Skip to content

Latest commit

 

History

History
113 lines (106 loc) · 9.64 KB

File metadata and controls

113 lines (106 loc) · 9.64 KB

Writeup outline and claims inventory

Structured summary of the project's results. Every claim must trace to a row in the claims inventory below; every row traces to a committed artifact. Update this file as sections 10.5 (second capacity) and 10.6 (held-out probe) of FINDINGS.md resolve.

Structure

  1. Introduction (from ITASORL.md Part I)
    • The reframed question: authentic base dynamics vs generative surrogate, detection read out, never rewarded.
    • The white space: social simulacra and evaluation-awareness work is semantically contaminated; this is substrate-grounded and tabula rasa.
    • Contribution: (a) the detectability-vs-encoding gap as a measurable object, (b) a robust L2 negative under survival pressure at scale, (c) a controlled L3 positive with a behavior-independent component, (d) the preregistered, gate-driven methodology itself.
  2. The world and the surrogate ladder (from ITASORL_world_spec.md; FINDINGS 1)
    • "A Patch of Earth" v0; ~146-dim observation; invariants (tabula rasa, readout-not-reward, substrate-grounded).
    • Ladder L0-L4; difficulty calibrated by an oracle band, not by fiat.
  3. Methods (from PREREGISTRATION*.md; FINDINGS 2, 9, 10.1, 11)
    • Experiment A oracles, matched pairs, leakage-audit battery.
    • Experiment B probes; three-agent design (untrained/predictor/survival) on an identical trunk; pooled + matched-pair readouts.
    • Gates: oracle band, engagement, L0 equivalence (TOST/ROPE), leakage.
    • Preregistration and deviation logging as first-class method.
  4. Results
    • 4.1 Detectability ceilings (FINDINGS 2): L1 1.000, L2 0.993, L3 gate 0.928.
    • 4.2 The L2 negative arc (FINDINGS 3, 4, 9): chance across channels, horizons, probes; survival coupling 0.610 [0.585, 0.634]; capacity ceiling 0.596.
    • 4.3 The L3 positive (FINDINGS 10): 0.752 [0.698, 0.807] survival-only; controls; behavior mediation; 0.726 behavior-independent (holds at 0.723 once absolute position and heading join the control basis, FINDINGS 10.4.1).
    • 4.4 Replications: second capacity resolved (FINDINGS 10.5): the behavior-independent signal replicates (0.722 vs 0.726) but the survival-vs-predictor dissociation is artifact-conditional; held-out fingerprint resolved (FINDINGS 10.6): capacity-variant transfer passes (0.773 vs untrained 0.569; same recipe and data - robustness within one recipe, per the 10.6 scope note; the generalization claim rides on the cross-recipe probe) and the common-garden control, re-scored with the fixed estimator, passes the frozen rule on both directions (0.666 forward, 0.684 reverse), so the signal is a modest persistent world-identity component, not purely reactive (FINDINGS 10.6.1); cross-recipe resolved (FINDINGS 10.7): the direction also reads a different surrogate family (RFF ridge, 0.684 vs untrained 0.548, rule passes), while the constant-drag family proved uncalibratable (empty gate-0 window) - coefficient-bias artifacts cannot be made subtle-but-detectable, independently motivating learned-texture surrogates.
  5. Limitations (FINDINGS 11, verbatim spirit)
  6. Related work (ITASORL.md Part I literature sections, incl. the 2026-07 axis-by-axis cross-reference: evaluation awareness behavioral and probe-based (arXiv:2309.00667, 2407.04694, 2505.23836, 2507.01786, 2509.13333), incidental-encoding probes (AtariARI arXiv:1906.08226; Othello-GPT arXiv:2210.13382), agent-side mismatch detection (GalilAI arXiv:2110.15489; RAPT arXiv:2602.01515), and necessity-of-world-models theory (Richens et al. arXiv:2506.01622). State the novelty claim exactly as worded in ITASORL.md's novelty posture; list the five preempted overclaim phrasings as what we do not claim.)
  7. Reproducibility statement (FINDINGS 12; artifacts/expB2/ promotion history; CITATION.cff)

Claims inventory

# Claim Number Doc section Artifact Figure
1 L0 control at chance (oracle) AUROC 0.523 FINDINGS 2.1 artifacts/expA/summary.json expA_ceiling.png
2 L1 detectable (oracle) AUROC 1.000 FINDINGS 2.1 artifacts/expA/summary.json expA_ceiling.png
3 L2 detectable (oracle) AUROC 0.993 FINDINGS 2.2 artifacts/expA/summary.json expA_L2_ceiling.png
4 L2 not encoded (recurrent state) 0.510 ± 0.039 @ drift 0.45 FINDINGS 3.1 artifacts/expB/summary.json expB_incidental.png
5 L2 surprise channel weak 0.596 ± 0.007 @ drift 0.45 FINDINGS 3.2 artifacts/expB/summary.json expB_channels.png
6 No liftoff with horizon 0.48-0.51 across 0/8/16 FINDINGS 3.3 artifacts/expB/summary.json (recorded rerun; see 3.3 correction) expB_kstep.png
7 L2 null robust (nonlinear probe) 0.482 ± 0.031 FINDINGS 3.4 artifacts/expB/summary.json -
8 Survival coupling does not rescue L2 0.523 ± 0.045 (replication) FINDINGS 9 artifacts/expB2/expB2_results.json expB2_survival.png
9 B-v3 regime negative at scale 0.610, 90% CI [0.585, 0.634], n=10 FINDINGS 7.1 artifacts/expB2/bv3_n10_summary.json -
10 L2 capacity ceiling below bar 0.596, 90% CI [0.577, 0.616], n=10 FINDINGS 7.1 artifacts/expB2/sysid_ceiling_n10_summary.json -
11 L3 gate frozen in-band oracle 0.928, floor 0.483 FINDINGS 10.1 PREREGISTRATION_L3.md sec. 12 -
12 L3 encoded by survival only 0.752, t 90% CI [0.698, 0.807], 8/10 seeds FINDINGS 10.2 artifacts/expB2/behavior_audit_l3_h8_traces.json -
13 L3 predictor baseline near chance 0.573 [0.546, 0.599] FINDINGS 10.2 artifacts/expB2/behavior_audit_l3_h8_traces.json -
14 L3 untrained floor at chance 0.488 [0.461, 0.514] FINDINGS 10.2 artifacts/expB2/behavior_audit_l3_h8_traces.json -
15 Reward leak clean 0.541, clean 10/10 FINDINGS 10.3 PREREGISTRATION_L3.md sec. 12 (n=10 audited entry) -
16 No survivorship asymmetry 0 early deaths, 110/110 all pools FINDINGS 10.3 PREREGISTRATION_L3.md sec. 12 -
17 Behavior alone decodes world trace 0.803 [0.763, 0.840] FINDINGS 10.4 artifacts/expB2/behavior_audit_l3_h8_traces.json -
18 Behavior-independent component (four-channel; position/heading added in 10.4.1) 0.726 [0.685, 0.765], 9/10; quad 0.721; position/heading-controlled 0.723 [0.682, 0.760], 8/10, quad 0.700 FINDINGS 10.4, 10.4.1 artifacts/expB2/behavior_audit_l3_h8_traces.json; artifacts/expB2/behavior_audit_l3_covar_n10.json -
19 Control neither manufactures nor spares signal untrained resid 0.498; predictor 0.574 FINDINGS 10.4 artifacts/expB2/behavior_audit_l3_h8_traces.json -
20 Second capacity: behavior-independent signal replicates 0.722 [0.678, 0.763], 8/10; quad 0.704 FINDINGS 10.5 artifacts/expB2/behavior_audit_l3_h7_traces.json -
21 Second capacity: dissociation NOT met (artifact-conditional) survival 0.737 [0.688, 0.780] vs predictor 0.714 [0.687, 0.740]; lead +0.023 < +0.05 FINDINGS 10.5 artifacts/expB2/behavior_audit_l3_h7_traces.json -
22 Gate 0 re-validated per capacity; hidden=7 frozen oracle 0.922, floor 0.566; hidden=8 regression exact (0.928/0.482); hidden=4 uninformative FINDINGS 10.5 PREREGISTRATION_L3.md sec. 12 + scripts/run_expA_l3.py -
23 Held-out capacity-variant transfer passes (same recipe/data; robustness within one recipe, 10.6 scope note) survival 0.773 [0.722, 0.824], 9/10 vs untrained floor 0.569; rule passes FINDINGS 10.6 artifacts/expB2/heldout_l3_h8_summary.json -
24 Common-garden control: re-scored, MODEST PERSISTENT (frozen rule passes both directions) survival cg_tail 0.666 forward / 0.684 reverse (>= 0.65 and > untrained + 0.05); late tail 0.586/0.577 decays below bar FINDINGS 10.6.1 artifacts/expB2/heldout_l3_h8_cg_rescore.json -
25 Reverse transfer direction-dependent: coarse-to-subtle FAILS survival 0.638, frozen rule fails FINDINGS 10.6 artifacts/expB2/heldout_l3_h7_reverse_summary.json -
26 Cross-recipe transfer: GENERALIZES across surrogate families (the generalization claim) survival 0.684, t 90% CI [0.654, 0.715], 7/10 vs untrained 0.548; rule passes, margin +0.034 FINDINGS 10.7 artifacts/l3_crossrecipe/summary.json -
27 Exp C milestone-3 pilot null INVALIDATED (wrong-world + biased estimator); H3 open pending re-run contrast +0.020 (not a valid measurement) FINDINGS 13, 13.C artifacts/expC/emergence_pilot_summary.json (provenance caveat: run predates artifact's stamped commit) -

Known gaps

  • Claims 1-7 resolved (2026-07-16): the recorded 06302026 e2e bundle's step metrics (plus the 2026-07-13 k-step rerun and the across-seed stds recovered from the bundle log) are promoted to artifacts/expA/summary.json and artifacts/expB/summary.json by scripts/promote_ab_summaries.py.
  • Claims 9-10 resolved (2026-07-16): per-seed pooled targets promoted to artifacts/expB2/bv3_n10_summary.json and artifacts/expB2/sysid_ceiling_n10_summary.json by the same script; scripts/audit_stats_recheck.py re-verifies every inventory number against the committed artifacts.
  • Rows 23-24 resolved (spec 2026-07-14, per-seed summary committed as artifacts/expB2/heldout_l3_h8_summary.json): transfer generalizes, and the re-scored common-garden control passes the frozen rule on both directions (row 24; forward 0.666, reverse 0.684; a modest persistent world-identity component, not purely reactive, FINDINGS §10.6.1); the writeup reports both as generality checks, not as a new headline. Rows 20-22 resolved 2026-07-14: the second capacity replicates the behavior-independent signal and bounds the survival-specific claim to the subtler artifact.