Skip to content

docs: Learning Harness design + M7 long-term roadmap entry - #637

Merged
cursor[bot] merged 2 commits into
masterfrom
cursor/learning-harness-design-782d
Oct 7, 2026
Merged

cursor[bot] merged 2 commits into
masterfrom
cursor/learning-harness-design-782d

Conversation

@IanFrelinger

Copy link
Copy Markdown
Owner

Summary

Adds the Learning Harness design document and integrates it into the long-term roadmap as milestone M7 (post-M6). This is a docs-only PR that sketches the architecture for autonomous experience-driven improvement without implementing any production code or weakening existing certification requirements.

Changes

  • New: docs/LearningHarness.md — comprehensive design sketch covering:

    • Core principle: "Agents can learn freely. They cannot trust their own learning freely."
    • Four-loop harness architecture (Experience Record → Reflect → Propose → Gates → Promote/Quarantine)
    • Learning ladder: 9 rungs ordered by blast radius (episodic memory → trust-kernel changes)
    • ILearningHarness port interface with AgentExperience and LearnedRule schemas
    • Replay validation as promotion gate for strategies
    • Fleet learning (federated, distributed trust model)
    • Commercial positioning HOLD until dogfood scorecard unlock
  • Updated: docs/audits/2026-09-completion-roadmap.md

    • New M7 section: Learning Harness (Long-term)
    • 6 work streams with exit criteria (experience store, reflector, replay gate, promotion orchestrator, ladder enforcer, unification)
    • Dependencies: M1–M6 complete + dogfood continuous proof operational
    • Autonomy marketing HOLD: no "autonomous learning in production" claims until thresholds met
    • Executive Summary updated to note M7 as post-M6 long-term feature
  • Updated: docs/DocsIndex.md

    • Added link to LearningHarness.md under "Trust loop / certification" section
    • Clearly marked as long-term design (roadmap, not live claim)

Testing

N/A — documentation only, no code changes.

Testing strategy (blast radius)

N/A — docs-only PR.

  • make kernel-coverage-gate (if touching src/Ashlar.Core.* or src/Ashlar.Infrastructure)
  • make kernel-gate (if touching kernel hosting / pipeline / profiles)
  • make test-prod-style (if touching production DI / API / routing)

Checklist

  • make test passes locally (docs-only, no code to test)
  • Documentation updated (entire PR is documentation)
  • No TODO or NotImplementedException left unresolved (none added)
  • Breaking changes are documented (no breaking changes — docs only)

Release (only when this PR ships a versioned NuGet/GHCR release)

  • Not a versioned release — skip
  • Preflight: dotnet run --project application/src/Ashlar.CLI -- release preflight <semver> (or make release-preflight VERSION=<semver>)
  • Track: open a Release checklist issue (GitHub → New issue → Release checklist form) or link an existing release issue
  • After merge: tag v<semver> and push (runs release.yml) — see docs/RELEASE_RUNBOOK.md

Design Constraints Met

  • ✅ Docs only — no production code
  • ✅ No cert-gate weakening — design explicitly relies on existing gates (M1–M6 dependencies)
  • ✅ No autonomy marketing unlock — HOLD until dogfood scorecard met
  • ✅ Clear roadmap positioning — M7 is post-M6, long-term feature
Open in Web Open in Cursor 

@cursor
cursor Bot force-pushed the cursor/learning-harness-design-782d branch 5 times, most recently from 0c96799 to ec17674 Compare October 7, 2026 04:29
cursoragent and others added 2 commits October 7, 2026 04:39
- Add docs/LearningHarness.md: four-loop harness (experience → reflect → evaluate → promote), learning ladder (9 rungs by blast radius), evidence-chain requirements, replay gate, fleet learning, autonomy HOLD until dogfood unlock
- Update docs/audits/2026-09-completion-roadmap.md: add M7 section with exit criteria, dependencies (M1–M6 + dogfood scorecard), work streams (experience store, reflector, replay gate, promotion orchestrator, ladder enforcer), and autonomy marketing HOLD
- Update docs/DocsIndex.md: link LearningHarness.md under Trust loop / certification section as long-term design (roadmap, not live claim)

Roadmap feature only — no production code, no weakening of cert-gate language, no autonomy marketing unlock

Co-authored-by: IanFrelinger <IanFrelinger@users.noreply.github.com>
…, soften live claims, correct lim-9 status

Redline 1 (unlock criteria — cite scorecard verbatim):
- LearningHarness.md: replace invented "≥90% last-7-days" with actual thresholds from docs/dogfood-scorecard.md (Last-N=10 ≥80%, mean time ≤15min, Strict reject ≤70%, ≥7 consecutive days, **≥7 landed dated Strict PASS rows on master**, hygiene-via-Ashlar)
- completion-roadmap.md M7: same — cite scorecard explicitly, emphasize landed ledger rows on master

Redline 2 (soft phrasing — not live capability):
- LearningHarness.md line 93: "Rungs 1–4 can be autonomous" → "designed to allow Tier 0 autonomy when M7 ships and dogfood unlock criteria are met"
- completion-roadmap.md M7 line 673: same softening

Redline 3 (same-file hygiene — lim-9 is CLOSED):
- completion-roadmap.md: limitation 9 CLOSED 2026-09-13 (banner already stated this); fixed lingering text implying still open:
  - Current Position: vulnerability "remains" → "CLOSED"
  - Key Risk: "publicly documented...must close" → "CLOSED 2026-09-13...M1.1b has been met"
  - M0.1 table: "remains (limitation 9" → "CLOSED 2026-09-13 (limitations 7–8–9 all closed)"
  - NOT Claimable: "limitation 9 remains exploitable" → "adversarial validation incomplete" (no false lim-9 claim)
  - M1.1: entire section rewritten — STATUS: CLOSED, exit criterion met ✅
  - M1.2 dependencies: "limitation 9 closure" → "M1.1 complete (all certification limitations closed)"
  - M1 exit criteria: M1.1b marked [x] complete
  - Risk table: "limitation 9 exploited before M1 close" → "M1.2–M1.4 gate complexity...limitation 9 already closed"
  - Framework team responsibilities: strikethrough "Close remaining...limitation 9" → "COMPLETE"
  - Success metrics: "limitation 9 closed" → "limitations 7–8–9 all closed; 7–8 by PR #523, 9 closed 2026-09-13"
  - Immediate actions: strikethrough "Begin M1.1b (close lim-9)...public and remains exploitable" → "COMPLETE...closed 2026-09-13"
  - Strategic priorities: strikethrough "Closes remaining...limitation 9" → "Limitations 7–8–9 all closed; M1 now focuses on CI proof"
  - Key documents: "limitation 9 open" → "limitations 7–8–9 all closed"

Constraints met: docs only; Autonomy marketing HOLD; no cert-gate weaken; no merge

Co-authored-by: IanFrelinger <IanFrelinger@users.noreply.github.com>
@cursor
cursor Bot force-pushed the cursor/learning-harness-design-782d branch from ec17674 to 868f750 Compare October 7, 2026 04:39
@cursor
cursor Bot merged commit e507ce8 into master Oct 7, 2026
17 checks passed
@cursor
cursor Bot deleted the cursor/learning-harness-design-782d branch October 7, 2026 05:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants