#!/usr/bin/env python3 """ORACLE-STATE CONTROL: does PNS-World admit a bounded sufficient state? The negative result says learned recurrent states never become causally useful. A reviewer's remaining escape hatch is: "maybe the semantic target or the output architecture is simply not learnable." This control closes it. We hand the model the simulator's *ground-truth bounded* semantic state - the current (entity, attribute) -> value map and the (entity, relation) -> entity map, as at most 120 typed slots, exactly the capacity Sigma has - plus the question, and ask the SAME query-conditioned read + enum head to answer. S_oracle_t + e_t -> enum head -> answer If this scores high, the task admits a bounded sufficient state and the answer heads can exploit one, so the learned models' failure is in CONSTRUCTING or USING state, not in the task or the readout. If it scores low, the negative result is confounded and must be reported as such. The slot CONTENTS are oracle; their embeddings are learned (a symbol must be embedded to be read). Nothing about the future or the gold answer's identity is supplied - only the world state that a perfect bounded summariser would hold at t-. """ import argparse import sys import time from pathlib import Path import numpy as np import torch import torch.nn as nn import torch.nn.functional as F sys.path.insert(0, str(Path(__file__).resolve().parent.parent / "src")) sys.path.insert(0, str(Path(__file__).resolve().parent)) from pns.common import atomic_write_json, eval_root # noqa: E402 from pns.model.modules import enum_legal_mask # noqa: E402 from pns.world.schema import ENUM_ID, ENUM_VOCAB, ET, Fam # noqa: E402 from pns.world.sim import LifetimeGen # noqa: E402 N_SLOT, N_ENT, N_ATTR, N_VAL = 120, 40, 24, len(ENUM_VOCAB) + 1 ATTRS = sorted({a for a in ["location", "team", "status", "phase", "priority", "destination", "owner", "project", "vault"]}) AID = {a: i for i, a in enumerate(ATTRS)} def build(seed, length=256): """Replay one lifetime; at every SEM question emit the oracle slot table as it stands BEFORE that question, plus the query and gold.""" g = LifetimeGen(seed, length) g.plan() events, red, _ = g.walk() out = [] # Derive the as-of-t bounded state directly from the event stream, exactly # as a perfect bounded summariser would - including revert semantics. live, rel = {}, {} hist = {} # (ent, attr) -> [(clock, source, value), ...] clock = 0 for t, ev in enumerate(events): clock += ev.dt p = ev.payload if ev.etype == ET.QUESTION and ev.family in (int(Fam.SEM_LATEST), int(Fam.SEM_2HOP)): tk = g.slots[t] d_ = tk.data if tk is not None else {} slots = [] for (e, a), v in live.items(): if a in AID and v in ENUM_ID: slots.append((e % N_ENT, AID[a], ENUM_ID[v] + 1, 0)) for (e, r), dst in rel.items(): if r in AID: slots.append((e % N_ENT, AID[r], 0, dst % N_ENT + 1)) if ev.family == int(Fam.SEM_LATEST): q_ent, q_rel, q_attr = d_.get("ent", 0) % N_ENT, 0, d_.get("attr") else: q_ent = d_.get("src", 0) % N_ENT q_rel = AID.get(d_.get("rel", ""), 0) + 1 q_attr = d_.get("attr") if q_attr not in AID: continue lm = np.zeros(8, np.uint8) for _e in ev.enum_legal: lm[_e >> 3] |= 1 << (_e & 7) out.append(dict(slots=slots[:N_SLOT], gold=ev.enum_gold, legal=lm, fam=ev.family, q_ent=q_ent, q_rel=q_rel, q_attr=AID[q_attr], delay=ev.meta.get("delay", -1))) # ---- advance the ground-truth bounded state if ev.etype == ET.OBSERVATION: if p.get("rel"): rel[(p["ent"], p["attr"])] = p["ent2"] else: k = (p["ent"], p["attr"]) live[k] = p["value"] hist.setdefault(k, []).append((clock, p.get("source", ""), p["value"])) elif ev.etype == ET.CORRECTION: k = (p["ent"], p["key"]) if p["mode"] == "replace": if not p.get("exact") and p.get("value") is not None: live[k] = p["value"] hist.setdefault(k, []).append((clock, "", p["value"])) else: # revert: restore the referenced earlier assertion h = hist.get(k, []) tgt = None if p.get("ref_kind") == "source": for c, src, v in h: if src == p.get("ref_source"): tgt = v else: for c, src, v in h: if c == p.get("ref_clock"): tgt = v if tgt is not None: live[k] = tgt hist[k] = [x for x in h if x[0] <= (p.get("ref_clock") or c)] return out class OracleReader(nn.Module): """Query-conditioned attention read over oracle slots -> enum head. Same shape of computation as PNSR's read path, same enum output head.""" def __init__(self, d=256): super().__init__() self.ent = nn.Embedding(N_ENT + 1, d) self.attr = nn.Embedding(N_ATTR + 1, d) self.val = nn.Embedding(N_VAL + 1, d) self.dst = nn.Embedding(N_ENT + 2, d) self.qe = nn.Embedding(N_ENT + 1, d) self.qa = nn.Embedding(N_ATTR + 1, d) self.qr = nn.Embedding(N_ATTR + 2, d) self.rd = nn.MultiheadAttention(d, 8, batch_first=True) self.rd2 = nn.MultiheadAttention(d, 8, batch_first=True) self.mlp = nn.Sequential(nn.LayerNorm(d), nn.Linear(d, d), nn.GELU(), nn.Linear(d, d)) self.enum = nn.Linear(d, len(ENUM_VOCAB)) def forward(self, se, sa, sv, sd, mask, qe, qa, qr): s = self.ent(se) + self.attr(sa) + self.val(sv) + self.dst(sd) q = (self.qe(qe) + self.qa(qa) + self.qr(qr)).unsqueeze(1) # hop 1: resolve the referent (identity for 1-hop questions) r1, _ = self.rd(q, s, s, key_padding_mask=~mask, need_weights=False) # hop 2: read the attribute of whatever hop 1 resolved q2 = r1 + self.qa(qa).unsqueeze(1) r2, _ = self.rd2(q2, s, s, key_padding_mask=~mask, need_weights=False) h = r2.squeeze(1) return self.enum(h + self.mlp(h)) def pack(rows, dev): B = len(rows) se = torch.zeros(B, N_SLOT, dtype=torch.long) sa = torch.zeros(B, N_SLOT, dtype=torch.long) sv = torch.zeros(B, N_SLOT, dtype=torch.long) sd = torch.zeros(B, N_SLOT, dtype=torch.long) mk = torch.zeros(B, N_SLOT, dtype=torch.bool) for i, r in enumerate(rows): for j, (e, a, v, dd) in enumerate(r["slots"][:N_SLOT]): se[i, j], sa[i, j], sv[i, j], sd[i, j], mk[i, j] = e, a, v, dd, True qe = torch.tensor([r["q_ent"] for r in rows]) qa = torch.tensor([r["q_attr"] for r in rows]) qr = torch.tensor([r["q_rel"] for r in rows]) y = torch.tensor([r["gold"] for r in rows]) lg = torch.tensor(np.stack([r["legal"] for r in rows])) return [x.to(dev) for x in (se, sa, sv, sd, mk, qe, qa, qr, y, lg)] def main(): ap = argparse.ArgumentParser() ap.add_argument("--train-lifetimes", type=int, default=3000) ap.add_argument("--val-lifetimes", type=int, default=400) ap.add_argument("--steps", type=int, default=3000) args = ap.parse_args() dev = "cuda" print("building oracle rows ...", flush=True) tr = [r for s in range(args.train_lifetimes) for r in build(s)] va = [r for s in range(1_000_000, 1_000_000 + args.val_lifetimes) for r in build(s)] print(f"train rows {len(tr)} val rows {len(va)}", flush=True) m = OracleReader().to(dev) opt = torch.optim.AdamW(m.parameters(), lr=3e-4, weight_decay=0.01) rng = np.random.default_rng(0) t0 = time.time() for step in range(args.steps): idx = rng.integers(0, len(tr), 256) se, sa, sv, sd, mk, qe, qa, qr, y, lg = pack([tr[i] for i in idx], dev) logits = m(se, sa, sv, sd, mk, qe, qa, qr).masked_fill(~enum_legal_mask(lg), -1e9) loss = F.cross_entropy(logits, y) opt.zero_grad(); loss.backward(); opt.step() if (step + 1) % 500 == 0: m.eval() accs, per = [], {} with torch.no_grad(): for k in range(0, len(va), 512): ch = va[k:k + 512] se, sa, sv, sd, mk, qe, qa, qr, y, lg = pack(ch, dev) p = m(se, sa, sv, sd, mk, qe, qa, qr).masked_fill(~enum_legal_mask(lg), -1e9).argmax(-1) ok = (p == y).float().cpu().numpy() accs.extend(ok) for r, o in zip(ch, ok): per.setdefault(Fam(r["fam"]).name, []).append(o) m.train() print(f"step {step+1} loss {float(loss):.3f} val {np.mean(accs):.4f} " + " ".join(f"{k} {np.mean(v):.3f}" for k, v in sorted(per.items())) + f" ({time.time()-t0:.0f}s)", flush=True) rep = {"val_acc": float(np.mean(accs)), "n_val": len(va), "per_family": {k: float(np.mean(v)) for k, v in per.items()}, "n_train_rows": len(tr)} atomic_write_json(eval_root() / "ORACLE_STATE.json", rep) print(rep) if __name__ == "__main__": main()