[PIR][WP19] Self-distilled entity-grounded skill synthesis in the Darwin loop (SkillForge pattern, ADR-321)

Part of the RuV Perpetual Intelligence Runtime (PIR), Wave 2. Epic: #837. Plan: docs/research/perpetual-intelligence-runtime/07-wave2-program-plan.md · evidence: 06-wave2-evidence-review.md (both land on main with the feat/pir-wave2-adrs PR).

ADR mapping

Proposed ruvector ADR-321 — self-distilled, entity-grounded skill synthesis inside the Darwin mutation loop. Extends merged ruvector ADR-313 (SHAPER, WP9 #841) directly: a mutation-generation strategy that synthesizes test-covered-functionality issues from the target repo itself (rather than waiting for real incidents), distills entity-grounded skills, and feeds them into the existing frozen-weight harness-evolution loop as one more candidate source.

Evidence

SkillForge (arXiv:2608.18933) — "SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution," submitted 2026-08-19. Grade B+ — the mechanism is confirmed verbatim, but the headline "beats issue-resolution baselines" claim is qualitative only: no percentage figure appears in the abstract, so the magnitude of improvement is not independently verifiable.

⚠️ Because of the B+ grade, no performance number may be quoted for SkillForge anywhere in this WP's deliverables — not in the ADR, code comments, commit messages, or status updates. There is no verified number to quote; any figure that appears would be fabricated or sourced from unverified material. The only citable performance claim is this program's own research-gate-measured delta.

Artifact available now: github.com/cslsolow/SkillForge (verified via GitHub API: public, 2.5 MB, real content — distilling/, synthesis/, src/, evaluation_result/) — adapt the released implementation rather than rebuilding from the paper.

Scope

• Add synthetic-issue-from-test-coverage skill distillation as a Darwin mutation-candidate source.
• Surfaces: harness/src/darwin.ts, examples/mragent scorePolicy — the same surfaces WP9 already owns; coordinate with WP9's remaining scope.

Candidate-mutation rule (applies to all of Wave 2)

Candidate mutation, not trusted prior art. Promotion requires this program's own research-gate paired-bootstrap-recomputed delta over the pre-WP baseline — never the paper's self-reported (here: nonexistent) numbers.

Acceptance criteria

• SkillForge-pattern candidate source wired into the Darwin loop (ADR-321)
• research-gate-measured delta over the pre-WP baseline recorded before any promotion
• Zero quoted SkillForge performance numbers anywhere in the deliverable (review-enforced)
• security scan after landing (touches the mutation-proposal surface)
• ADR-321 merged with repo-qualified citations per ADR-305 §4

Depends on: WP9 (#841) — the SHAPER-pattern loop must exist to receive this as a candidate source. Phase W2-3.

Start with harness/src/darwin.ts and the examples/mragent scorePolicy surface, then read ADR-313, ADR-321, and the Wave 2 program plan. Review the released SkillForge artifact before defining the candidate source. Done means the source is wired into the Darwin loop, the research-gate delta is recorded, no SkillForge performance numbers are quoted, ADR-321 is merged, and a security scan passes.