Producer checks after the oh-my-pi audit
Updated on 2026-09-05 after the operator asked whether the machinery was ready for a rerun. The preceding triage identified the application failures and missing clarifications. The procedure changes below are implemented; Git history records their commit boundary.
Deployed behavior and interfaces
The canonical producer skill is loaded through the existing skill projections. Its input/output paths, authority, lens interfaces and publication commands are unchanged. Its final integrated check now requires scope agreement, coverage of every qualifying learning route, justified push selection, and stable canonical identities. The producer records checked routes and material dispositions in its existing Semantic verification section. An unsupported known classification blocks publication; explicit uncertainty within a coherent scope can remain valid.
The result contract now makes the relevant distinctions explicit: local/display material versus opaque payloads, generated continuation summaries versus raw logs, actual selection versus names on requested objects, and canonical identity before final acceptance. The C10 audit applies the same questions to published results. The producer performs its check on the unfinished exact result; it does not call C10's published-input reader before publication.
Producer step 2 now requires bounded source ranges and aggregate output budgeting. Step 3 specifies new IDs for split records after sharing, keeping the combined record's identity explicit. Sparse overlays and the existing execution-preflight policy remain the applicable rules for repetition and unexecuted tests.
Composition search covered the epistemic lens, legacy writer, transfer scan, placement pass, landscape synthesis and C10. Their argument/result protocols remain compatible. C10 consumes the refined classification semantics and was updated with the producer and type. No schema, Python implementation, publication gate, extra review artifact or skill metadata changed.
Verification boundary
Current method SHA-256:
| File | SHA-256 |
|---|---|
Producer SKILL.md |
1ff5cf0013a5eb193c8a806be0e69b9f756246219e02d9085543650c9d55c24e |
| Main-result type | 144931a68679f793c7748c17d8be58fdd4faa6a5711408033c7841753254d598 |
| C10 instruction | 0f762399374ee51e9ec5bf1c8ac0da091ec85b452933773fba9503c2ee547def |
Inputs for the bounded semantic exercise are the full retained results from
AAS-2026-09-05-oh-my-pi-01 and AAS-2026-09-05-apache-maka-01. Their exact
result and public-review hashes match the preceding
oh-my-pi audit and
Maka acceptance.
The strict reader still accepts both structurally. Neither input is corrected
or republished by this exercise.
uv run pytest: 723 passed. Existing instruction-composition, result-type,
publication and comparison tests passed. Changed library Markdown validates
without warnings. No wording-matching tests were added.
The skill-creator generic quick validator rejects argument-hint, context,
model, type and user-invocable. The exact same rejection occurs for the
committed baseline, and frontmatter is byte-identical. These existing
Commonplace fields pass the repository validator; this is a generic-validator
compatibility limit, not a regression introduced by the edit.
Independent behavioral verification completed. The checker received the revised producer/type and both exact results, without the session report, workshop audit, expected findings or proposed replacement values. It performed a report-only check and bounded pinned-source reads, without changing inputs, running target code or publishing.
| Exercise | Independent finding |
|---|---|
| oh-my-pi scope/form | Block publication: the profile's complete form set and the local-summary lens scope do not account consistently for RTE-4's provider replay/archive branch. The opaque form itself remains unknown. |
| oh-my-pi trace learning | Block publication: RTE-4/OBJ-2/BAP-1 establish a qualifying retained-summary route that learning_scope.note excludes as reshaping. Assess its dependent axes; do not guess its task horizon. |
| oh-my-pi push selection | RTE-3's automatic branch/entry selection supports identifier. Catalog names alone do not; clarify RTE-12 attribution without removing the independently supported token. |
| Maka task horizon | Block publication under the clarified rule: learning_scope at result lines 76–81 and lens line 282 infer per-task from session-bounded continuation. RTE-13 and its source anchor establish session/event selection, not a task boundary. Further evidence or explicit uncertainty is required; cross-task is not an automatic replacement. |
| Maka learning branches | Block publication: the lens at line 282 attributes learning to text checkpoints, while distilled_form at lines 88–93 includes the opaque provider branch. RTE-12/OBJ-4/RTE-13 need a consistent account of which branches qualify and contribute to every dependent axis. |
| Legitimate uncertainty | Both results' explicit limits remain valid. Maka's opaque form/lineage, uninspected curation and faithfulness, and weaker afforded pull basis are not defects merely because they prevent stronger claims. |
| Canonical split hypothetical | A combined canonical claim already supplied to a worker keeps its referent and is marked superseded with replacement references; both parts receive fresh monotonic IDs. Neither inherits the old ID. |
For Maka, the checker inspected pinned
packages/runtime/src/prior-run-context.ts:39-74 and
packages/runtime/src/history-compaction.ts:303-325,378-399. Those spans support
session-based selection and shared event/checkpoint/replay machinery; they do
not decide the missing task horizon or opaque representation. This is a
follow-up diagnosis under the clarified method, not a silent rewrite of the
earlier pilot's finding.
The exercise therefore supplies four real blocking diagnoses and supported uncertainty/selection cases. It establishes that the instructions can expose these issues, not that every future producer will resolve them. Remaining route-specific task definitions and coarse/name-filter mappings can be recorded as uncertain; no additional universal taxonomy rule is needed here.
Final method and input hashes match the recorded exercise boundary. The
acceptance and updated workshop records validate without warnings, and
git diff --check passes.
Rerun boundary
Use a new run ID and reload the revised producer skill and result type. Reinspect the frozen oh-my-pi revision under the normal producer workflow; the old completed result supplies questions to investigate, not source evidence to copy. The source boundary, integrated classifications and compact projection must agree before normal publication.
These changes improve the instructions applied by an agent. They do not make the deterministic validator judge semantic support or add an independent reviewer to every production run. The existing oh-my-pi result and the affected Maka learning assessments need regeneration; no rerun has been performed by this verification. Current readers still accept both structurally, so withhold their affected complete comparison claims until source regeneration resolves them.