Define RAVEL recursive experience and causal-learning substrate

[epi13]

Summary

Define an additive post-RAVEL-0.6 experience substrate so recursive development is not reduced to feeding errors into the next candidate.

What changed

• add a full architecture document for experience-governed recursion;
• define five memory classes: episodic, causal, semantic, procedural, and negative;
• define six linked record types:
  • experience_episode;
  • causal_hypothesis;
  • intervention_record;
  • causal_attribution;
  • learned_principle; and
  • strategy_record;
• require successful, erroneous, neutral, and abstaining episodes rather than error-only memory;
• separate episode outcomes, hypothesis disposition, principle maturity, transfer status, strategy reuse, candidate disposition, and verifier/evaluator status;
• define immediate, enabling, delayed-descendant, transfer, retention, and negative-downstream credit classes;
• require competing hypotheses, falsifiers, bounded probes, pre-evaluation predictions, attribution alternatives, transfer-gated reuse, and known strategy failure modes;
• add a machine-readable architecture profile and a linked reference record bundle;
• add an offline standard-library validator and ten deterministic negative fixtures;
• wire the checks into the repository root make check target and the Core Standard CI workflow;
• add strict documentation navigation, RAVEL research links, and a changelog entry.

Why

The existing recursive architecture track governs candidate lineage, evaluation, rollback, policy replacement, and authority. Its feedback object is still primarily structured failure plus predicted and actual effects. That is enough for controlled recursive repair, but not for a system to learn what an event means, distinguish competing causes, preserve successful behavior, understand why an intervention worked or failed, assign delayed lineage credit, or determine when a lesson transfers.

This PR defines that missing substrate without implementing a hidden optimizer or allowing the recursive system to rewrite its judge.

Core lifecycle

observations
  -> structured experience episodes
  -> competing causal hypotheses
  -> bounded diagnostic and counterfactual probes
  -> intervention candidates with predicted effects
  -> immutable evaluation
  -> causal attribution
  -> provisional learned principles
  -> transfer tests
  -> reusable adaptation strategies

Authority and evidence boundary

• Frozen RAVEL 0.4 and 0.5 artifacts are unchanged.
• The RAVEL 0.6 preregistration and final-material boundary are unchanged.
• The existing recursive-architecture comparison preregistration is unchanged.
• The experience system cannot modify evaluator, thresholds, partitions, resource policy, custody, or promotion authority.
• Failed and rejected records cannot be silently deleted.
• Provider or verifier PASS | FAIL | UNKNOWN results cannot be redefined by diagnostic interpretation.
• Compact identity-bearing records are used instead of requiring private chain-of-thought transcripts.
• Formal MNCS and MNCDS status remain UNKNOWN; independent evaluation, protected custody, certification, and promotion remain unestablished.

Executable checks

The validator requires:

• all five memory classes;
• all six record types and required fields;
• all four episode outcome classes;
• linked hypothesis, intervention, attribution, principle, and strategy identities;
• hypotheses recorded before intervention;
• non-in-place candidate replacement and rollback binding;
• explicit attribution credit classes;
• supporting attributions and falsifiers for principles;
• transfer evidence before supported global strategy reuse;
• retained rejected hypotheses and known strategy failure modes; and
• all declared causal-feedback controls.

Negative fixtures must reject:

1. evaluator authority expansion;
2. error-only memory;
3. deleted failure memory;
4. post-hoc hypotheses;
5. unsupported principles;
6. strategies without failure modes;
7. global reuse without transfer support;
8. attribution without a credit class;
9. aggregate-only causal promotion; and
10. access to future-final material before freeze.

Scope

This is architecture, vocabulary, reference data, and semantic validation. It does not yet implement:

• runtime RAVEL episode emission;
• Forge probe operations;
• an intervention generator;
• causal inference algorithms;
• experience retrieval or caching;
• strategy application;
• policy-recursive candidate generation; or
• external final evaluation.

Those require later bounded candidate epochs and their own preregistration.

Validation

Current-head Core Standard results completed so far:

• Ruff formatting and lint: PASS;
• strict mypy: PASS;
• release-gap issue-map validation: PASS;
• recursive experience profile and ten negative fixtures: PASS;
• strict MkDocs and whitespace checks: PASS;
• clean wheel build/install: PASS;
• pinned Python/Rust interoperability: PASS;
• release-candidate checks: PASS.

The ordinary Python-version test/corpus matrix is still completing. The immediately preceding head passed Python 3.12, 3.13, and 3.14 tests, packaged-schema validation, conformance corpus, MNCDS corpus, and examples; the only failure on that head was the six mechanical Ruff findings corrected in the current head.

Review focus

Please focus review on:

• whether the six records are the minimum useful vocabulary;
• whether positive, neutral, and abstention memory are represented strongly enough;
• whether causal attribution is conservative enough about alternatives and controls;
• whether delayed lineage credit remains explicit rather than becoming a hidden reward;
• whether transfer-gated strategy reuse is sufficiently fail-closed;
• whether the boundary between RAVEL, Forge, MNCDS, MNCS, and the recursion governor remains clear; and
• whether this belongs as a post-0.6 companion track rather than being folded into the existing 0.6 hypothesis.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.