AModels / Code
MOAT v5g Benchmark Code · Stage 1–2d Results · Full Relay Log
appendix/moat_v5g_results_summary.py
MOAT v5g — Results Summary
Current
Consolidated results summary across all Stage 1–2d experiments.
Aggregates AUC metrics, attribution angle statistics, and directional energy contrasts.
Python · .py
appendix/moat_v5g_stage2d_fixed.py
MOAT v5g — Stage 2d (Fixed)
Corrected multi-directional replay table. Fixes criterion from v_Q < isotropic
to v_B > v_Q; renames "discriminative oracle" to v_B-oracle.
Python · .py
appendix/moat_v5g_stage2d.py
MOAT v5g — Stage 2d
Superseded
Multi-directional replay table: six action sources replayed into both environments.
Initial version prior to criterion correction.
Python · .py
appendix/moat_v5g_stage2c.py
MOAT v5g — Stage 2c
Policy-matched replay experiment. Demonstrates AUC drop 0.762 → 0.553
under neutral B̂ = I; confirms policy geometry as the AUC source.
Python · .py
appendix/moat_v5g_stage2b.py
MOAT v5g — Stage 2b
Attribution angle experiment. Compares dominant direction of B_est − I
against v_B and v_Q at episode end; confirms 88.8% misattribution rate under H_Q.
Python · .py
appendix/moat_v5g_stage2.py
MOAT v5g — Stage 2a
Endogenous directional depletion experiment. Removes wrong_strength;
SRAAgent runs from B̂₀ = I and generates DE_B contrast 0.420 autonomously.
Python · .py
The code and experiment descriptions were generated and revised through the relay. Claude produced the main implementation drafts; Codex and ChatGPT provided review, claim-boundary corrections, and interpretation checks. stage2d_fixed and v5g_results_summary are the current reference versions.
BRelay Materials
Relay Log — From Observer-Agent Theory to MOAT v5g
log/moat_log.txt
Full Relay Log
Complete dialogue relay documenting the transition from the earlier Observer-Agent /
SRA framing to the MOAT v5g benchmark. Includes the pruning of consciousness,
observer, and high-level philosophical claims; the shift toward ABHT-adjacent
benchmark framing; and the Stage 1–2d experimental reinterpretation of attribution collapse.
Full text