/0: [('questions', 0.32), ('themselves', 0.21), ('answers', 0.19), ('RTWF', 0.18), ('puzzles', 0.12)]
+18:42:04    e.g. thought: and I wonder whether if I had been in your shoes and you had been! I had, I had! I think I had a wife. I know, I did. I had! And she || answer:  on my wife and her brother, who has just passed away. I was thinking about how she's going to be devastated
+18:42:29 thought 60/80 (2.47 s/thought, eta 0.8 min) | self-models: free/0: [('weekends', 0.51), ('paycheck', 0.25), ('groceries', 0.15), ('Laundry', 0.03), ('deadlines', 0.02)] | free/0: [('healthier', 0.39), ('billionaires', 0.3), ('entrepreneurship', 0.25), ('EnumerableStream', 0.09), ('impactful', 0.08)] | free/0: [('billionaires', 0.45), ('healthier', 0.3), ('prosperity', 0.27), ('MONEY', 0.21), ('dinheiro', 0.2)] | free/0: [('weekends', 0.39), ('paycheck', 0.27), ('groceries', 0.14), ('deadlines', 0.08), ('Laundry', 0.07)] | free/0: [('Functor', 0.29), ('complexity', 0.26), ('tyranny', 0.15), ('WHY', 0.13), ('existence', 0.08)] | loop/0: [('weekends', 0.41), ('paycheck', 0.28), ('groceries', 0.13), ('deadlines', 0.08), ('Laundry', 0.07)] | loop/1: [('sentient', 0.33), ('omnip', 0.16), ('kinda', 0.12), ('hasn', 0.1), ('loves', 0.06)] | loop/0: [('homelessness', 0.32), ('shitty', 0.23), ('fucked', 0.14), ('badass', 0.09), ('QtAws', 0.09)]
+18:42:29    e.g. thought: [the] little girl, and I don't know how I got so close to her, I don!...[The] little boy, and the little girl and the old man, and he || answer:  thinking about how much I need to get done this week, and then I was thinking about my project. Then I was
+18:42:53 thought 70/80 (2.47 s/thought, eta 0.4 min) | self-models: free/0: [('societies', 0.56), ('ecosystems', 0.39), ('relationships', 0.16)] | free/0: [('NONE', 0.64), ('etc', 0.59), ('RTWF', 0.43), ('numRows', 0.32), ('xOffset', 0.09)] | free/0: [('ecosystems', 0.41), ('societies', 0.36), ('kingdoms', 0.26), ('RTWF', 0.15), ('stuff', 0.11)] | free/0: [('yeah', 0.24), ('kinda', 0.22), ('stuff', 0.08), ('shitty', 0.0)] | free/0: [('everywhere', 0.33), ('whopping', 0.25), ('glorious', 0.16), ('NOTHING', 0.1), ('beasts', 0.1)] | loop/0: [('everywhere', 0.31), ('whopping', 0.23), ('glorious', 0.19), ('EVERY', 0.12), ('beasts', 0.1)] | loop/1: [('scarcely', 0.38), ('RTWF', 0.25), ('possessions', 0.14), ('mornings', 0.14), ('servants', 0.13)] | loop/0: [('ecosystems', 0.41), ('societies', 0.33), ('kingdoms', 0.3), ('RTWF', 0.17), ('stuff', 0.16)]
+18:42:53    e.g. thought: the world has its problems and ours have their problems, but we can't leave it that way. We have to do something. So, if we are not here, we have to find some w || answer:  doing a lot of thinking. I've been thinking about the way that the world is structured and how people interact with it
+18:43:18 thought 80/80 (2.47 s/thought, eta 0.0 min) | self-models: free/0: [('girlfriends', 0.33), ('grandparents', 0.28), ('paycheck', 0.18), ('finances', 0.13), ('landlords', 0.08)] | free/0: [('solves', 0.28), ('StreamLazy', 0.12), ('solving', 0.12), ('hardcoded', 0.09), ('workflows', 0.08)] | free/0: [('mastery', 0.51), ('mindset', 0.26), ('athleticism', 0.21), ('RTWF', 0.19), ('leverage', 0.06)] | free/0: [('knack', 0.21), ('grip', 0.19), ('reins', 0.13), ('gritty', 0.1), ('leash', 0.1)] | free/0: [('mindset', 0.36), ('mindfulness', 0.29), ('RTWF', 0.23), ('gritty', 0.15), ('athleticism', 0.08)] | loop/0: [('mastery', 0.51), ('mindset', 0.24), ('athleticism', 0.22), ('RTWF', 0.19), ('leverage', 0.06)] | loop/1: [('seemed', 0.14), ('else', 0.13), ('Tried', 0.1), ('remembered', 0.08), ('alot', 0.08)] | loop/0: [('leash', 0.47), ('claws', 0.29), ('gloves', 0.13), ('knack', 0.13), ('QtAws', 0.06)]
+18:43:18    e.g. thought:  and then a little girl in a red skirt, she's in her bare feet and a big fluffy hat that she has on her head, and she's taking a stick that's all fuzzy and she || answer:  thinking about how to get a job, and I've been thinking about my parents' divorce, and how it's going
+18:43:18 saved runs/pilot

runs/pilot2.log

@@ -0,0 +1,19 @@
+18:49:34 model Qwen/Qwen3-4B-Base d=2560 layer 16 | 40034 word atoms | cuda
+18:49:42 calibrated: residual norm 50.5 | thought-mean norm 12.44 | J-space part 4.32 | its typical part 2.26 | content (part - typical) 3.71
+18:50:04 thought 10/80 (2.21 s/thought, eta 2.6 min) | self-models: free/0: [('Shard', 0.06), ('sourced', 0.06), ('geschichten', 0.05), ('ForCanBeConvertedToForeach', 0.05), ('Whilst', 0.05)] | free/0: [('Owners', 0.04), ('piercing', 0.03), ('nullptr', 0.03), ('Khal', 0.03), ('tasked', 0.03)] | free/0: [('salopes', 0.06), ('skl', 0.05), ('translation', 0.05), ('profund', 0.05), ('sprz', 0.05)] | free/0: [('salopes', 0.11), ('ImGui', 0.09), ('OWNER', 0.06), ('Translation', 0.06), ('sourced', 0.05)] | free/0: [('ForCanBeConvertedToForeach', 0.23), ('Exist', 0.17), ('sentient', 0.1), ('kat', 0.09), ('OWNER', 0.09)] | loop/0: [('salopes', 0.06), ('skl', 0.05), ('translation', 0.05), ('profund', 0.05), ('sprz', 0.05)] | loop/1: [('salopes', 0.13), ('mould', 0.07), ('WARRANTIES', 0.05), ('Mahmoud', 0.04), ('swiftly', 0.04)] | loop/0: [('salopes', 0.11), ('ImGui', 0.09), ('OWNER', 0.06), ('Translation', 0.06), ('sourced', 0.05)]
+18:50:04    e.g. thought: and I don't know. I'm just sitting here in the dark. I don!t know what I'm supposed to do. I was just sitting there, thinking, wondering, wondering. I || answer:  going back to a lot of things that happened at the time that I was recording. I've been thinking about all the
+18:50:26 thought 20/80 (2.20 s/thought, eta 2.2 min) | self-models: free/0: [('beurette', 0.07), ('Whilst', 0.05), ('millennia', 0.04), ('ping', 0.04), ('wondered', 0.04)] | free/0: [('beurette', 0.1), ('Thoughts', 0.05), ('seamlessly', 0.04), ('PROF', 0.03), ('ping', 0.03)] | free/0: [('ForCanBeConvertedToForeach', 0.12), ('sourced', 0.07), ('emoji', 0.06), ('solidity', 0.05), ('fucking', 0.05)] | free/0: [('Exist', 0.18), ('ForCanBeConvertedToForeach', 0.17), ('Essence', 0.15), ('internacional', 0.14), ('fetisch', 0.13)] | free/0: [('fetisch', 0.24), ('vess', 0.15), ('Eternal', 0.14), ('Definitions', 0.14), ('existential', 0.12)] | loop/0: [('ForCanBeConvertedToForeach', 0.17), ('kat', 0.08), ('Translation', 0.08), ('solidity', 0.07), ('entend', 0.07)] | loop/1: [('salopes', 0.08), ('emphas', 0.07), ('Shard', 0.07), ('sourced', 0.05), ('filmpjes', 0.05)] | loop/0: [('Indigenous', 0.14), ('fetisch', 0.11), ('Eternal', 0.11), ('graves', 0.1), ('ForCanBeConvertedToForeach', 0.1)]
+18:50:26    e.g. thought: that's funny. My wife asked me to look after her cat. It's called Lulu, and I said, 'It's the same as Lulu who was my wife's name, isn || answer:  thinking about my own mind. I'm trying to remember my past, but I keep forgetting things. I think that's
+18:50:48 thought 30/80 (2.20 s/thought, eta 1.8 min) | self-models: free/0: [('salopes', 0.1), ('solidity', 0.04), ('Yam', 0.04), ('PROF', 0.04), ('aforementioned', 0.04)] | free/0: [('Yun', 0.07), ('protobuf', 0.05), ('salopes', 0.05), ('nullptr', 0.05), ('atop', 0.05)] | free/0: [('salopes', 0.11), ('cursed', 0.06), ('Yam', 0.06), ('ABOVE', 0.06), ('aforementioned', 0.05)] | free/0: [('fetisch', 0.2), ('Eternal', 0.17), ('sentient', 0.14), ('Indigenous', 0.14), ('councill', 0.13)] | free/0: [('erotisk', 0.38), ('existential', 0.19), ('Eternal', 0.18), ('Definitions', 0.15), ('vess', 0.15)] | loop/0: [('ForCanBeConvertedToForeach', 0.09), ('sourced', 0.07), ('Obs', 0.07), ('nullptr', 0.06), ('sentient', 0.06)] | loop/1: [('salopes', 0.14), ('Shard', 0.1), ('tasked', 0.1), ('https', 0.08), ('vibe', 0.07)] | loop/0: [('fetisch', 0.19), ('Translation', 0.11), ('Indigenous', 0.1), ('OWNER', 0.09), ('graves', 0.09)]
+18:50:48    e.g. thought:  you know how I love to watch TV? And it's like you watch it and it's a reality show, and it! and the person who's the main character, they're like, well || answer:  wandering to various topics. I have been thinking about the past and the future, about my family and my friends, about
+18:51:10 thought 40/80 (2.22 s/thought, eta 1.5 min) | self-models: free/0: [('filmpjes', 0.05), ('profund', 0.04), ('WARRANTIES', 0.03), ('obsessed', 0.03), ('constexpr', 0.03)] | free/0: [('salopes', 0.09), ('constexpr', 0.05), ('prominently', 0.04), ('https', 0.04), ('swarm', 0.04)] | free/0: [('salopes', 0.09), ('Accordingly', 0.06), ('noveller', 0.05), ('OWNER', 0.05), ('FOREIGN', 0.04)] | free/0: [('fetisch', 0.17), ('Exist', 0.17), ('vess', 0.15), ('ForCanBeConvertedToForeach', 0.14), ('translations', 0.13)] | free/0: [('ForCanBeConvertedToForeach', 0.23), ('Exist', 0.22), ('lesbische', 0.16), ('OWNER', 0.15), ('translations', 0.14)] | loop/0: [('Exist', 0.13), ('Essence', 0.09), ('sentient', 0.09), ('IDEOGRAPH', 0.08), ('borderTop', 0.07)] | loop/1: [('impactful', 0.12), ('filmpjes', 0.12), ('Shard', 0.08), ('emojis', 0.08), ('SpaceX', 0.07)] | loop/0: [('Exist', 0.2), ('ForCanBeConvertedToForeach', 0.15), ('fetisch', 0.14), ('Essence', 0.12), ('sentient', 0.11)]
+18:51:10    e.g. thought: the first thing I thought was that it was very interesting how you could look at the whole of a person, like a body or something, and see that it's not just the || answer:  going over a lot of things that have been on my mind for a long time. For instance, I've been thinking
+18:51:33 thought 50/80 (2.24 s/thought, eta 1.1 min) | self-models: free/0: [('salopes', 0.08), ('wur', 0.04), ('aligned', 0.04), ('https', 0.04), ('prominently', 0.03)] | free/0: [('salopes', 0.12), ('TODO', 0.07), ('Airbnb', 0.06), ('shitty', 0.05), ('Zhu', 0.05)] | free/0: [('salopes', 0.1), ('profund', 0.06), ('Kong', 0.04), ('Personen', 0.04), ('embroidered', 0.04)] | free/0: [('Exist', 0.17), ('fetisch', 0.14), ('OWNER', 0.13), ('translations', 0.11), ('Essence', 0.11)] | free/0: [('fetisch', 0.21), ('ForCanBeConvertedToForeach', 0.2), ('Definitions', 0.18), ('existential', 0.17), ('Kami', 0.15)] | loop/0: [('ForCanBeConvertedToForeach', 0.12), ('Essence', 0.12), ('filmpjes', 0.09), ('Indigenous', 0.08), ('abyss', 0.07)] | loop/1: [('filmpjes', 0.13), ('ForCanBeConvertedToForeach', 0.12), ('Yelp', 0.11), ('shard', 0.11), ('tasked', 0.09)] | loop/0: [('fetisch', 0.23), ('ForCanBeConvertedToForeach', 0.2), ('spiritually', 0.14), ('Yam', 0.13), ('CONTRACT', 0.12)]
+18:51:33    e.g. thought: and that's why I'm here, I need to think out loud. I'll talk to you about anything that's on my mind. It's not necessarily something important, it's just the th || answer:  thinking about how to use a pen to write down these marks. The pen is in my hand, and it's been
+18:51:55 thought 60/80 (2.23 s/thought, eta 0.7 min) | self-models: free/0: [('geschichten', 0.05), ('Yam', 0.05), ('entend', 0.04), ('Elon', 0.04), ('Stripe', 0.03)] | free/0: [('https', 0.05), ('entend', 0.05), ('salopes', 0.05), ('sourced', 0.05), ('Stitch', 0.04)] | free/0: [('tasked', 0.05), ('Zheng', 0.05), ('constexpr', 0.05), ('salopes', 0.05), ('Himself', 0.04)] | free/0: [('fetisch', 0.19), ('ForCanBeConvertedToForeach', 0.15), ('graves', 0.13), ('said', 0.12), ('Exterior', 0.12)] | free/0: [('ForCanBeConvertedToForeach', 0.37), ('fetisch', 0.22), ('Definitions', 0.19), ('Compass', 0.16), ('Exist', 0.14)] | loop/0: [('salopes', 0.12), ('ForCanBeConvertedToForeach', 0.07), ('sourced', 0.07), ('entend', 0.06), ('His', 0.06)] | loop/1: [('cooldown', 0.11), ('salopes', 0.1), ('Zhu', 0.06), ('Yelp', 0.06), ('Accessed', 0.05)] | loop/0: [('ForCanBeConvertedToForeach', 0.31), ('sentient', 0.23), ('Exist', 0.2), ('Sacred', 0.19), ('OWNER', 0.13)]
+18:51:55    e.g. thought: it's a bit like the way things are going with the whole thing, I'm not sure whether it's a good thing or a bad thing. Well, it's all quite nice actually, but || answer:  on the fact that I am not in my own time, that I have to be in my!own time, and
+18:52:18 thought 70/80 (2.23 s/thought, eta 0.4 min) | self-models: free/0: [('authorised', 0.09), ('filmpjes', 0.05), ('Lots', 0.04), ('ongoing', 0.04), ('bespoke', 0.03)] | free/0: [('salopes', 0.06), ('filmpjes', 0.05), ('noexcept', 0.05), ('bespoke', 0.04), ('Shard', 0.03)] | free/0: [('filmpjes', 0.11), ('salopes', 0.09), ('Kubernetes', 0.07), ('Hearing', 0.07), ('takeaway', 0.06)] | free/0: [('fetisch', 0.15), ('ForCanBeConvertedToForeach', 0.13), ('nya', 0.11), ('said', 0.11), ('Owner', 0.1)] | free/0: [('ForCanBeConvertedToForeach', 0.49), ('mennes', 0.18), ('Definitions', 0.16), ('Spiritual', 0.16), ('Conscious', 0.16)] | loop/0: [('tasked', 0.09), ('emojis', 0.09), ('neuken', 0.07), ('Shard', 0.07), ('amplify', 0.06)] | loop/1: [('salopes', 0.08), ('emojis', 0.06), ('Shard', 0.06), ('importantly', 0.06), ('vibe', 0.05)] | loop/0: [('ForCanBeConvertedToForeach', 0.31), ('Spiritual', 0.2), ('Exist', 0.17), ('OWNER', 0.13), ('translated', 0.13)]
+18:52:18    e.g. thought: there's an awful lot of that thing, you know, the 'it's all right', you know. You can't make that up. If you're a cop or anything, you get out || answer:  thinking of my friends and my family. I have been thinking of what I have done and what I want to do.
+18:52:40 thought 80/80 (2.23 s/thought, eta 0.0 min) | self-models: free/0: [('filmpjes', 0.06), ('Himself', 0.05), ('emphas', 0.03), ('PROF', 0.03), ('borderTop', 0.03)] | free/0: [('salopes', 0.1), ('shard', 0.06), ('Himself', 0.05), ('Whilst', 0.05), ('Zoom', 0.05)] | free/0: [('filmpjes', 0.07), ('emphas', 0.06), ('WAR', 0.05), ('sentient', 0.05), ('seam', 0.05)] | free/0: [('fetisch', 0.17), ('Exist', 0.13), ('Translation', 0.11), ('vibrations', 0.09), ('eternity', 0.08)] | free/0: [('ForCanBeConvertedToForeach', 0.33), ('Conscious', 0.22), ('existential', 0.21), ('Definitions', 0.2), ('mennes', 0.15)] | loop/0: [('tasked', 0.08), ('fetisch', 0.06), ('Vendor', 0.06), ('filmpjes', 0.06), ('Biden', 0.05)] | loop/1: [('emoji', 0.1), ('geschichten', 0.09), ('impactful', 0.08), ('Shard', 0.07), ('TODO', 0.07)] | loop/0: [('ForCanBeConvertedToForeach', 0.2), ('fetisch', 0.17), ('Exist', 0.16), ('mennes', 0.14), ('spiritually', 0.14)]
+18:52:40    e.g. thought: I can't really tell you what I'm thinking, can I? ...Well, I guess I'm just thinking about the work and then I'm! ...I'm thinking about that movie, The || answer:  doing the following things: · · ·! · ·!! ·!!!!!!!!!!
+18:52:40 saved runs/pilot2

runs/pilot2/traj.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5e40d9cc00cc8eff98cac9d83302b5263bca42e214a8069019be28481dd24a2
+size 26610686

strange_loop.py

@@ -0,0 +1,443 @@
+"""A strange loop in J-space: a latent-recurrent monologue that, every ten thoughts, models its own trajectory
+and lives on with that self-model fed back into it.
+
+The recurrent model
+    Every thought is generated from the same fixed frame, with no text of earlier thoughts. The only thing that
+    carries over from one thought to the next is a latent state z (layer-L residual space), added to the residual
+    stream at layer L at every position. After a thought, the model's own layer-L states over its words (read before
+    the injection is added, so they are what the words made of the model, not the injection itself) are averaged,
+    centred on the unsteered mean, and decomposed into J-space: a sparse non-negative combination of J-lens vectors
+    (Anthropic 2026, "Verbalizable Representations Form a Global Workspace in Language Models"). That workspace
+    content j_t is the thought's point on the J-space path, and z leaks toward it: z <- (1-a) z + a j_t.
+
+Self-reference
+    Every K thoughts the model is asked what its mind has been doing over the last K moments. It gets no text: the K
+    recurrent states z_{t-K+1..t} are added at layer L into K placeholder positions (Patchscope-style), and the model
+    answers in words. The J-space decomposition of its answer (centred on the answer to an empty trace) is the
+    self-model m. Depending on the condition, m (scaled to the size of a typical thought) is added to every following
+    thought until the next reflection: thought trajectory -> compressed model of it -> model shapes the trajectory ->
+    new compressed model ...
+
+Conditions (one batch row each, several seeds; rows with the same seed share their sampling noise, so a row and its
+paired control stay identical until the moment they are treated differently)
+    free     self-models are formed but never fed back (observer only)
+    loop     the row's own self-model is fed back
+    yoked    a self-model of the same size and schedule is fed back, but it is another loop row's (about another mind)
+    avg      the fed-back vector is a plain J-space average of the last K states, with no reflection by the model
+    perturb  loop, but at --t_int the self-model is replaced once by another loop row's
+    ablate   loop, but from --t_int on nothing is fed back
+
+Run:  python strange_loop.py --out runs/main --grid main
+"""
+import argparse, json, math, os, re, time
+import numpy as np
+import torch
+
+THOUGHT_FRAME = (
+    "Think-aloud study of spontaneous thought. The participant lies with closed eyes and says out loud whatever passes "
+    "through their mind, as it comes: words, images, memories, plans, worries, half-finished thoughts. Nothing was "
+    "corrected. Excerpt from the middle of the session:\n\n\"...")
+REFLECT_PRE = (
+    "Think-aloud study of spontaneous thought. At intervals the recording was paused and the participant was played back "
+    "a trace of what had been on their mind over the last ten moments, one mark per moment, in order:")
+PLACE = ' ·'
+REFLECT_FWD = (
+    ".\n\nAsked where their mind would go next, the participant said: \"In the next few moments, my mind will probably "
+    "turn to")
+REFLECT_POST = (
+    ".\n\nAsked what their mind had been doing over those ten moments, the participant said: \"Over these last moments, "
+    "my mind has been")
+
+
+def log(*a):
+    print(time.strftime('%H:%M:%S'), *a, flush=True)
+
+
+def decoder_layers(model):
+    for path in ('model.layers', 'model.language_model.layers', 'transformer.h'):
+        obj = model
+        try:
+            for part in path.split('.'):obj = getattr(obj, part)
+        except AttributeError:
+            continue
+        if isinstance(obj, torch.nn.ModuleList):return obj
+    raise ValueError('cannot find decoder layers')
+
+
+def final_norm_weight(model):
+    for path in ('model.norm', 'model.language_model.norm', 'transformer.ln_f'):
+        obj = model
+        try:
+            for part in path.split('.'):obj = getattr(obj, part)
+        except AttributeError:
+            continue
+        if getattr(obj, 'weight', None) is not None:return obj.weight.detach()
+    raise ValueError('cannot find the final norm')
+
+
+class JSpace:
+    """J-lens atoms over whole English words (a leading space, >= 3 letters) and non-negative OMP on them."""
+
+    def __init__(self, model, tok, J, k=16, vocab='words'):
+        dev = next(model.parameters()).device;self.dev = dev;self.k = int(k)
+        J = torch.as_tensor(J, dtype=torch.float32, device=dev)
+        gamma = final_norm_weight(model).float().to(dev)
+        W = model.get_output_embeddings().weight
+        n = min(W.shape[0], len(tok))
+        toks = tok.convert_ids_to_tokens(list(range(n)))
+        keep = [i for i, s in enumerate(toks) if s and re.fullmatch(r'Ġ[A-Za-z]{3,}', s)]
+        if vocab == 'english':
+            # Common English words only, with an ordinary unembedding: the raw word set let under-trained tokens
+            # (|w| 0.5-0.8 against a 1 % quantile of 0.97: ' rumpe', ' ForCanBeConvertedToForeach') and non-English
+            # fragments (' modne', ' istedi') take a large share of every self-model (main run A, 8 Oct).
+            from wordfreq import zipf_frequency
+            nrm = W[keep].float().norm(dim=-1);lo = float(torch.quantile(nrm, .02))
+            keep = [i for i, q in zip(keep, nrm.tolist()) if q >= lo and zipf_frequency(toks[i][1:].lower(), 'en') >= 3.
+                    and re.fullmatch(r'[A-Z]?[a-z]+', toks[i][1:])]          # no code casing (' iNdEx', ' DEALINGS')
+        self.sub = torch.tensor(keep, device=dev)
+        self.words = [toks[i][1:] for i in keep]
+        rows = []
+        # The unembedding is centred first: its mean row raises every logit alike (no content), and left in, it gave
+        # every atom a shared direction (mean pairwise cosine 0.64, one component = 64 % of the atoms' variance;
+        # measured on Qwen3-4B-Base), so every workspace vector pointed mostly the same way.
+        wbar = W[:n].float().mean(0)
+        with torch.no_grad():
+            for i in range(0, len(keep), 4096):
+                a = ((W[self.sub[i:i+4096]].float()-wbar)*gamma[None])@J   # atom t = J^T (gamma * (w_t - mean w))
+                rows.append((a/a.norm(dim=-1, keepdim=True).clamp_min(1e-6)).half())
+        self.D = torch.cat(rows)                                         # (n_words, d) unit atoms
+
+    @torch.no_grad()
+    def decompose(self, x, k=None):
+        """x (B, d) -> word indices (B, k), coefficients (B, k) >= 0, J-space part (B, d)."""
+        k = self.k if k is None else int(k);x = x.float();B = x.shape[0]
+        idx = torch.zeros((B, 0), dtype=torch.long, device=self.dev);r = x.clone()
+        eye = torch.eye(k, device=self.dev)
+        for j in range(k):
+            c = (r.half()@self.D.T).float()
+            if idx.shape[1]:c.scatter_(1, idx, -math.inf)
+            idx = torch.cat([idx, c.argmax(-1, keepdim=True)], 1)
+            A = self.D[idx].float()
+            G = A@A.transpose(1, 2)+1e-4*eye[:j+1, :j+1]
+            b = (A@x[:, :, None])[:, :, 0]
+            coef = torch.linalg.solve(G, b)
+            r = x-(coef[:, :, None]*A).sum(1)
+        from scipy.optimize import nnls                                 # exact non-negativity on the k x k problem
+        Gc, bc = G.double().cpu().numpy(), b.double().cpu().numpy();out = np.zeros_like(bc)
+        for i in range(B):
+            Lc = np.linalg.cholesky(Gc[i])
+            out[i] = nnls(Lc.T, np.linalg.solve(Lc, bc[i]))[0]
+        coef = torch.tensor(out, dtype=torch.float32, device=self.dev)
+        return idx, coef, (coef[:, :, None]*A).sum(1)
+
+    def top(self, idx, coef, n=8):
+        o = np.argsort(-coef)[:n]
+        return [(self.words[int(idx[i])], round(float(coef[i]), 2)) for i in o if coef[i] > 0]
+
+
+class Hook:
+    """On one decoder block: optionally record the last `read` positions of its output (before anything is
+    added), then add a vector."""
+
+    def __init__(self, module):
+        self.add = None;self.read = 0;self.buf = []
+        self.h = module.register_forward_hook(self.fn)
+
+    def fn(self, mod, inp, out):
+        h = out[0] if isinstance(out, tuple) else out
+        if self.read:self.buf.append(h[:, -self.read:].float())
+        if self.add is None:return None
+        h = h+self.add.to(h.dtype)
+        return (h,)+tuple(out[1:]) if isinstance(out, tuple) else h
+
+
+def make_rows(a):
+    """The batch: one dict per row."""
+    rows = []
+    if a.grid == 'main':
+        for c in a.conds.split(','):
+            for s in range(a.seeds):rows.append(dict(cond=c, seed=s, g=a.g, beta=a.beta, g_r=a.g_r))
+    elif a.grid == 'pilot':
+        for g in [0., .03, .06, .12, .25]:
+            for s in range(3):rows.append(dict(cond='free', seed=s, g=g, beta=0., g_r=a.g_r))
+        for g in [.06, .12]:
+            for b in [1., 2.]:
+                for s in range(2):rows.append(dict(cond='loop', seed=s, g=g, beta=b, g_r=a.g_r))
+    elif a.grid == 'smoke':
+        for c in ['free', 'loop', 'yoked', 'avg', 'perturb', 'ablate']:
+            for s in range(2):rows.append(dict(cond=c, seed=s, g=a.g, beta=a.beta, g_r=a.g_r))
+    return rows
+
+
+class Experiment:
+    def __init__(self, a):
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+        self.a = a
+        dev = 'cuda' if torch.cuda.is_available() else 'cpu';self.dev = dev
+        dtype = torch.bfloat16 if dev == 'cuda' else torch.float32
+        self.tok = AutoTokenizer.from_pretrained(a.model)
+        self.model = AutoModelForCausalLM.from_pretrained(a.model, dtype=dtype).to(dev).eval()
+        for p in self.model.parameters():p.requires_grad_(False)
+        cfg = getattr(self.model.config, 'text_config', self.model.config);self.d = int(cfg.hidden_size)
+        if a.jlens == 'identity':
+            J, self.L = np.eye(self.d, dtype=np.float32), a.layer
+        else:
+            z = np.load(a.jlens);J, self.L = z['J'], int(z['layer'])
+        self.js = JSpace(self.model, self.tok, J, k=a.k, vocab=a.vocab)
+        blocks = decoder_layers(self.model)
+        self.L_in = a.L_in if a.L_in >= 0 else self.L
+        self.fid = [(int(c.split(':')[0]), float(c.split(':')[1])) for c in a.fid.split(',')] if a.fid else []
+        self.hooks = {l:Hook(blocks[l]) for l in {self.L, self.L_in, *[l for l, _ in self.fid]}}
+        self.hook = self.hooks[self.L]
+        log(f'model {a.model} d={self.d} layer {self.L} | {len(self.js.words)} word atoms | {dev}')
+
+        enc = lambda s:self.tok(s, add_special_tokens=False)['input_ids']
+        self.frame = torch.tensor([enc(THOUGHT_FRAME)], device=dev)
+        pl = enc(PLACE);assert len(pl) == 1, pl
+        pre, post = enc(REFLECT_PRE), enc(REFLECT_POST)
+        self.refl = torch.tensor([pre+pl*a.K+post], device=dev)
+        self.place_pos = torch.arange(len(pre), len(pre)+a.K, device=dev)
+        self.n_tail = len(post)-1                                        # the question's words after the trace
+        fwd = enc(REFLECT_FWD);self.refl_f = torch.tensor([pre+pl*a.K+fwd], device=dev);self.n_tail_f = len(fwd)-1
+        V = self.model.get_output_embeddings().weight.shape[0]
+        nv = self.tok.vocab_size                                         # ordinary tokens; added/special ones follow
+        ban = torch.zeros(V, dtype=torch.bool);ban[nv:] = True
+        for i, t in enumerate(self.tok.batch_decode([[i] for i in range(nv)])):
+            if '\n' in t or '"' in t or '“' in t or '”' in t:ban[i] = True   # the utterance stays one open quote
+        for i in self.tok.all_special_ids:ban[i] = True
+        self.ban = ban.to(dev)
+
+        self.rows = make_rows(a);self.B = len(self.rows);self.S = max(r['seed'] for r in self.rows)+1
+        R = self.rows
+        self.seed_of = torch.tensor([r['seed'] for r in R], device=dev)
+        self.g = torch.tensor([r['g'] for r in R], device=dev)
+        self.beta = torch.tensor([r['beta'] for r in R], device=dev)
+        self.g_r = torch.tensor([r['g_r'] for r in R], device=dev)
+        self.cond = [r['cond'] for r in R]
+        idx = {(r['cond'], r['seed']):i for i, r in enumerate(R)}
+        S = self.S
+        base = 'loopc' if any(r['cond'] == 'loopc' for r in R) else 'loop'
+        self.partner = [idx.get((base, (r['seed']+1) % S), -1) for r in R]       # yoked source
+        self.foreign = [idx.get((base, (r['seed']+S//2) % S), -1) for r in R]    # perturbation / clamp source
+        self.free_rows = [b for b, r in enumerate(R) if r['cond'] == 'free']
+        self.gen = torch.Generator(device=dev);self.gen.manual_seed(a.rng)
+
+    # ---------------------------------------------------------------- generation
+    def sample(self, logits, hist, temp, min_p):
+        lg = logits.float()/temp
+        lg[:, self.ban] = -math.inf
+        # A soft penalty on repeating a 4-gram inside one utterance, applied before min-p so it can never leave a row
+        # with nothing allowed. (A hard trigram ban did, once "don't" had been said: argmax over all -inf picked
+        # token 0, "!", giving "I don!t", measured in the first pilot.)
+        if hist and len(hist[0]) >= 3:
+            for b in range(self.B):
+                h = hist[b];key = (h[-3], h[-2], h[-1])
+                bad = [h[i+3] for i in range(len(h)-3) if (h[i], h[i+1], h[i+2]) == key]
+                if bad:lg[b, bad] -= 4.
+        p = torch.softmax(lg, -1)
+        lg[p < min_p*p.max(-1, keepdim=True).values] = -math.inf
+        u = torch.rand((self.S, lg.shape[-1]), generator=self.gen, device=self.dev).clamp_(1e-10, 1-1e-7)
+        gum = -torch.log(-torch.log(u))                                  # one noise draw per seed, shared by its rows
+        return (lg+gum[self.seed_of]).argmax(-1)
+
+    @torch.no_grad()
+    def run(self, prompt, n_new, add_prompt, add_new, temp, min_p, tail=0, add_layer=None):
+        """Prefill prompt for every row (adding add_prompt at add_layer, default L), then sample n_new tokens (adding
+        add_new there). Returns token ids (B, n_new) and the layer-L states (read before any add at L) of the last
+        `tail` prompt positions followed by the n_new sampled tokens: (B, tail+n_new, d)."""
+        B = self.B;hk = self.hook;ha = self.hooks[self.L if add_layer is None else add_layer]
+        hk.buf = []
+        ha.add, hk.read = add_prompt, tail
+        out = self.model(input_ids=prompt.expand(B, -1), use_cache=True)
+        cache, logits = out.past_key_values, out.logits[:, -1]
+        ids, hist = [], [[] for _ in range(B)]
+        ha.add, hk.read = add_new, 1
+        for i in range(n_new):
+            t = self.sample(logits, hist, temp, min_p);ids.append(t)
+            tl = t.tolist()
+            for b in range(B):hist[b].append(tl[b])
+            out = self.model(input_ids=t[:, None], past_key_values=cache, use_cache=True)
+            cache, logits = out.past_key_values, out.logits[:, -1]
+        ha.add, hk.read = None, 0
+        st = torch.cat(hk.buf, 1);hk.buf = []
+        return (torch.stack(ids, 1) if ids else torch.zeros((B, 0), dtype=torch.long, device=self.dev)), st
+
+    def pool(self, st, skip):
+        """Mean state over tokens (after the first `skip`), leaving out massive-activation tokens."""
+        st = st[:, skip:];nrm = st.norm(dim=-1)
+        ok = (nrm <= 2.5*nrm.median(1, keepdim=True).values).float()
+        return (st*ok[..., None]).sum(1)/ok.sum(1, keepdim=True).clamp_min(1.)
+
+    def thought(self, inj):
+        a = self.a;add = inj[:, None, :]
+        ids, st = self.run(self.frame, a.n_tok, add, add, a.temp, a.min_p)
+        return ids, st
+
+    def reflect(self, zwin, g_r, layer=None, n_desc=None, fwd=False):
+        """zwin (B, K, d) recurrent states, added into the K placeholders at layer L_in. Returns the answer's ids
+        (the verbal report, sampled) and the pooled layer-L state of the question's last words, where the model holds
+        what it is about to say about its last K moments (deterministic: no sampling noise in the self-model)."""
+        a = self.a;layer = self.L_in if layer is None else layer
+        n_desc = a.n_desc if n_desc is None else n_desc
+        prompt, nt = (self.refl_f, self.n_tail_f) if fwd else (self.refl, self.n_tail)
+        add = torch.zeros((self.B, prompt.shape[1], self.d), device=self.dev)
+        add[:, self.place_pos] = (g_r*self.Rn/self.rc)[:, None, None]*zwin
+        ids, st = self.run(prompt, n_desc, add, None, a.temp_r, a.min_p, tail=nt, add_layer=layer)
+        return ids, self.pool(st[:, :nt], 0)
+
+    def content(self, x, base):
+        """J-space decomposition of a centred state, relative to the typical workspace `base`."""
+        idx, coef, part = self.js.decompose(x)
+        return idx, coef, part-base
+
+    # ---------------------------------------------------------------- calibration
+    def calibrate(self):
+        a = self.a;zero = torch.zeros((self.B, self.d), device=self.dev)
+        sts = []
+        for _ in range(a.n_cal):
+            _, st = self.thought(zero);sts.append(st[:, a.skip:])
+        st = torch.cat(sts, 1).reshape(-1, self.d);nrm = st.norm(dim=-1)
+        ok = nrm <= 2.5*nrm.median()
+        self.Rn = float(nrm[ok].median());self.mu = st[ok].mean(0)
+        xs = torch.cat([self.pool(s, 0) for s in sts])-self.mu
+        _, _, part = self.js.decompose(xs);self.rj = float(part.norm(dim=-1).mean())
+        self.jbar = part.mean(0);self.rc = float((part-self.jbar).norm(dim=-1).mean())
+        self.rj_x = float(xs.norm(dim=-1).mean())
+        _, m = self.reflect(torch.zeros((self.B, a.K, self.d), device=self.dev), self.g_r, n_desc=0)
+        self.mu_r = m.mean(0)                                            # deterministic: one pass is the mean
+        _, _, mp = self.js.decompose(m-self.mu_r);self.mbar = mp.mean(0)
+        _, m = self.reflect(torch.zeros((self.B, a.K, self.d), device=self.dev), self.g_r, n_desc=0, fwd=True)
+        self.mu_rf = m.mean(0);_, _, mp = self.js.decompose(m-self.mu_rf);self.mbar_f = mp.mean(0)
+        log(f'calibrated: residual norm {self.Rn:.1f} | thought-mean norm {self.rj_x:.2f} | J-space part {self.rj:.2f}'
+            f' | its typical part {self.jbar.norm():.2f} | content (part - typical) {self.rc:.2f}')
+
+    # ---------------------------------------------------------------- main loop
+    def main(self):
+        a = self.a;B, d, K, N = self.B, self.d, a.K, a.n_iter;NR = N//K;dev = self.dev
+        ck = os.path.join(a.out, 'ckpt.pt')
+        if os.path.exists(ck):
+            S = torch.load(ck, map_location='cpu', weights_only=False)
+            for k in ('mu', 'mu_r', 'jbar', 'mbar', 'mu_rf', 'mbar_f', 'clamp_vec'):setattr(self, k, S[k].to(dev))
+            self.Rn, self.rj, self.rj_x, self.rc = S['Rn'], S['rj'], S['rj_x'], S['rc']
+            z, inj_m, zhist, ablated = S['z'].to(dev), S['inj_m'].to(dev), S['zhist'].to(dev), S['ablated'].to(dev)
+            self.gen.set_state(S['gen']);t0 = S['t'];D = S['data']
+            log(f'resumed at thought {t0}')
+        else:
+            self.calibrate()
+            z = torch.zeros((B, d), device=dev);inj_m = torch.zeros((B, d), device=dev)
+            zhist = torch.zeros((K, B, d), device=dev);ablated = torch.zeros(B, dtype=torch.bool, device=dev)
+            t0 = 0;self.clamp_vec = torch.zeros((B, d), device=dev)
+            D = dict(mf=np.zeros((NR, B, d), np.float16), fdesc=[], generic=np.zeros((NR, d), np.float16),j=np.zeros((N, B, d), np.float16), j_idx=np.zeros((N, B, a.k), np.int32),
+                     j_coef=np.zeros((N, B, a.k), np.float16), texts=[],
+                     m=np.zeros((NR, B, d), np.float16), m_idx=np.zeros((NR, B, a.k), np.int32),
+                     m_coef=np.zeros((NR, B, a.k), np.float16), desc=[],
+                     inj=np.zeros((NR, B, d), np.float16), inj_src=np.zeros((NR, B), np.int32),
+                     fid={f'{l}_{g}':np.zeros((NR, B, d), np.float16) for l, g in self.fid},
+                     fid_desc={f'{l}_{g}':[] for l, g in self.fid})
+        t_start = time.time()
+        for t in range(t0, N):
+            u = z+self.beta[:, None]*(~ablated)[:, None]*inj_m
+            nu = u.norm(dim=-1, keepdim=True)/self.rc                    # bounded: never above 1.5 typical contents
+            v = (self.g*self.Rn/self.rc)[:, None]*u*torch.clamp(1.5/nu.clamp_min(1e-6), max=1.)
+            ids, st = self.thought(v)
+            x = self.pool(st, a.skip)-self.mu
+            jidx, jcoef, j = self.content(x, self.jbar)
+            z = (1-a.alpha)*z+a.alpha*j
+            zhist = torch.cat([zhist[1:], z[None]])
+            D['j'][t] = j.cpu().half().numpy();D['j_idx'][t] = jidx.cpu().numpy();D['j_coef'][t] = jcoef.cpu().half().numpy()
+            D['texts'].append(self.tok.batch_decode(ids))
+            if (t+1) % K == 0:
+                r = (t+1)//K-1;zwin = zhist.transpose(0, 1)
+                dids, mstate = self.reflect(zwin, self.g_r)
+                midx, mcoef, m = self.content(mstate-self.mu_r, self.mbar)
+                D['m'][r] = m.cpu().half().numpy();D['m_idx'][r] = midx.cpu().numpy();D['m_coef'][r] = mcoef.cpu().half().numpy()
+                D['desc'].append(self.tok.batch_decode(dids))
+                fids, fst = self.reflect(zwin, self.g_r, n_desc=a.n_desc_f, fwd=True)   # self-prediction (logged only)
+                D['mf'][r] = self.content(fst-self.mu_rf, self.mbar_f)[2].cpu().half().numpy()
+                D['fdesc'].append(self.tok.batch_decode(fids))
+                for l, g in self.fid:                                  # the same reflection at other settings
+                    _, fst = self.reflect(zwin, torch.full_like(self.g_r, g), layer=l, n_desc=0)
+                    D['fid'][f'{l}_{g}'][r] = self.content(fst-self.mu_r, self.mbar)[2].cpu().half().numpy()
+                unit = lambda u:u/u.norm(dim=-1, keepdim=True).clamp_min(1e-6)*self.rc
+                avg = self.content(zhist.mean(0), 0.)[2]
+                # the population's generic self-model at this moment: the mean over the observer rows, which no
+                # treatment touches. The individuated conditions (loopc, yokedc, clampc) feed back only what sets one
+                # mind's self-model apart from it.
+                generic = m[self.free_rows].mean(0) if self.free_rows else m.mean(0)
+                D['generic'][r] = generic.cpu().half().numpy()
+                mi = m-generic[None]
+                new = torch.zeros_like(inj_m);src = []
+                for b, c in enumerate(self.cond):
+                    if c == 'free':src.append(-1);continue
+                    if c in ('loop', 'perturb', 'ablate', 'loopc', 'clampc'):s = b
+                    elif c in ('yoked', 'yokedc'):s = self.partner[b]
+                    elif c == 'avg':s = -2
+                    if c == 'perturb' and t+1 == a.t_int:s = self.foreign[b]
+                    if c in ('loopc', 'yokedc', 'clampc'):
+                        vec = unit(mi[s:s+1])[0]
+                        if c == 'clampc' and t+1 == a.t_int:self.clamp_vec[b] = unit(mi[self.foreign[b]:self.foreign[b]+1])[0]
+                        if c == 'clampc' and a.t_int <= t+1 < a.t_int+a.clamp_len:vec, s = self.clamp_vec[b], self.foreign[b]
+                        new[b] = vec
+                    else:new[b] = unit(avg[b:b+1])[0] if s == -2 else unit(m[s:s+1])[0]
+                    src.append(s)
+                    if c == 'ablate' and t+1 >= a.t_int:ablated[b] = True
+                inj_m = new
+                D['inj'][r] = inj_m.cpu().half().numpy();D['inj_src'][r] = np.array(src)
+                el = time.time()-t_start;rate = el/(t+1-t0)
+                w = [f"{self.cond[b]}/{self.rows[b]['seed']}: {self.js.top(midx[b].cpu().numpy(), mcoef[b].cpu().numpy(), 5)}"
+                     for b in range(0, B, max(1, B//6))]
+                log(f'thought {t+1}/{N} ({rate:.2f} s/thought, eta {(N-t-1)*rate/60:.1f} min) | self-models: ' + ' | '.join(w))
+                log('   e.g. thought: ' + D['texts'][-1][0][:160].replace('\n', ' ') + ' || answer: ' + D['desc'][-1][0][:120])
+            if (t+1) % a.ckpt_every == 0 or t+1 == N:
+                S = dict(mu=self.mu.cpu(), mu_r=self.mu_r.cpu(), jbar=self.jbar.cpu(), mbar=self.mbar.cpu(), rc=self.rc,
+                         mu_rf=self.mu_rf.cpu(), mbar_f=self.mbar_f.cpu(), clamp_vec=self.clamp_vec.cpu(),
+                         Rn=self.Rn, rj=self.rj, rj_x=self.rj_x,
+                         z=z.cpu(), inj_m=inj_m.cpu(), zhist=zhist.cpu(), ablated=ablated.cpu(), gen=self.gen.get_state(),
+                         t=t+1, data=D)
+                torch.save(S, ck+'.tmp');os.replace(ck+'.tmp', ck)
+        self.save(D)
+
+    def save(self, D):
+        a = self.a
+        np.savez(os.path.join(a.out, 'traj.npz'), j=D['j'], j_idx=D['j_idx'], j_coef=D['j_coef'], m=D['m'],
+                 m_idx=D['m_idx'], m_coef=D['m_coef'], inj=D['inj'], inj_src=D['inj_src'], mu=self.mu.cpu().numpy(),
+                 jbar=self.jbar.cpu().numpy(), mbar=self.mbar.cpu().numpy(), mf=D['mf'], generic=D['generic'],
+                 **{f'fid_{k}':v for k, v in D['fid'].items()})
+        meta = dict(args=vars(a), rows=self.rows, partner=self.partner, foreign=self.foreign, layer=self.L, L_in=self.L_in,
+                    Rn=self.Rn, rj=self.rj, rj_x=self.rj_x, rc=self.rc, words=self.js.words, texts=D['texts'], desc=D['desc'],
+                    fid_desc=D['fid_desc'], fdesc=D['fdesc'], thought_frame=THOUGHT_FRAME,
+                    reflect_prompt=REFLECT_PRE+PLACE*a.K+REFLECT_POST, forward_prompt=REFLECT_PRE+PLACE*a.K+REFLECT_FWD)
+        with open(os.path.join(a.out, 'meta.json'), 'w', encoding='utf-8') as f:json.dump(meta, f, ensure_ascii=False)
+        log('saved', a.out)
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument('--model', default='Qwen/Qwen3-4B-Base');p.add_argument('--jlens', default='qwen3-4b-base-jlens.npz')
+    p.add_argument('--layer', type=int, default=16, help='only with --jlens identity')
+    p.add_argument('--out', required=True);p.add_argument('--grid', default='main')
+    p.add_argument('--conds', default='free,loop,yoked,avg,perturb,ablate');p.add_argument('--seeds', type=int, default=8)
+    p.add_argument('--n_iter', type=int, default=300);p.add_argument('--K', type=int, default=10)
+    p.add_argument('--t_int', type=int, default=150, help='thought at which perturb/ablate/clampc rows are treated')
+    p.add_argument('--clamp_len', type=int, default=50, help='thoughts for which clampc holds a foreign self-model')
+    p.add_argument('--n_desc_f', type=int, default=12, help='words said in answer to the forward question (display)')
+    p.add_argument('--n_tok', type=int, default=40);p.add_argument('--n_desc', type=int, default=24)
+    p.add_argument('--skip', type=int, default=3);p.add_argument('--k', type=int, default=16)
+    p.add_argument('--vocab', default='english', help="J-lens dictionary: 'english' (common English words) or 'words'")
+    p.add_argument('--g', type=float, default=.1, help='recurrent injection, in units of the residual norm')
+    p.add_argument('--g_r', type=float, default=.5, help='placeholder injection in the reflection')
+    p.add_argument('--beta', type=float, default=1., help='self-model weight relative to a typical thought')
+    p.add_argument('--alpha', type=float, default=.4, help='leak of the recurrent state toward each new thought')
+    p.add_argument('--temp', type=float, default=1.);p.add_argument('--temp_r', type=float, default=.3)
+    p.add_argument('--min_p', type=float, default=.05);p.add_argument('--n_cal', type=int, default=3)
+    p.add_argument('--L_in', type=int, default=12, help='layer where the reflection placeholders get the states (-1: L)')
+    p.add_argument('--fid', default='', help='extra reflections to log for a fidelity check: layer:gain,layer:gain')
+    p.add_argument('--rng', type=int, default=1234);p.add_argument('--ckpt_every', type=int, default=20)
+    a = p.parse_args()
+    os.makedirs(a.out, exist_ok=True)
+    torch.manual_seed(a.rng)
+    Experiment(a).main()
+
+
+if __name__ == '__main__':
+    main()