Skip to content

PREREGISTRATION_L3.md

Latest commit

 

History

History
673 lines (627 loc) · 52.7 KB

PREREGISTRATION_L3.md

File metadata and controls

673 lines (627 loc) · 52.7 KB
Raw

Pre-Registration - ITASORL Experiment L3 (Learned-World-Model Surrogate, Incidental Detection)

Status: pre-registered before the confirmatory run. Date: 2026-07-10. Gate: the L2/pooled line is a confirmed negative under scale - B-v3 regime n=10 survival pooled 0.610 (90% CI [0.585, 0.634]) and the sysid-aux capacity ceiling n=10 pooled 0.596 (90% CI [0.577, 0.616]) both sit entirely below the 0.65 bar, so the pre-registered "proceed to L3 only if L2 holds under scale" condition is met. Configuration below is frozen at commit time; any later change is logged in section 12.

1. Background and the contrast

Experiments B-v2 (L2 drift, AR(1)) and B-v3 (L2 regime, per-episode constant offset) asked whether a survival actor-critic incidentally encodes world identity when the surrogate differs from the authentic world by a hand-specified dynamics parameter (the drag). It does not, and the negative is now near-architectural: even a trunk supervised directly on drag (the sysid-aux ceiling) reads the pooled persistent-direction probe at only ~0.60, while the matched-pair detectability channel reaches ~0.70 - identity is decodable when forced in, but the pooled readout saturates below the bar (docs/FINDINGS.md section 7.1, section 9).

L3 changes what makes the surrogate a surrogate. Instead of a hand-tuned drag parameter, the surrogate world is produced by a separate neural network trained to imitate the authentic dynamics. Its differences from the authentic world are the generative fingerprint of that imitation: smoothing / blur, missing fine detail, a characteristic learned noise texture, and mode collapse (rare events that should occur never do) (docs/ITASORL.md L3, ladder section). This is a structurally different, more realistic surrogate than a single dynamics knob, and it is the project's headline target (docs/ITASORL.md).

2. Hypothesis

  • H_L3 (primary): a from-scratch survival agent living across authentic and L3-surrogate worlds incidentally encodes world identity in its recurrent state above the prediction-only and untrained baselines and above the SESOI, read out but never rewarded.
  • Null: pooled target ~0.50, no better than baselines - even a rich generative fingerprint is not incidentally encoded under readout-not-reward.

Either outcome is substantive. The result is conditional on this surrogate's difficulty (section 7 oracle gate): it says what kind of substrate artifact survival will or will not encode, not that no fingerprint could ever be encoded.

3. Agents - identical trunk and identical readout; only the objective differs

Unchanged from B-v2/B-v3: untrained (mechanical floor), predictor (next-step prediction on this trunk), survival (recurrent A2C + GAE + world-model decoder aux). "L3 fingerprint induces encoding" requires survival > predictor > untrained and survival >= SESOI. The probe is identical across all three.

4. World and the L3 surrogate (readout-not-reward preserved)

Signal targeted (decision, section 11): the organism test targets Signal A (latent geometry) - the agent lives in both authentic and L3-surrogate worlds and the probe reads world identity from its recurrent state - keeping the "incidental, readout-not-reward" framing identical to B-v2/B-v3. Signal B (prediction-residual of a world model frozen on authentic then dropped into the surrogate) is used only inside the section 7 oracle check, not as the organism headline.

Surrogate construction:

  1. Two-stage construction (C -> B), resolved after a code-level integration review (2026-07-10). G is a small predictive world-model (RecurrentWorldModel) trained on authentic PatchOfEarthV0 rollouts and then frozen. Because G predicts OBSERVATIONS while the survival world needs full internal STATE (energy, food, death), "wrap G as the World" is realized in two stages:
    • Stage 1 - oracle gate (Option A, observation-channel). Authentic physics is kept intact (survival stays well-defined); an L3-surrogate trajectory is an authentic rollout whose OBSERVATIONS are passed through frozen G, so the tell is G's generative reconstruction fingerprint (blur, missing fine detail, mode collapse). This is what the section-7 Experiment-A oracle detects. Cheap; no change to the survival dynamics.
    • Stage 2 - organism run (Option B, dynamics-level), CONTINGENT on the gate. Only if Stage 1's oracle lands in [0.85, 0.95] do we build the dynamics-level surrogate: G (trained on logged STATE transitions) drives the world's dynamics, with energy/food/death computed from G's predicted state, plugged into itasorl/experiment_b2.py::make_world as the new l3 drift-mode where the L2 path lives. If G's learned dynamics cannot sustain survival (engagement gate fails), that is logged in section 12 and the organism construction is re-scoped before any multi-seed run - a stated empirical risk, not assumed away. The authentic world (drift off) is unchanged and byte-identical to B-v2/B-v3 authentic in both stages.
  2. Difficulty is calibrated by outcome, not architecture (decision). We choose the smallest-capacity G whose Experiment-A oracle detectability (section 7) lands in the pre-set band AUROC in [0.85, 0.95]: clearly detectable (the fingerprint exists) but not trivially separable like L2's ~0.99. Capacity is the single difficulty knob.
  3. Fixed data budget (decision). G trains on a fixed budget of authentic rollouts (frozen in section 9); we do NOT trade data against capacity - capacity alone sets difficulty, so the surrogate's fidelity is reproducible.
  4. Same harsh metabolism, denser food, and potential-based food shaping as B-v2/B-v3 (SURVIVAL_METAB / SURVIVAL_FOOD); reported returns use the TRUE reward.
  5. World identity (authentic vs G-surrogate) is never in the observation or the reward; only the probe ever sees it.

5. Readouts (episode = unit of inference; GroupKFold)

Identical harness to B-v3:

  • PRIMARY - pooled (Experiment B frame): target AUROC of a linear probe on episode_features(h_t) = [mean h, final h], plus additive target_var / target_full and per-probe selectivity = target - shuffled.
  • SECONDARY - matched-pair recurrent: detectability of the fingerprint in the agent state from a bit-identical shared prefix; reported, not the headline.

6. Estimand and smallest effect of interest

  • Estimand: per-episode selectivity-corrected AUROC of the pooled linear world-identity probe, averaged over seeds.
  • SESOI (decision - reuse 0.65): claim incidental encoding only if the survival pooled target is >= 0.65 AND exceeds both predictor and untrained by >= 0.05 - the same bar as B-v2/B-v3, so L3 is directly comparable to the earlier rungs.

7. Gates - ALL must pass before the survival target is interpreted

The four B-v3 gates, PLUS a new L3-specific manipulation gate that must pass FIRST:

  1. Oracle-detectability gate (L3-specific, decision - oracle-first). BEFORE any organism run, an Experiment-A oracle (a privileged discriminator with direct access to observations / prediction residuals) must detect authentic vs G-surrogate at AUROC in [0.85, 0.95]. Below 0.85: the surrogate is too good, there is no fingerprint to encode - fix G, do not run the organism. Above 0.95: too easy (approaching L2's trivial regime) - increase G's capacity. This gate both confirms the tell exists and freezes the difficulty.
  2. Engagement: trained TRUE return beats random and scripted by ENGAGE_MARGIN with lifetime not worse by LIFE_TOL.
  3. L0 control: authentic-vs-authentic pooled target equivalent to 0.5 by TOST.
  4. Positive control: speed probe >= 0.75.
  5. Leakage audit: reward-sum / length / lifetime each within 0.1 of 0.5.

8. Decision matrix

  • Encoding induced (conditional on this fingerprint): gate 0 passes AND all gates pass AND survival target >= 0.65 AND > predictor+0.05 AND > untrained+0.05.
  • Strengthened negative: gate 0 passes AND all gates pass AND survival ~= 0.50 ~= predictor ~= untrained. (The strongest form of the project's core claim: even a rich learned-surrogate fingerprint, verified detectable by an oracle, is not incidentally encoded under survival pressure.)
  • Uninformative: gate 0 fails (surrogate mis-calibrated) or engagement fails.

9. Fixed configuration (frozen at commit)

  • Surrogate net G: family = small recurrent predictor (GRU core). Fixed training config, frozen here: next-state MSE loss, Adam lr=1e-3, a fixed number of epochs to convergence (early-stop on held-out authentic MSE, patience 5), seed 0. The ONLY free knob is hidden size, swept ascending over {8, 16, 24, 32, 48, 64}; the FIRST size whose section-7 oracle AUROC lands in [0.85, 0.95] is FROZEN (selected value + oracle AUROC recorded in section 12 before any organism run).
  • Capacity-sweep fallback (no valid capacity in the grid). Because AUROC decreases with capacity, the sweep can step OVER the band (e.g. hidden=16 gives 0.97, hidden=24 gives 0.82). If no grid point lands in [0.85, 0.95]: (a) bisect hidden size between the two bracketing points; if still no fit, (b) hold the smallest capacity that is >= 0.85 and REDUCE the authentic-data budget (a secondary difficulty knob) until oracle AUROC enters the band. Whichever path is used, the final G config is frozen in section 12 and the organism run does not start until gate 0 passes. This guarantees gate 0 is reachable.
  • G training data budget: fixed authentic-rollout budget frozen here (initial proposal: the same collection scale as B-v2 collect_pool, i.e. on the order of a few hundred authentic episodes); no data/capacity trade.
  • Organism config (mirrors B-v3): drifts [0.0, surrogate]; seeds [0,1,2] confirmatory, power extension 0..9; updates 300; n_eps=16; max_steps=80; hidden=96; ray_steps=5; shaping_coef=1.0; pooled n=110, steps=24; matched-pair pairs=60, prefix=20, branch=24. Adam lr=3e-4, gamma=0.99, lam=0.95, ent_coef=0.01, vf_coef=0.5, wm_coef=1.0.

10. Analysis plan

Mean +/- std over seeds. Primary contrast: survival vs predictor vs untrained pooled target across authentic vs G-surrogate. Report every gate including the oracle gate. Same figure/JSON schema as B-v2/B-v3. Also report the sysid-aux CEILING for L3 as the capacity reference, and the matched-pair detectability as the "decodable-when-forced" comparison, exactly as for L2.

Rigor carried from the B-v3 audit (2026-07-10):

  • Honest seed-level CI. The primary "clears / misses 0.65" adjudication uses a t-based (or BCa) 90% CI on the per-seed AUROCs, not the percentile-of-mean bootstrap alone (at n=10 the percentile-of-mean interval under-covers at the boundary). Report both; the decision follows the t/BCa interval.
  • Per-world survivorship. Report authentic vs surrogate survivor counts / death rates per cell, so the pooled survivorship asymmetry (section 11) is quantified, not assumed.
  • Multiplicity. The volatility readouts (target_var, target_full) are secondary / exploratory; only the level target enters the pre-registered SESOI decision. Any volatility-based claim is flagged exploratory and not multiplicity-corrected.

11. Key interpretive caveat

  • This is a different, higher rung, not the L2 claim. A positive L3 result answers "does survival encode a learned-world-model generative fingerprint", not "does survival encode L2 drift". A negative is the strongest form of detectable != learned: the fingerprint is oracle-verified present (gate 0), yet the survival organism does not form a persistent world-identity direction from it.
  • Signal A vs Signal B. The organism headline is Signal A (latent geometry from living in both worlds); Signal B (frozen-world-model prediction-residual) is the oracle-side check only. Conflating them would confuse detectability with incidental encoding.
  • Difficulty is a design choice, made explicit. The oracle band [0.85, 0.95] fixes how hard the tell is; the result is conditional on that band, this architecture, this world family, and this surrogate net.
  • Fingerprint character is confounded with capacity, and we bound it rather than assume it away. A low-capacity G and a high-capacity G can both land in the oracle band while producing qualitatively different artifacts (heavy blur vs mode collapse vs a learned noise texture). The band fixes difficulty, not artifact type. We therefore (a) record which fingerprint dominates at the frozen capacity (via the oracle's most informative feature channel) in section 12, and (b) run the organism test at TWO in-band capacities when the sweep yields two candidates, reporting whether the encoding verdict is stable across artifact type. A single-capacity result is reported as conditional on that artifact.
  • Pooled readout is Exp-B-comparable, not confound-clean. As in B-v2/B-v3, the pooled probe drops early deaths per world, so a harder surrogate that kills more episodes yields a survivorship-selected surrogate pool. The matched-pair channel (bit-identical shared prefix, equal-length truncation) is the confound-clean detectability guard; the pooled target is read as directly comparable to Experiment B, not as the confound-controlled estimate. This run additionally reports per-world survivor counts / death rates (section 10) so the asymmetry is visibly bounded.

12. Deviations from pre-registration

  • 2026-07-10, milestone-1 empirical finding (Stage-1 oracle calibration, GPU). The Stage-1 observation-channel construction (Option A: authentic physics, observations passed through frozen G) was built and its oracle calibrated on a from-scratch G capacity sweep (hidden 8-128; one-step reconstruction AND open-loop imagination; n=200-300 authentic episodes; local GPU). Result: the artifact is trivially detectable (oracle AUROC ~1.0) at every capacity and every imagination horizon, and a matched sensor-noise / variance floor did NOT restore a calibratable gradient. Reason: a strong privileged oracle separates "real sensor data" from "any neural-net reconstruction of it" via G's systematic per-dimension manifold, not via G's fidelity - so capacity is not a working difficulty knob and the [0.85, 0.95] band is unreachable without deliberately weakening the oracle (which would game the gate). Per the pre-agreed escalation, the Stage-1 obs-channel construction is RETIRED and the dynamics-level construction becomes primary: the surrogate's observations are produced by the REAL sensor model applied to G's predicted STATE, so observations live on the authentic manifold and the sole tell is G's dynamics error, which capacity controls. Section 4 Stage 1 is superseded by this entry; the section-7 oracle gate now runs against the dynamics-level surrogate, with capacity re-validated as the difficulty knob before any organism run.

  • 2026-07-10, GATE 0 MET (dynamics-level oracle calibration, GPU). Built the dynamics-level construction: G_motion (learned velocity law, (vel,a)->vel_next, drag withheld) via itasorl/surrogate_l3.py, the l3 world hook, and the L2-style residual oracle (itasorl/experiment_a_l3.py, matched pairs, exact per-step authentic law from the logged transitions). Capacity is a clean MONOTONE difficulty knob; the section-7 [0.85, 0.95] band is reachable. Frozen gate-0 configuration: sensor-noise floor sigma=0.05, capacity hidden=8 -> oracle AUROC 0.890 (centered in-band; hidden=4 -> 0.940 also in-band; hidden>=16 -> <0.75). Leakage: MECHANICAL channels (length, metadata) clean at 0.5; reward legitimately leaks (~0.86 at hidden=4; different dynamics -> different movement cost -> different reward, the documented dynamics-rung consequence, docs/FINDINGS.md section 2.2). The residual oracle uses ONLY the velocity residual, so it does not exploit reward - BUT the reward coupling is a readout-not-reward CONSIDERATION for the organism run: the B-v2/B-v3 leakage audit + matched-pair equal-length truncation must be re-checked at the frozen difficulty (smaller dynamics error at hidden=8 should shrink the reward leak vs hidden=4). Gate 0 is met; the organism run may proceed with these controls reported.

  • 2026-07-10, ORGANISM RUN n=3 (confirmatory) + an unresolved measurement discrepancy. Ran run_expB2 --drift-mode l3 --l3-hidden 8 (n=3 seeds, 300 updates, local GPU, fullruns/l3_n3). At drift 0.45 the pooled target was survival 0.801 (90% CI [0.760, 0.842]), predictor 0.779, untrained 0.706; L0 control clean (0.514); manipulation check: artifact survival-relevant. Primary H_B2 NOT met by the SESOI (survival only +0.022 over predictor, < 0.05). TWO problems make this verdict NOT trustworthy yet: (1) high organism mechanical floor - the untrained (mechanical) agent already reads 0.706, so there is little headroom to isolate incidental encoding; (2) a re-calibration sweep (untrained floor vs oracle AUROC across G capacity) shows the tension is structural: larger G drops the untrained floor toward chance (hidden 32 -> 0.47, 128 -> 0.46) but the oracle signal COLLAPSES below the band (hidden 32 -> 0.70, 128 -> 0.63), so a clean-floor-AND-detectable capacity does not robustly exist above hidden=8. UNRESOLVED: an isolated untrained-floor probe at hidden=8 gives 0.548 reproducibly (identical on CPU and GPU training, so G is not the noise source), which CONTRADICTS the full run's 0.706 and lies OUTSIDE the run's CI [0.66, 0.80]. The full pipeline and the isolated measurement disagree for an unidentified reason. The L3 organism verdict is ON HOLD pending reconciliation of that discrepancy (a focused debugging task, not a re-run); only then is a clean-floor difficulty and a real H_L3 adjudication meaningful.

  • 2026-07-10, DISCREPANCY RESOLVED - a world-params bug; the confound was an artifact. Root cause of the 0.548-vs-0.706 divergence: run_expB2 runs in P = WorldParams(k_land=1.5, k_water=1.5, gravity=0.4), but setup_l3_surrogate trained G_motion on bare WorldParams() DEFAULTS - so G learned the WRONG world's dynamics and was deployed in another, inflating its systematic error and faking the high mechanical floor. Confirmed on GPU: G-trained-on-default read on P reproduces the run's 0.734 exactly; G-trained-on-P drops the untrained floor to 0.483 (chance). Fixed by passing params=P at both setup sites (regression test test_g_motion_wrong_world_diverges_more). The n=3 organism result above (untrained 0.706, survival 0.801) is INVALID - a bug artifact - and is retracted. Re-calibrated GATE 0 on the CORRECT world P: at hidden=8 the dumb/untrained floor is ~0.48 (chance) at every capacity 8-32, and the oracle lands in-band at sigma=0.02 -> AUROC 0.928 (leakage clean). So a clean L3 difficulty DOES exist - oracle-detectable fingerprint AND a chance-level organism mechanical floor. Frozen gate 0 (world P): hidden=8, sigma=0.02, oracle 0.928, untrained floor 0.483. Next: re-run the organism with the fix (G on P) for a real H_L3 adjudication - this time the mechanical floor is at chance, so survival-specific encoding can be isolated.

  • 2026-07-10, CORRECTED ORGANISM RUN n=3 - a PRELIMINARY POSITIVE (not yet decisive). Re-ran run_expB2 --drift-mode l3 --l3-hidden 8 (n=3, 300 updates, local GPU, fullruns/l3_n3_fixed, commit eeee512) with G trained on the correct world. At drift 0.45 the pooled target is untrained 0.482 +/- 0.055 (mechanical floor AT CHANCE - the fix holds; per-seed 0.483/0.414/0.549), predictor 0.573 +/- 0.057, survival 0.777 +/- 0.122 (per-seed 0.853/0.636/0.841). This is the project's FIRST genuine survival-specific dissociation: survival beats predictor by +0.204 and untrained by +0.295 (both >> the 0.05 SESOI), the shuffled negative control is at chance for every arm (survival 0.556, so it is real signal not probe overfitting), matched-pair 0.906, speed positive control 0.966, engagement 100%, manipulation check survival-relevant. The pipeline verdict prints H MET (encoding induced, conditional on gates). HONEST CAVEATS (do not oversell): the mean 0.777 clears 0.65, but the t-based 90% CI [0.648, 0.906] just STRADDLES the bar (the printed percentile bootstrap [0.705, 0.849] excludes it, but that interval under-covers at the boundary - same issue flagged for the ceiling); 1/3 seeds (0.636) is below 0.65; variance is high (sd 0.122); the L0 TOST is underpowered at n=3 (mean 0.514, not shown equivalent); and the dynamics-rung reward-coupling is not yet ruled out for the organism (needs the leakage audit + a held-out probe). VERDICT: a strong, real, encouraging POSITIVE SIGNAL - the first sign that L3 (a learned-dynamics fingerprint) reverses the L2 nulls - but NOT a settled result at n=3. The n=10 power extension is running (fullruns/l3_n10) to test whether the CI clears 0.65 decisively; the reward/held-out controls remain to be reported before any headline claim.

  • 2026-07-12, AUDITED ORGANISM RUN n=10 + POST-HOC CONFOUND AUDITS - POSITIVE, PARTLY BEHAVIOR-MEDIATED. Re-ran run_expB2 --drift-mode l3 --l3-hidden 8 (n=10, seeds 0..9, 300 updates, local GPU, fullruns/l3_n10_audited) with the pooled-path leakage audit added (PR #39). At drift 0.45 the survival pooled target is 0.752, honest t-based 90% CI [0.698, 0.807] (EXCLUDES the 0.65 bar; per-seed 0.853/0.636/0.841/0.823/0.830/0.573/0.705/0.782/0.759/0.723; 8/10 seeds >= 0.65), with predictor 0.573 and untrained 0.488 (mechanical floor at chance) - survival beats both baselines by well over the 0.05 SESOI, and the L0 authentic-vs-authentic control is 0.517 (chance). This is a determinism reproduction of the fullruns/l3_n10 headline (identical 0.752). CONTROLS THAT PASS: (1) POOLED reward-leakage audit - world identity is NOT decodable from summed reward: reward-leak AUROC mean 0.541, range [0.517, 0.559], clean 10/10 seeds (peer-review item 1, RESOLVED). (2) Per-world survivorship - 0 early deaths in every pool (110/110 authentic and surrogate, all seeds), so dropping early deaths introduces no asymmetry (peer-review item 2, RESOLVED). (3) Nonlinear-probe check (post-hoc, on the dumped states): a random-forest probe on the UNTRAINED agent reads only 0.517 (its shuffled baseline 0.506; L0 untrained 0.483), so the survival-vs-untrained dissociation is NOT a linear-probe artifact - the world is genuinely near-chance-decodable in the untrained state even nonlinearly. CONTROL THAT PARTLY MEDIATES - BEHAVIORAL MEDIATION (post-hoc, on the dumped states): the agent moves and forages differently in the two worlds, so per-episode behavior alone (speed/energy/food/drag) decodes the world at ~0.69 (0.689 linear, 0.704 nonlinear). Controlling for behavior CLEANLY (behavior->state regression fit IN-FOLD, no leakage) leaves a behavior-independent world-signal of ~0.66 (0.676 linear basis, 0.659 quadratic; per-seed quad 0.71/0.58/0.71/0.69/0.73/0.54/0.62/0.70/0.67/0.64, 6/10 >= 0.65) - clearly above the 0.488 untrained floor but only AT the 0.65 bar. So behavior mediates ~0.09 of the 0.752 headline; a real behavior-independent component survives, but it is weak-to-moderate, not the abstract world-identity direction the raw 0.752 suggests. (An in-SAMPLE residualization gave a deflated 0.56-0.63 by over-removing via leakage; the in-fold ~0.66 is the defensible figure. Still a soft UPPER bound: only 4 coarse per-episode MEANS were controlled, so per-timestep behavior would likely lower it further.) The reward audit passed because reward is one scalar; behavior is multivariate and a much stronger separator. VERDICT: a real, reward- and survivorship-controlled, nonlinear-robust POSITIVE - L3 (a learned-dynamics fingerprint) reverses the L2 nulls and the survival agent (uniquely) carries world-discriminative state, read out and never rewarded. But it is only PARTLY behavior-independent (~0.66 after clean behavior control, right at the 0.65 bar), so the strong "abstract world-identity direction at 0.752" reading is NOT supported; the honest headline is "survival-relevant state that differs by world, plus a modest (~0.66) behavior-independent world component." STILL OWED: (a) a richer per-timestep behavior control to tighten the behavior-independent estimate; (b) the pre-registered hidden=4 second-capacity replication (section 11); (c) the held-out/common-garden probe (a world-signal that transfers to an UNSEEN fingerprint is much harder to dismiss as behavior).

  • 2026-07-12 - BEHAVIOR AUDIT MADE REPRODUCIBLE (code, no new runs). The ad hoc mediation audit above is now canonical committed code (itasorl/behavior_audit.py + scripts/audit_behavior_mediation.py, in-fold controls, control properties unit-tested on synthetic ground truth). Re-run on the same dumps it reproduces the published survival d=0.45 figures exactly: target 0.752, behavior-only 0.689 linear / 0.705 nonlinear (prose said 0.704; random-forest seed wiggle), controlled 0.676 linear / 0.659 quadratic (artifact: artifacts/expB2/behavior_audit_l3_n10.json). The dump now also persists PER-TIMESTEP behavior traces (bta/bts), and the script runs the strictly stronger per-timestep control when traces are present, with a decision rule fixed in advance (docs/specs/2026-07-12-l3-behavior-audit-design.md): survival per-timestep-controlled mean >= 0.65 strengthens the claim; [0.60, 0.65) weakens it to a below-bar trace; < 0.60 means largely behavior-mediated. Owed items (a) and (b) reduce to two human-launched runs (scripts/README.md, "L3 owed runs"): hidden=8 with traces (headline capacity) and hidden=4 n=10.

  • 2026-07-13 - PER-TIMESTEP BEHAVIOR CONTROL: CLAIM STRENGTHENED (hidden=8 re-run with traces). Re-ran the frozen hidden=8 n=10 protocol with trace-extended dumps (fullruns/l3_h8_traces, local GPU). The pipeline is deterministic end-to-end: every published figure reproduced exactly (survival target 0.752 [0.704, 0.797]; behavior-only 0.689/0.705; per-episode controlled 0.676/0.659; L0 mean 0.517, ROPE equivalence accepted; reward-leak clean 10/10; 0 deaths, all pools 110/110). The pre-registered decision rule then evaluated the strictly stronger per-timestep control: survival resid_trace = 0.726 (90% CI [0.685, 0.765], 9/10 seeds >= 0.65; quadratic variant 0.721 [0.678, 0.760]) -> >= 0.65, STRENGTHENS, and unlike the per-episode estimate the CI EXCLUDES the 0.65 bar. Honesty checks on real data: the untrained agent's resid_trace is 0.498 (exact chance) even though untrained BEHAVIOR alone decodes the world at 0.645, so the control neither manufactures nor spares signal; predictor resid_trace 0.574 (the survival-only dissociation is preserved under the control). Two notable structural facts: (1) the full behavior trace alone decodes the world at 0.803, BETTER than the state probe itself - behavior is highly world-discriminative; (2) the per-episode-mean control had been OVER-removing (0.676), exactly the attenuation the synthetic tests predicted, while the surgical per-timestep control leaves ~0.73. The honest headline improves to: reward- and survivorship-controlled, nonlinear-robust, with a behavior-independent world-signal of ~0.73 whose CI clears the pre-registered bar. Caveats: the residualization is linear/quadratic over phi = [b_t, b_(t-1), cummean(b)]; a full-history or nonlinear control could in principle remove more. Artifact: artifacts/expB2/behavior_audit_l3_h8_traces.json. STILL OWED: the hidden=4 second-capacity replication (section 11) and the held-out/common-garden probe.

  • 2026-07-13 - HIDDEN=4 RUN UNINFORMATIVE (gate failure); GATE 0 RE-VALIDATED PER CAPACITY AND THE SECOND CAPACITY RE-FROZEN AT HIDDEN=7. The owed hidden=4 n=10 run (fullruns/l3_h4_traces, local GPU) completed and FAILS its gates at drift 0.45: untrained mechanical floor 0.891 (not chance), pooled reward-leak 0.637 with the clean gate passing in 0/10 seeds, engagement passing in only 30% of seeds (the drift-trained policy barely recovers: train@0.45 eval@0.45 return -1.069, vs -0.219 at hidden=8). Per the section-8 decision matrix the run is UNINFORMATIVE (gate-0 / engagement failure), NOT a negative; the L0 control (0.517, TOST and ROPE both accept equivalence) confirms the apparatus itself is fine. ROOT CAUSE: hidden=4's in-band status (oracle 0.940) was frozen from the 2026-07-10 calibration made on the WRONG world at sigma=0.05, BEFORE the world-params fix; the post-fix re-calibration froze gate 0 only at hidden=8 and swept the untrained floor over capacities 8-32, so hidden=4 was never re-validated on world P (its reward leak was itself anticipated in the gate-0 entry above). FIX: gate 0 is now a committed, runnable check (scripts/run_expA_l3.py) validating BOTH halves per capacity on world P at the frozen sigma=0.02 - the residual oracle (band [0.85, 0.95], mechanical leakage) AND the organism-side untrained floor (|target - 0.5| < 0.1 at drift 0.45, arm built exactly as run_expB2 builds it) - with a hidden=8 regression check and a selection rule frozen in advance: the smallest in-band capacity below 8 with a clean floor (section-9 fallback (a), bisection between the bracketing capacities 4 and 8). RECALIBRATION RESULT (fullruns/l3_gate0_recal/calibration.json): hidden=8 regression EXACT (oracle 0.928, floor 0.482); hidden=4 oracle 0.932 in-band but floor 0.896 (independently reproducing the organism run's 0.891 from 3 seeds); hidden=5 oracle 0.972 (OUT of band); hidden=6 oracle 0.946 in-band but floor 0.647; hidden=7 oracle 0.922 in-band, mechanical leakage clean, floor 0.566 -> second capacity RE-FROZEN at hidden=7. Honest note: the hidden=7 floor (0.566) is elevated relative to hidden=8 (0.482) though within the frozen tolerance; the SESOI already requires survival > untrained + 0.05, so the elevated floor is priced into adjudication. The hidden=7 n=10 organism run with trace dumps is running (fullruns/l3_h7_traces); its result and the per-timestep behavior audit will be recorded here when complete.

  • 2026-07-14 - SECOND-CAPACITY REPLICATION AT HIDDEN=7 (n=10): THE BEHAVIOR-INDEPENDENT SURVIVAL SIGNAL REPLICATES (~0.72, CI clears the bar) BUT THE SURVIVAL-VS-PREDICTOR DISSOCIATION DOES NOT - THE FULL "ENCODING INDUCED" VERDICT IS ARTIFACT-CONDITIONAL. The re-frozen hidden=7 n=10 run (fullruns/l3_h7_traces, local GPU, frozen protocol, trace dumps) completed with ALL gates passing: engagement 10/10 seeds at drift 0.45 (pooled train@0.45 eval@0.45 return -0.734, a partial recovery sitting between hidden=8's -0.219 and hidden=4's failing -1.069; the per-seed engagement criterion passes in every seed), L0 control 0.517 (TOST p=0.010 and ROPE P=0.999 both accept equivalence), speed positive control 0.959, pooled reward-leak 0.567 clean in 10/10 seeds, 0 deaths in every pool (110/110 both worlds, all seeds), and the untrained mechanical floor pooled at 0.586 (dev 0.086, inside the frozen <0.1 tolerance and matching the recalibration's 0.566, though per-seed it is violated in s3 0.732 and s9 0.683 with s6 0.605 marginal). PRIMARY at drift 0.45: survival pooled target 0.737, 90% CI [0.688, 0.780], 8/10 seeds >= 0.65 (per-seed 0.540/0.748/0.820/0.799/0.696/0.623/0.749/0.768/0.770/0.852) - above the 0.65 bar and clearing untrained (0.586) by +0.151, BUT predictor reads 0.714 [0.687, 0.740], so survival leads the predictor by only +0.023 and the section-8 requirement (> predictor + 0.05) is NOT met: the clean hidden=8 dissociation (survival 0.752 vs predictor 0.573) does not replicate at hidden=7. BEHAVIOR AUDIT (artifacts/expB2/behavior_audit_l3_h7_traces.json, frozen decision rule): survival resid_trace = 0.722 (90% CI [0.678, 0.763], 8/10 seeds >= 0.65; quadratic 0.704) -> >= 0.65 with the CI excluding the bar, an almost exact replication of hidden=8's 0.726 - the behavior-independent survival world-signal is STABLE across the two in-band artifacts. The dissociation is not: predictor resid_trace 0.691 (vs 0.574 at hidden=8) and untrained resid_trace 0.579 (vs 0.498 = exact chance at hidden=8). READING (per the section-11 two-capacity clause): the hidden=7 artifact is qualitatively coarser - mechanically leakier (untrained floor 0.586/resid 0.579) and far more behaviorally salient (the behavior trace alone decodes the world at 0.762-0.796 in ALL arms, including untrained, vs 0.645 untrained at hidden=8) - so at this capacity every trained agent picks the fingerprint up and the survival-SPECIFIC part of the claim is conditional on the subtler hidden=8 artifact. The cross-capacity finding that survives both runs is: a reward-clean, survivorship-clean, behavior-independent world-signal of ~0.72 in the survival agent's state at both frozen capacities. Secondary notes: matched-pair survival mp_target 0.814; volatility readout target_var 0.706 / target_full 0.763 (above the 0.65 bar, consistent with the volatility-encoding observation on earlier rungs). LAST OWED ITEM, the held-out/common-garden probe (a world-signal that transfers to an UNSEEN fingerprint), is now run and recorded in the entry below.

  • 2026-07-14 - HELD-OUT FINGERPRINT + COMMON-GARDEN PROBE (n=10): THE WORLD-SIGNAL GENERALIZES TO AN UNSEEN SAME-RECIPE FINGERPRINT (TRANSFER POSITIVE) BUT DOES NOT SURVIVE A COMMON-GARDEN CONTROL (REACTIVE, NOT A PERSISTENT REPRESENTATION). Two readout-only evaluation channels on one frozen hidden=8 training run (fullruns/l3_h8_heldout, local GPU, frozen protocol, spec docs/specs/2026-07-14-l3-heldout-common-garden-probe-design.md); no change to training, the surrogate family, or the pre-registered headline probe. GATES REPRODUCED (third independent determinism check): survival pooled target 0.752 (90% CI [0.704, 0.797]), byte-matching the published hidden=8 headline; L0 control 0.517 (TOST p=0.010, ROPE P=0.999 both accept equivalence); pooled reward-leak 0.541 clean in 10/10 seeds; 0 deaths in every pool (110/110 both worlds, all seeds). CHANNEL 1, UNSEEN-FINGERPRINT TRANSFER: the world-identity direction fit against the trained hidden=8 fingerprint, frozen, then scored on a FRESH authentic pool vs the held-out hidden=7 fingerprint the agent never trained against. Survival transfer_target = 0.773 (90% CI [0.722, 0.824], 9/10 seeds >= 0.65), predictor 0.633 (3/10), untrained mechanical floor 0.569 (0/10). Frozen decision rule (survival >= 0.65 AND > untrained + 0.05 = 0.619) PASSES on both clauses -> GENERALIZES beyond the single trained fingerprint instance. Scope caveat, stated in the spec and honored here: hidden=7 is the SAME surrogate recipe at a different capacity, not a different surrogate family; cross-recipe transfer is out of scope for this run. CHANNEL 2, COMMON GARDEN: 20-step prefix in either the authentic or the hidden=8 surrogate world, then both groups continue under IDENTICAL authentic dynamics for a 24-step tail; the probe reads tail-only state, labels = prefix world, GroupKFold. Survival cg_tail_target = 0.557 (90% CI [0.492, 0.622], 1/10 seeds >= 0.65), predictor 0.409, untrained 0.377. Frozen decision rule (survival >= 0.65 AND > untrained + 0.05) FAILS on the first clause. DECAY CHECK on the last 8 tail steps: survival cg_latetail_target = 0.492 (90% CI [0.431, 0.553]), at chance -> the prefix-world signal washes out along the shared tail. READING (frozen rules applied): the L3 world-signal is NOT an overfit to the one artifact instance G_0 (transfer generalizes), but once the felt dynamics are made identical it does not persist in tail-only state - so the signal reads as REACTIVE tracking of the currently-felt dynamics, not a persistent stored world-identity representation. This is an informative negative for the representational reading (spec: "not a gate failure"), and it resolves the long-standing reactive-vs-representational ambiguity (section-11 caveats) toward reactive. This was the last owed post-hoc item; the L3 arc is complete. Provenance addendum (2026-07-16): the behavior audit re-run on this bundle's fresh --save-agents dumps reproduces resid_trace 0.726 (90% CI [0.685, 0.765], 9/10 seeds) identically; per-seed values committed as artifacts/expB2/behavior_audit_l3_h8_heldout.json. Tooling: audit_behavior_mediation.py skips the heldout sibling dumps (*_h7transfer.npz, *_cg.npz) it does not parse.

  • 2026-07-14 - REVERSE-TRANSFER RUN ANNOUNCED (frozen BEFORE launch). The spec's staged follow-up condition is met (forward transfer positive), so the reverse direction runs next: train at hidden=7, hold out the SUBTLER hidden=8 fingerprint (fullruns/l3_h7_heldout, n=10, frozen protocol, --save-agents, trace dumps). Decision rules are IDENTICAL in form to the forward spec and frozen here: transfer - survival transfer_target >= 0.65 AND > untrained transfer + 0.05 -> the world-signal generalizes to a subtler unseen fingerprint (the stricter version of the 0.773 claim); common garden - same form, second data point on persistence at the coarser training artifact. INTERPRETATION LIMIT, stated in advance: at hidden=7 the survival-vs-predictor dissociation did not hold (two-capacity entry above), so this run reads the survival arm against the untrained floor only and cannot support any survival-SPECIFICITY claim. Determinism expectation: the standard-probe half must reproduce the fullruns/l3_h7_traces table exactly (survival 0.737 [0.688, 0.780]); any deviation invalidates the run. Result to be recorded here when complete.

  • 2026-07-16 - REVERSE-TRANSFER RUN RECORDED (completed 2026-07-15): THE COARSE-TRAINED DIRECTION READS THE SUBTLER UNSEEN FINGERPRINT ONLY PARTIALLY - THE FROZEN RULE FAILS (INFORMATIVE NEGATIVE); HELD-OUT TRANSFER IS DIRECTION-DEPENDENT. Executes the 2026-07-14 freeze above (fullruns/l3_h7_heldout, n=10, --save-agents; per-seed summary committed as artifacts/expB2/heldout_l3_h7_reverse_summary.json). DETERMINISM GATE PASSES: the standard-probe half reproduces the fullruns/l3_h7_traces table exactly - survival pooled target 0.737 (boot 90% CI [0.688, 0.780]), L0 0.517 (TOST p=0.010 / ROPE accept equivalence), reward-leak 0.567 clean 10/10, engagement 10/10, 0 deaths per pool - so the run is valid per the freeze. TRANSFER (train hidden=7, hold out the subtler hidden=8): survival transfer_target = 0.638 (t 90% CI [0.600, 0.676], 4/10 seeds >= 0.65) vs untrained mechanical floor 0.525 and predictor 0.603. The frozen rule (>= 0.65 AND > untrained + 0.05 = 0.575) FAILS on the absolute bar; the floor-margin clause alone passes (+0.063). Adjudicated NEGATIVE per the frozen rule. COMMON GARDEN: survival cg_tail_target = 0.598 (t 90% CI [0.547, 0.649], 4/10 seeds >= 0.65), decaying to 0.489 on the last-8 window; predictor 0.504, untrained 0.456. Rule FAILS -> REACTIVE, a second independent data point matching the forward run. READING (frozen rules applied; the freeze's interpretation limit honored - no survival-specificity claim at hidden=7, and predictor transfer 0.603 indeed sits near survival 0.638): held-out transfer is DIRECTION-DEPENDENT. Fit on the subtle hidden=8 fingerprint the direction reads coarser unseen artifacts (same-recipe 0.773, cross-recipe 0.684); fit on the coarse hidden=7 fingerprint it reads the subtler one only partially (0.638, above the floor but below the bar). The published "fingerprint-GENERAL" wording is qualified accordingly in FINDINGS 10.6 and README.

  • 2026-07-16 - RESEARCH-INTEGRITY AUDIT: FREEZE-TIMING NOTE FOR THE HIDDEN=7 SELECTION. A commit-level audit of this log found: the capacity-fallback RULE was frozen in the original pre-registration commit (0217263, 2026-07-10), strictly before the hidden=4 gate failure it was later applied to. The specific "RE-FROZEN at hidden=7" statement, however, landed in the same commit as the hidden=7 n=10 result artifact (85229c2, 2026-07-14), so the capacity selection and its result are commit-simultaneous rather than sequential. The honest reading: hidden=7 was selected by a rule frozen before the failure, but the selection itself has no committed timestamp preceding the run. The per-timestep behavior-control spec has clean sequential timing (spec 1b5985e 2026-07-12, results bf8a74e 2026-07-13). Every published number was re-verified against committed per-seed artifacts by scripts/audit_stats_recheck.py (117 checks, all passing); all 15 cited arXiv IDs were resolved against arxiv.org with matching titles and authors. See docs/AUDIT_2026-07.md.

  • 2026-07-16 - CROSS-RECIPE TRANSFER PROBE (n=10, readout-only): THE WORLD-SIGNAL READS A DIFFERENT SURROGATE FAMILY (RULE PASSES); THE CONSTANT-DRAG FAMILY IS UNCALIBRATABLE AND DROPPED. Executes the pre-registered spec docs/specs/2026-07-15-l3-crossrecipe-transfer-probe-design.md (frozen before any run; PR #51) against the SAVED fullruns/l3_h8_heldout agents; no training anywhere. Committed artifact: artifacts/l3_crossrecipe/summary.json. GATE-0 CALIBRATION: the G_rff family's round-1 D sweep {8..128} stepped over the band (0.975 at D=64, 0.624 at D=128); the pre-registered bisection fallback landed D=80 in-band (oracle 0.887, leakage pass, untrained floor 0.538). The G_cd family has an EMPTY calibration window: by the eps where its oracle reaches the band (eps=3.2: 0.854; eps=6.4: 0.907) the untrained mechanical floor exceeds 0.6 (0.859, 0.913) - a drag-coefficient bias large enough to detect is felt so grossly that any recurrent state separates the worlds. DROPPED per the pre-stated rule (recorded, no transfer claim, no penalty to the primary). Both cd calibration double-runs produced byte-identical JSONs. INTEGRITY GATE (fourth determinism check): all 60 reloaded agents regenerated their standard pools bit-identically against the saved dumps and the drift-0.45 pooled survival mean reproduced the published 0.752 exactly. CHANNEL (PRIMARY, G_rff D=80): survival transfer_rff_target = 0.684 (boot 90% CI [0.657, 0.710]; t 90% CI [0.654, 0.715], lower bound above the bar; 7/10 seeds >= 0.65, per-seed min 0.584), predictor 0.574 [0.554, 0.593] (0/10), untrained mechanical floor 0.548 [0.538, 0.557] (0/10). Frozen rule (survival >= 0.65 AND > untrained + 0.05 = 0.598) PASSES on both clauses, machine-checked in the runner aggregate (rff_rule_pass true, margin +0.034). READING: the world-identity direction fit against the trained MLP fingerprint reads a different-recipe fingerprint (cosine-basis ridge texture) the agent never lived with, survival-specifically; attenuated versus same-recipe transfer (0.773 -> 0.684), as expected for a farther family. Scope, stated honestly: the reactive interpretation of the common-garden entry above is UNCHANGED - this extends the generality of the reactive world-signal across surrogate recipes, it does not revisit persistence. The cd drop is itself informative: coefficient-bias artifacts cannot be made subtle-but-detectable in world P, independently motivating the learned-texture construction of the L3 rung.

  • 2026-07-18 - CORRECTION (OUTCOME-BLIND): COMMON-GARDEN CHANNEL RE-SCORING UNDER THE FIXED ESTIMATOR. The 2026-07-14/16 held-out and reverse-run adjudications' common-garden channel (cg_tail 0.557 / late-tail 0.492; reverse 0.598/0.489) was computed with the pre-fix pair-splitting cg_probe (per-episode CV groups; bias toward AUROC 0 whenever the surviving pair count is not a multiple of 5 - fixed in commit 1633bca; see FINDINGS 13.C). The committed artifacts carry the bias signature (drift-0 cg floors 0.001-0.27; pair counts 96-110). The cg channel is being re-scored from the saved _cg.npz tail dumps with the fixed estimator (scripts/reanalyze_cg_states.py); the frozen decision rules are UNCHANGED and will re-adjudicate the corrected numbers. The transfer channel (frozen-fit train/test, no CV) and the mp channel (pair counts 60/25, fold-safe; no pair drops) are unaffected and stand. This entry is recorded BEFORE the re-score's outcome is known; the "REACTIVE" adjudication is provisional until it lands.

  • 2026-07-18 - OPERATIVE-ESTIMAND ERRATA. Section 6 registers the "selectivity-corrected AUROC" as the estimand; every adjudication (0.752, 0.737, 0.726, 0.722 vs the 0.65 bar) used the raw pooled/resid target with the shuffled probe as a gate-style control. No verdict could have flipped (shuffled at chance throughout); the operative definition is the raw target adjudicated at the bar, selectivity reported as a control.

  • 2026-07-22 - CORRECTION: THE MP CHANNEL WAS NOT FOLD-SAFE. The 2026-07-18 entry above asserted the matched-pair channel "(pair counts 60/25, fold-safe; no pair drops)" was unaffected by the 1633bca estimator fix. That assertion was checked (per FINDINGS 10.6's "confirm, do not assume" banner) and is FALSE: under the sklearn in use (1.5.2), GroupKFold with the pre-fix singleton groups splits matched-pair twins at both pair counts (33/60, 24/25), and the recorded drift-0.00 mp floors carry the 13.C bias signature (0.134-0.137 instead of ~0.5). Both held-out bundles' mp channels were re-scored from rollouts regenerated off the saved agent bundles (scripts/reanalyze_mp_readout.py; deterministic - all 120 cells reproduce their stored pre-fix values bit-for-bit under singleton-group scoring). Under the fixed pair-grouped estimator all drift-0.00 floors return to 0.500 exactly; corrected drift-0.45 survival means are 0.893 (forward, was 0.883) and 0.855 (reverse, was 0.814 - the value quoted in the 2026-07-16 entry). The bias deflated, never inflated. The mp channel is secondary/demoted (no frozen rule reads it), so no adjudication changes; the corrected aggregates are committed as artifacts/expB2/heldout_l3_h8_mp_rescore.json and artifacts/expB2/heldout_l3_h7_reverse_mp_rescore.json. The transfer channel's exemption in the 2026-07-18 entry stands (frozen-fit train/test, no CV grouping).

  • 2026-07-24 - A2 OBSERVATION-CHANNEL LOCALIZATION (n=10, readout-only): WORLD-IDENTITY SIGNAL IS VISUAL, NOT INTEROCEPTIVE at hidden=8. Executes scripts/run_l3_obs_localization.py against the saved fullruns/l3_h8_heldout agents; no training. The frozen agent's observations are channel-masked before the running norm and new drift-0.45 pools are collected under the hidden=8 fingerprint. Baseline (no mask) survival mean 0.753 reproduces the headline. Masking vision (120/146 dims) drops the signal to 0.686 (7/10 seeds >= 0.65). Masking interoception (14/146 dims) leaves it at 0.756 (8/10). Masking all observations collapses it to 0.500. READING: the world-identity signal is carried by the visual stream / visually-shaped behavior, not by the explicit velocity/energy interoceptive channels. This is consistent with section 10.4's behavior- mediation finding; it does not distinguish passive visual echo from a visually- guided behavior plan. The probe's positive controls (speed, energy, food) remain readable in every mask condition, so the chance collapse is not a dead-probe artifact. Local artifact: fullruns/l3_h8_obs_localization/aggregate.json.

  • 2026-07-24 - A2 OBSERVATION-CHANNEL LOCALIZATION (n=10, readout-only): HIDDEN=7 SIGNAL IS ROBUST TO SINGLE-CHANNEL MASKING. Executes scripts/run_l3_obs_localization.py with --hidden 7 against the saved fullruns/l3_h7_heldout agents; no training. The frozen agent's observations are channel-masked before the running norm and new drift-0.45 pools are collected under the hidden=7 fingerprint. Baseline (no mask) survival mean 0.737 reproduces the headline. Masking vision (120/146 dims) leaves the signal at 0.764 (9/10 seeds >= 0.65). Masking interoception (14/146 dims) leaves it at 0.742 (8/10). Masking all observations collapses it to 0.500. Unlike hidden=8, where vision masking causes a clear drop, the hidden=7 signal is robust to single-channel lesion: the coarser surrogate's artifacts propagate into both visual and interoceptive streams. The probe's positive controls (speed, energy, food) remain readable in every mask condition, so the chance collapse is not a dead-probe artifact. Local artifact: fullruns/l3_h7_obs_localization/aggregate.json.

  • 2026-07-25 - A3 L1 H2 SUBSTRATE-GROUNDING BATTERY (n=10): ORGANISM ENCODING NEGATIVE at matched in-band discretization. Extends the L3 H2 battery to the L1 observation-quantization rung. GATE 0: fine Δ sweep under B-v2 world params freezes Δ = 0.023 (oracle AUROC 0.873, leakage clean; fullruns/l1_calib.json). Noise comparator gate freezes σ_o = 0.01 (oracle AUROC 0.873, leakage clean; variance/high-frequency oracle, not the L1 grid residual; fullruns/l1_noise_calib.json). ORGANISM RUN: B-v2 protocol with --drift-mode l1 --l1-delta 0.023 --sensor-sigma 0.01, drifts {0.0, 0.023}, seeds 0..9, three arms, --save-agents + state dumps (fullruns/l1_heldout). Survival pooled target 0.533 (90% CI [0.509, 0.556], 0/10 ≥ 0.65); predictor 0.489; untrained 0.494. L0 ROPE accepts chance; engagement/leakage/survivorship clean; ceilings alive. Primary H_B2 NOT met - oracle-detectable L1 leaves no incidental world-identity encoding. A1 GRADED-SEAM: integrity gate bit-matches dumps and reproduces 0.533; Δ ladder stays near chance (0.522 / 0.519 / 0.528 / 0.547 / 0.533). A2 NOISE KNOCKOUT: transfer onto matched-band iid obs noise gives survival 0.525 / untrained 0.509 / predictor 0.516 (frozen rule fails both clauses). A2 CHANNEL LOCALIZATION: baseline 0.533; intero mask ~0.51; all-mask 0.500; vision mask elevates survival and untrained alike (~0.70) - non-specific lesion artifact given the null baseline. READING: L1 does not reproduce the L3 positive; H2 substrate-grounding remains L3-specific. Detectability of an observation-level seam is not sufficient for incidental encoding. Artifacts: fullruns/l1_heldout/, fullruns/l1_h2_ablations/, fullruns/l1_obs_localization/.

  • 2026-07-23 - H2 TEXTURE-KNOCKOUT: STRUCTURE + DOSE-RESPONSE ABLATIONS (n=10, readout-only): H2 CONFIRMED IN TEXTURE-SPECIFIC FORM at hidden=8 and hidden=7. Executes the frozen spec docs/specs/2026-07-22-h2-substrate-grounding-ablations-design.md (PR #99) against the saved fullruns/l3_h8_heldout and fullruns/l3_h7_heldout agents; no training. GATE-0 CALIBRATION: G_gn (Gaussian-jitter velocity law) freezes at sigma_v=0.01 with oracle AUROC 0.865 (band pass), leakage pass, and untrained floor 0.448 (within tolerance). The same-recipe MLP capacity ladder {h16, h32, h64} is calibrated sub-band with clean floors and leakage: oracle AUROCs 0.788, 0.656, 0.603. INTEGRITY GATE (sixth determinism check at hidden=8, seventh at hidden=7): all 60 reloaded agents per capacity regenerated their standard pools bit-identically and the drift-0.45 pooled survival means reproduced the published 0.752 (hidden=8) and 0.737 (hidden=7) exactly. CHANNEL 1 (structure knockout, PRIMARY): at hidden=8 survival transfer_gn_target = 0.539 (0/10 seeds >= 0.65), predictor 0.556, untrained floor 0.542; at hidden=7 survival 0.510, predictor 0.539, untrained floor 0.520. The frozen positive rule fails both clauses at both capacities, so the direction does NOT read matched-band unstructured jitter. CHANNEL 2 (dose-response, SECONDARY): at hidden=8 survival transfer means h16 0.701, h32 0.622, h64 0.541 (seeds >= 0.65: 8/10, 3/10, 0/10); at hidden=7 h16 0.574, h32 0.516, h64 0.513 (1/10, 0/10, 0/10). VERDICT: H2 is supported in texture-specific form across both in-band capacities; the world-signal loads on the learned structure of the substrate velocity-law surrogate, not on generic dynamics perturbation. The survival-specificity claim remains conditional on the subtler hidden=8 artifact (section 10.5).

13. How to run (milestones, in order)

  1. Build + calibrate the surrogate. Train G on authentic rollouts; wrap as a World; run the Experiment-A oracle; ascend capacity until oracle AUROC in [0.85, 0.95]; FREEZE and record in section 12. (CPU-friendly; no GPU sweep.)

  2. Gate check. Only if gate 0 passes, proceed. Otherwise re-tune G.

  3. Organism run (GPU, mirrors B-v3), dumping states for offline re-probing:

    python scripts/run_expB2.py --drift-mode l3 --dump-states runs/l3/states
python scripts/run_e2e.py --b2-drift-mode l3 --b2-dump-states runs/l3/states
# capacity ceiling reference:
python scripts/run_expB2.py --drift-mode l3 --sysid-aux
# power extension (n=10) via the seeds 0..9 profile

    (The l3 drift-mode and the G World are new code, gated behind this preregistration; no multi-hour organism sweep before milestone 1's oracle gate passes.)