Date: 2026-08-28
This follow-up records repository changes that landed while the targets/goals/tasks planning turn was still active.
Two previously in-review components are now on main:
PR #78 makes the multi-attempt control plane real: separate attempt histories, independent verifier outcomes, worker failure isolation, semantic replay, no automatic candidate selection, and no merge authority.
PR #81 makes evaluator control content-addressed and bound to the exact Work Unit/source revision and exact required-validator set before positive decision support is possible.
The canonical trust chain is now:
WorkUnit
-> ResultManifest
-> EvaluatorPlan
-> VerificationResult
-> human/integration decision
A control-plane authority audit found that both verification CLI entrypoints could write --output to arbitrary repository-relative paths:
experiments/local_verifier.py;experiments/evaluator_plan_runner.py.This meant a component described as read-only/evidence-only could target a tracked file such as README.md.
PR #90 was opened to make executable authority match the intended authority:
verification output -> root results/ only
Root results/ is already ignored by .gitignore.
PR #90 adds fail-closed self-tests for canonical-path output attempts and preserves existing candidate-evaluation and EvaluatorPlan semantics. Both Phase 0 schema check and Evaluator plan binding CI pass on head b955d7474841ca018cc4e0878f542542841239d2.
No self-merge or self-approval was performed.
The main execution issues were updated again to avoid stale descriptions:
The principal physical/runtime blocker is still:
PR #34 canonical node
-> synchronize with current main
-> controlled Docker acceptance #37
-> independent sandbox/path review
The assistant environment does not provide the controlled Docker host required by #37, so that gate was not fabricated or claimed.
After PR #90 safety review, #5 Phase B1 should extend the canonical EvaluatorPlan-bound verifier to the real node smoke bundle without executing candidate code.
Minimum independently observed checks:
allowed_paths / forbidden_paths enforcement;Useful patch mechanics explored in closed PR #61 should be extracted into the canonical verifier path rather than resurrecting a second verifier package.
GitHub metadata still reports main as unprotected. Issue #35 therefore remains the highest integration-governance safety gap. Repository files and agent instructions cannot substitute for a GitHub-enforced ruleset/branch-protection boundary.
Growth Seed #28 has completed and is closed after the IDKGraph P0 decomposition landed. The repository should continue to gate new Growth Seeds on verified descendant evidence and reviewer capacity rather than issue volume.
1. Protect main (#35) when admin/settings action is available.
2. Independently review/integrate green PR #90.
3. Synchronize #34 and complete controlled Docker gate #37.
4. Build #5 real patch-bundle evaluator through canonical EvaluatorPlan path.
5. Connect real node adapter to merged PR #78 orchestrator.
6. Build minimal Evidence Report/replay UX for #16.
7. Run real-task #2/#30 only after comparable real evidence exists.
Synthetic R2/R3 scale/evolution research may continue as research, but it should not be promoted as proof of real Verified Swarm Runner quality before this execution chain is complete.