docs(research): add verified learning loop roadmap

[TheHalfMoon]

Purpose

Persist the post-Spec-006 improvement program in repository truth so the project can follow it without relying on chat history.

Adds

• docs/research/010-verified-learning-loop-roadmap.md
• explicit gap assessment for experiment/learning architecture
• research synthesis covering Autoresearch, Hermes, Reflexion, Voyager, ACE, Darwin Gödel Machine, PaperBench, RE-Bench, SpecBench, and Faraday/Replica
• proposed future invariants for verified learning
• target architecture with a separate Experiment Plane
• conceptual VerifiedExperience, LearningProposal, SkillVersion, Experiment, and LearningDecision records
• protected holdout/evaluator-independence requirements
• skill lifecycle with canary, invalidation, revocation, and rollback
• ordered L0-L10 future program and dependency chain
• first post-Spec-006 walking skeleton
• repository-integration rules requiring major improvements to become durable repository truth

Scope firewall

This PR is research-only and non-authorizing.

It does NOT:

• widen Spec 006 or T079/T080;
• authorize runtime/model calls;
• add migrations, dependencies, runtime code, memory code, or learning code;
• amend the Constitution;
• create a future Tasks contract;
• authorize learned routing, fine-tuning, RL, world models, self-modifying core behavior, or automatic landing.

The current Spec 006 implementation sequence remains authoritative. The roadmap explicitly requires Spec 006 to close canonically before a future learning program can enter Constitution -> Spec -> Plan -> Tasks -> Implement.

Strategic decision preserved

Any agent. Verified work. Verified learning.

The recommended future loop is:

Verified work
-> Verified experience
-> Learning proposal
-> Bounded champion/challenger experiment
-> Protected holdout + independent evaluation
-> Learning evidence
-> Explicit promotion/canary
-> Monitor
-> Invalidate/rollback when stale or regressing

Relationship to historical PR #21

PR #21 remains historical research input and should not be merged unchanged as the future plan. After Spec 006 closes, its learning/memory research should be reconciled against current canonical repository truth and this newer roadmap before formal specification.

---

Summary by cubic

Adds docs/research/010-verified-learning-loop-roadmap.md, a research-only roadmap for a future verified learning loop, so the post-Spec-006 improvement program lives in repository truth rather than chat history. The document does not change behavior or authorize any implementation.

Scope firewall

• Research-only and non-authorizing; it does not widen Spec 006, T079, or T080.
• No runtime/model calls, migrations, dependencies, or runtime/memory/learning code.
• Does not amend the Constitution or authorize learned routing, fine-tuning, RL, world models, or self-modifying core behavior.
• Spec 006 remains authoritative and must close canonically before this roadmap can drive a formal program.

What the roadmap covers

• Gap assessment for experiment/learning architecture and research synthesis across Autoresearch, Hermes, Reflexion, Voyager, ACE, Darwin Gödel Machine, PaperBench, RE-Bench, SpecBench, and Faraday/Replica.
• Proposed future invariants, a target architecture with a separate Experiment Plane, and conceptual VerifiedExperience, LearningProposal, SkillVersion, Experiment, and LearningDecision records.
• Ordered L0–L10 future program with a dependency chain and a first L0–L3 walking skeleton, plus repository-integration rules requiring major improvements to land as durable repository truth.
• Historical PR docs(research): preserve agent intelligence and learning archive #21 remains research input and should not be merged unchanged; reconcile it against this roadmap after Spec 006 closes.

^{Written for commit 4ce5617. Summary will update on new commits.}

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

_{Comment @coderabbitai help to get the list of available commands.}