This repository contains the reference implementation, experimental framework, and supporting materials for Internal Attribution Collapse, a failure mode in adaptive systems where:
- External behavior remains distinguishable.
- Internal attribution becomes structurally incorrect.
- Policy adaptation alters the geometry of collected evidence.
The central result is:
External classifier distinguishability and agent-internal attribution correctness are independent.
An adaptive agent may produce trajectories that are externally classifiable while internally learning the wrong structural explanation for the observed phenomenon.
Traditional evaluation often assumes:
High predictive performance
↓
Correct understanding
This work demonstrates a counterexample.
An agent can:
- Adapt successfully.
- Produce apparently separable evidence.
- Remain structurally wrong about the underlying cause.
The resulting evidence becomes dependent on the geometry induced by the policy itself.
README.md moat_v5g_stage2d_fixed docs/ index.html theory.html appendix.html appendix/ moat_v5g_stage2 moat_v5g_stage2b moat_v5g_stage2c moat_v5g_stage2d moat_v5g_stage2d_fixed results/ moat_v5g_results_summary ja/ docs/ index.html theory.html appendix.html
A failure mode in which:
- evidence remains externally distinguishable,
- but updates are mapped into the wrong latent channel.
Observed separability may arise from policy-induced trajectory geometry rather than correct causal attribution.
A benchmark framework designed to investigate:
- attribution fidelity,
- directional collapse,
- adaptive policy feedback,
- policy-dependent evidence generation.
A constructive minimal counterexample demonstrating:
- endogenous directional depletion,
- structural misattribution,
- divergence between external and internal evaluation.
Internal attribution failure:
- 88.8% of adversarial episodes align with the wrong structural direction.
Policy-matched replay:
- Adaptive AUC: 0.762
- Replay AUC: 0.553
The apparent signal is largely policy-geometry dependent.
AUC varies as a function of action direction.
This suggests that evidence quality cannot be interpreted independently from policy geometry.
This project is not presented as a replacement for existing approaches such as Active Bayesian Hypothesis Testing (ABHT).
Instead, it provides:
- a benchmark,
- a failure-mode analysis,
- a constructive counterexample,
for studying the gap between:
- external evidence quality,
- internal attribution fidelity.
Stage 1–2d: Completed
Stage 2e:
Residual AUC collapse inside the adaptive loop remains an open question.
Both positive and negative outcomes are scientifically informative.
Draft status.
Multi-AI review relay:
- Claude
- Codex
- ChatGPT
- Gemini
- Perplexity
Baseline comparisons and further validation are ongoing.
Research prototype. Use at your own risk. The agent may be confidently wrong. That is, unfortunately, part of the point.