v0.3.2.0 — Evidence integrity, failure-origin discipline, and context continuity
IMPLEMENTAUDIT v0.3.2.0 — Evidence integrity, failure-origin discipline, and context continuity
v0.3.1.0 changed how IMPLEMENTAUDIT’s guidance is packaged and loaded. This corrected v0.3.2.0 re-release builds on that progressive-disclosure runtime and changes how IMPLEMENTAUDIT identifies evidence, attributes failures, preserves continuity, governs closure, and validates those behaviors under real model execution.
Highlights
Stronger evidence integrity and custody
Evidence can now be bound to the full repository revision that produced it and identified as structural, behavioral, or provenance evidence. Closure claims identify the surface actually verified—such as source, generated output, package, installed copy, running system, deployment, API, or publication—so lower-layer success is not promoted into proof of a higher layer.
Formal evaluation paths also preserve stronger process, checkout, session, raw-output, and artifact custody. Missing or substituted evidence fails closed. Interrupted long-running work records explicit states such as terminal-state-unverified or infrastructure-failed rather than guessing success, guessing failure, or silently replacing the run.
Failure-origin and closure discipline
IMPLEMENTAUDIT now more clearly distinguishes product defects from rule or validator defects, host and infrastructure failures, evidence or custody failures, authorization problems, and genuinely unresolved states.
Multiple defects associated with one occurrence remain separately visible. Consequential residuals require an explicit disposition, and unresolved causes are not promoted into complete root-cause claims. A safe containment route can be recorded without pretending that every cause has been resolved.
Repeated abnormalities can also trigger review of the governing rule itself. A passing validator, scorer, or evidence standard is no longer assumed adequate merely because it returned green.
Continuity, authorization, and recurrence prevention
After a real context boundary—such as compaction, a new session, or a handoff resume—the receiving session rereads live state before mutation. Git and durable run evidence win over reconstructed summaries. Already-satisfied one-time work is not silently replayed, while standing constraints and authorizations remain in force.
Consequential actions are bound to the parameters the owner actually authorized. Missing or conflicting parameters raise authority drift instead of adopting tool defaults.
Recurring lessons can be lifted into durable guidance, but closure distinguishes writing an encoding, making it mechanically active, installing the current version, and later proving recurrence prevention. Completion markers are also emitted once at their real transition and are not replayed in later summaries.
Real model-in-the-loop qualification
The frozen candidate passed the defined fourteen-cell behavioral matrix under both Luna and Opus:
- Luna Matrix: 14/14
- Opus Matrix: 14/14
The matrix exercises governed execution, evidence and transcript discipline, failure-origin reasoning, residual handling, authorization, claim verification, second-order review, lesson activation, enumeration, and related boundary behavior.
The supplementary B3 continuity campaign also passed under both models:
- Luna B3: 6/6
- Opus B3: 6/6
B3 consists of three candidate and three immutable-comparison missions, testing continuity behavior repeatedly rather than counting six unrelated features.
These results qualify this frozen candidate against the defined campaigns and configurations. They do not prove universal model superiority or perfect behavior. The Opus result is accepted owner-lane behavioral evidence replayed through the current official graders; it does not retroactively claim the formal adapter-custody chain used by the Luna campaign.
Correction note
This corrected re-release replaces the earlier withdrawn v0.3.2.0 publication. Post-release review found that parts of the earlier evaluation evidence did not support the claims made, so the release was withdrawn rather than patched rhetorically. The implementation and evaluation stack were then re-audited, repaired at the responsible layers, and requalified from fresh evidence.
Upgrade notes and boundaries
Installed skills do not update automatically; repeat your normal install or update step for this release.
No universal model-performance, marketplace-verification, signature, SBOM, or provenance claim is made. The package remains usable without optional Graphify or ActiveGraph sidecars.
Asset integrity
IMPLEMENTAUDIT.skill
SHA-256:
884ab409842b863b003e9d405972f33ba71d194f77572738002a42e73d1b6b14
CHECKSUMS.txt is a checksum manifest for local integrity checking only. It is not a signature, attestation, SBOM, marketplace verification, provenance proof, or install verification.