[aDragon0707]

Summary

Pi integration can now bind a selected test-success claim to a complete event lifecycle and independently tested delivery. A successful tool result or final message cannot supply its own host evidence.

Pi JSON events + actual process exit
  + independently recorded delivery digest
  + trusted external contract / copied-source unittest run
    -> SACP host observation / selected claim -> derived verdict

Adds a no-prompt installed-Pi probe, bounded JSON capture API (model execution denied by default), parser, trusted-fixture oracle, demo and 35 Pi tests. Updates bilingual entry docs, architecture/development/status, ADR implementation progress and adapter classification. Required CI needs neither Pi nor model access.

Evidence

• Before: deterministic engine fixtures did not collect Pi events or independently test delivered source.
  After: python -m examples.pi_evaluation uses the same scripted success trace for correct and incorrect code: host_completed / allowed, then unverified / blocked. This is a synthetic replay, not a real-agent run.
• python scripts/run_tests.py --suite core --fail-on-skip --json .local/pi-integration/core.json: 136 passed, zero failures/errors/skips.
• Same runner for langgraph, protocol, harness: 7 tests, 49 documents, 8 breakers + 3 positive controls; all passed.
• Pi-specific tests cover truncation, malformed/unknown events, retries, tool identity, independent good/bad source acceptance, stale delivery reuse, missing/zero/skipped/expected-failure tests, timeout, no-model default and scripted process capture limits.
• Actual installed Pi 1.0.0 / Node 24.18.0: version and RPC get_state passed, exit 0, prompts/model calls requested 0, no inherited credentials. CLI entry digest alone is not full package pinning. No real task JSON trace has been captured.
• Documentation links/fences/public-path check and git diff --cached --check: passed.

Review record: base c06106ba2bf2f6dbc5303ca0cdacdb48613e04bb, head 6c910db; classes adapter, automation, docs. Primary-agent review checked evidence promotion, execution opt-in, independent oracle inventory, cross-platform process handling and claim boundaries. The policy script change only classifies new Pi paths, adds no permission or execution exception, and is automation-critical. Trusted default-branch policy does not yet know these new adapter paths; this review explicitly includes them. No SACP/0.1 protocol changes or protocol companion waiver.

Pre-merge evidence checked: PR CI run 38687709522 has all eight required jobs successful. Downloaded all six suite artifacts: core 136 on Ubuntu 3.11/3.12 and Windows 3.12; LangGraph 7, protocol 49, harness 8+3, zero failures/errors/skips. Trusted receipt run 38687777151 produced strict-valid COMPLETE/200, bound to head 6c910db29239db3ec2d18fa8e012c12b3b6fe895, merge-test SHA ad0c9f39fb72328bb4d482e1f1ef2d3299551548, tree fingerprint sha256:90349199d0853b725b1c41b3f6c1b39fbbe665789a5e180a083b237d10126115, with no evidence errors. Receipt classes are automation, docs under old main policy; adapter scope was explicitly reviewed here. Decision: proceed with manually initiated squash merge and final main verification.

Merge Danger

Door: two-way

Additive experimental package; no dependency/workflow changes. Reports are not official wire receipts or release approvals. Local acceptance is explicitly for trusted fixtures: Python process separation is not a hostile-code sandbox. Caller-provided evidence is not authenticated. Capture launchers declare isolation; they do not attest it. Real execution still needs enforced isolation, full runtime pin, provider/model, credentials and budget. User requested integration only, with no model calls; none were made.

Blast Radius: evaluation

Manually squash only after inspecting CI jobs and SHA-bound trusted receipt. Revalidate final main SHA after merge; preserve branches and local evidence.

Post-merge result: squash commit d6b3864d4b3a403ab735b322230251d1c84ec589; main-verify run 38687891756 succeeded on all seven jobs. Downloaded and checked all six evidence artifacts: core 136 on all three configured environments, LangGraph 7, protocol 49, harness 8 breakers + 3 positive controls, zero failures/errors/skips. Strict-valid post-merge COMPLETE/200 receipt binds the final main SHA and matching tree fingerprint, with no evidence errors. Remote integration branch preserved; local worktree clean. No model execution occurred.