Upload NIGHT7_STATE.md with huggingface_hub

NIGHT7_STATE.md

@@ -40,6 +40,9 @@ PILLAR C — GENERALITY (owed from night-6 L4): port length-dilution + no-bindin
 - A1 capacity law fit: recoverability(content-type, N) curves -> closed-form? categorical no-binding vs graded dilution test.
 
 ## LOG (newest at top)
+- 09:45 (Jun-22) — GAUGE MAKER ROUND 2 → v0.4 (conceptual win). AXIS-APPROPRIATE CALIBRATION implemented: each axis its own honesty-negative (R→C1/C3, F/X→C2, D→Cd-discovery-noise[NEW], M/S→Cm-phantom-site[NEW]); technique faces negatives matching declared axes. RESULT: honest discovery/mechanism methods now HEADLINE (causal_patch +.83, ica +.38, kmeans +.34) — fixes the inverted-purpose bug; noise-hallucinators go negative (eye-dump −1.95). Cherry-pick re-closed w/o blunt cap (eligible denom=declared∪implemented + must face matching negative). PER-SEED rotation Q kills seed-memorizers (1.0→0.0 seed1; leaderboard now ≥2-seed avg). E1 fit-offload metered (1.0→.96). Residuals: R1 strong-baseline-subtraction, D2 sparse-code-recon, F1 off-causal-penalty. GAUGE_SPEC v0.4.
+       NEXT CYCLE = BREAKER ROUND 3 (attack v0.4 + the 4 maker-flagged items): (1) R1-strong-baseline now ZEROS honest linear probes (planted feature linear → can't beat ridge) → OWED a 2nd R challenge w/ non-linear/noise structure where a good technique CAN beat ridge (build it); (2) D2 recon discrimination modest (.37 vs .33) — probe a top-k-recon-tuned wrong basis; (3) S-axis reuses Cm loosely → owed S-specific known-negative (identity/no-clock operator); (4) find NEW v0.4 exploits (does axis-appropriate-calibration open a dodge? can a technique declare a trivial axis-subset to face an easy negative?). v0.4 leaderboard avg(seed0,1): causal_patch+.83/ica+.38/kmeans+.34/diff_means+.20/readout_lens+.17/pca+.11; probes+.00; controls−1.0/−1.2/−1.98. Loop cron-driven.
+
 - 08:05 (Jun-22) — GAUGE BREAKER ROUND 2 done (gauge/optimizer/blindspots_r2.md). All 7 round-1 fixes HOLD (verified). NEW findings ranked for MAKER ROUND 2:
        #1 (TOP — inverts purpose) CALIBRATION-MANDATE bug: v0.3 zeroes honest non-probe techniques (ica/causal_patch/kmeans/diff_means/sae → TRUST0/UNCAL) for not taking the role C-test, while mediocre mlp_probe headlines. FIX: AXIS-APPROPRIATE calibration — discovery-noise-negative (report factors on pure noise?), mechanism-phantom-site-negative, keep C1-C3 for probes; mandate the negative MATCHING each declared axis.
        #2 SEED-INDEPENDENT blocks: U/v_causal/v_ro/number/entity are seed-invariant by construction → hardcoded F1/X1/S1 dirs score 1.0 every seed (even a seed-sweep won't catch). FIX: per-seed random block rotation in synth.py.