docs: propose mutation falsification evidence harness

[tnunamak]

What

Proposes a design-only mutation-falsification evidence pilot. It changes no production code.

Version 1 is a trusted, local, operator-supervised experiment. It:

• preserves the existing migration oracle lifecycle and adds structured evidence only;
• separates intent, machine observations, computed projections, and independent triage;
• uses RFC 8785 plus SHA-256 for integrity binding without claiming authenticated identity;
• keeps focused GroupMe checks outside test-accounting authority and retains each complete authority bundle before deleting its independent clone;
• defines conservative killed, survived, and inconclusive projections, batch-scoped clean-baseline reuse, bounded evidence retention, and stop criteria;
• isolates writable clone, home, temporary, cache, dependency, Git, and accounting paths;
• defers any shared coordinator, sandbox, generic executor, StrykerJS, CI gate, agent-generated mutation, or test-deletion authority.

The earlier systemd/cgroup coordinator design was rejected. This revision removes it instead of weakening or overstating its guarantees. The developer harness makes no sandbox, automatic crash-recovery, or PDPP product-host portability claim.

Review status

Two fresh independent reviews of 65a4ff0f7 returned LAND:

• contract review, confidence 0.90;
• implementation-realism review against the current test-accounting and package-tooling seams, high confidence.

The prior fe3d965d8 LAND comment is superseded because later adversarial reviews found load-bearing containment and receipt-retention defects in that revision.

Validation

• pnpm exec openspec validate add-mutation-falsification-evidence --strict
• git diff --check
• DCO and Assisted-by: AI trailers on every commit
• clean worktree after commit

This draft targets the pinned red-team audit base. Retarget or transplant it to the published integration candidate before landing.

Assisted-by: AI

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
pdpp	Ready	Preview	Aug 11, 2026 9:11pm

[tnunamak]

Independent final re-review of fe3d965d8: LAND, confidence 0.93.

The revision closes the prior blockers:

• manager-owned wrapper claims and binds the attempt before adapter descendants start;
• delayed launch, launcher death, PID reuse, and descendant cleanup fail closed;
• Linux cgroup-v2/systemd containment is explicit and unsupported hosts refuse execution;
• workspace bytes are an observed soft threshold, not a false hard quota;
• clean and mutant backstop policy is consistent;
• structured migration-oracle evidence precedes the trusted GroupMe pilot and decision gate.

Implementation watchpoints: fsync the containing directory after active-marker link/replacement, test crash points around both operations, and describe CPUQuota as throttling rather than kill-on-overage.

[tnunamak]

Superseding the earlier fe3d965d8 LAND comment: subsequent independent adversarial reviews correctly returned REVISE. That version lost authority evidence with disposable workspaces, overclaimed systemd/resource semantics, and built a shared coordinator before the pilot justified one.

Revision 65a4ff0f7 removes the coordinator and narrows the change to a trusted, operator-supervised evidence pilot. It retains complete accounting bundles, defines total conservative projections and bounded baseline reuse, isolates writable state in independent clones, and explicitly disclaims sandbox, automatic crash-recovery, and product-host portability claims.

Fresh independent contract and implementation-realism reviews both return LAND. Named strict OpenSpec validation and diff checks pass. Keep this PR draft and design-only until it is retargeted to the published integration candidate.