C10 acceptance: main-review classification audit

Accepted on 2026-09-05 for the procedure migration and bounded checks below. The preceding production pilot and publication fix are committed in f586fb1d; this C10 migration and its follow-ups are recorded in Git history.

Consumption and correction boundary

The revised procedure reads complete retained main results through the existing strict reader, then checks whether classifications preserve the recorded mechanisms. It returns a diagnosis with input and method identities. It no longer patches legacy review prose, moves observation dates forward or refreshes the trace-learning survey.

Repository search found no automated caller, skill projection or CLI entry for this instruction. Mentions in the bulk-operations and self-improvement workshop inventories are descriptive. The existing path remains the manual invocation point. The conditional analyse-agentic-system interface already accepts the source identity, revision, target boundary and issue to inspect, so no caller or callee change was needed.

The audit distinguishes supported classifications, evidence limits, classification defects and gaps in the shared method. An existing rule can already cover an application error; regeneration then uses that rule unchanged. When corrections are commissioned, the producer performs fresh source inspection in a new run and publishes normally. The old result supplies the diagnosis, not replacement evidence. These checks exercised diagnosis and routing; they did not execute a production correction or source refresh.

Selected input and method identities

The explicit population is one system: Apache Maka's public main review and its exact retained result. Run: AAS-2026-09-05-apache-maka-01. Source: https://github.com/apache/maka. Recorded revision: ece69ab3e7a1629a6073831005711d8aa7160ca4. Cutoff: 2026-09-05. Evidence tier: code-grounded.

The comparison scope covers built-in retained RuntimeEvents and checkpoints, local MEMORY.md/PENDING.md, atomic MemoryItems, and their keys, provenance and extraction bookkeeping. Goal/skill control material, offloaded tool artifacts, arbitrary project files, external extensions and transient replay arrays are excluded. The audit read the full result's source register, shared records, lenses, reconciliation and limits, alongside its comparison profile.

SHA-256 identities at initial and final checks:

Input SHA-256
Public main review cf2f80113c2c21074cdc07e149b9a9cb3a764f0ff5b6fb39728deee633eac76c
Retained exact result fcd16d145d4ee6730eedab994478c8a320fd98f79123dd2df145c3cb6b8d3c18
kb/instructions/refresh-agent-memory-review-taxonomy.md b384fd14ab74bb18fbac62bedd9f3bc28c8f01f57a8282017b963974c5080baf
kb/instructions/analyse-agentic-system/SKILL.md 7dcb69c3f9d29d940f6d3f627949819e89a6c493340e419e3418ec58d6d390ba
kb/types/agentic-system-analysis-result.md df54e009e603fa9718c478360134160b39bff128e356b832b42ebb43209ce7b0
kb/types/agentic-system-analysis-result.schema.yaml 2c330378e5ff6a5e9f8810c30338b3455aaa1f7a6b3d5ce483626dc5162049d9
src/commonplace/lib/systems_matrix.py 3346be5e1b31ce49921a9fe9697fd810b15b8758794928e3326fa6b415d866e0
kb/notes/definitions/representational-form.md 2779941570fe6726a2832965ce711df45b22e515282443853803a623a7eee5a0
kb/notes/definitions/behavioral-authority.md 9c25df74af0e2df7dae032286300537b7e29b71ddf71b5932fdd3994d46809c9

Bounded live diagnosis

The applicable comparison contract requires a known set to cover the declared scope and use the weakest basis supporting that union. Missing evidence remains explicit. The following are analytical judgments from the retained records, not conclusions computed by the structural reader.

Classification Disposition and support in the retained result
Storage substrate Supported. Files and SQLite preserve the distinct stores in OBJ-1–OBJ-4.
Representational form Evidence limit. The opaque provider checkpoint in OBJ-4 prevents a complete aggregate; the result keeps known natural-language and symbolic parts explicit without guessing its encoding.
Lineage Evidence limit. Opaque, replacement and import provenance prevent a complete scoped set. Established extraction routes do not decide every object's lineage.
Behavioral authority Supported. BAP-2–BAP-4 distinguish advisory payload from validation, routing and enforcement in scoped metadata and write controls. Authority is attached to consumers and operative parts.
Read-back direction Supported. RTE-8/RTE-13 wire push routes; RTE-11 affords a pull API and ABS-1 records no production caller. The union correctly uses the weaker afforded basis.
Trace learning and its dependent axes Supported within the recorded limit. RTE-12/RTE-13 wire automatic text checkpoints into later task continuation, supporting per-task online learning from the event stream. This does not establish cross-task consumption of atomic SQLite memory. Opaque distilled form remains not determinable.
Curation and faithfulness Evidence limit. Recorded individual operations do not decide the full curation set. Faithfulness testing remains uninspected; the audit cannot turn that into either a tested dependence claim or an assertion that no such test exists.

No classification defect was established within this retained boundary. That finding does not establish current upstream behavior, operational success or coverage beyond the selected system. No regeneration was required by the live diagnosis, and no generated result or public review was changed.

Procedure exercise and rejection checks

The Python preflight was extracted from the instruction and executed with the explicit Maka path. Its output recorded both file hashes and the full population identity. The project interpreter ran the snippet because the sandbox's uv launcher could not acquire its required capabilities.

Case Observed result
Live explicit Maka selection Accepted; final population and input hashes matched the initial check.
Isolated copy of the same retained evidence and reader method inputs Accepted without a legacy corpus or local run-state directory.
Explicit Pond; default all-generated population Both blocked on Pond's missing or mismatched retained-result metadata; no silent subset selection.
Explicit legacy review path Rejected as not a main-review path.
Missing copied retained result Blocked on the absent file.
Changed retained bytes with unchanged public digest Blocked on retained-result SHA-256 mismatch.
Public-review drift after initial load Final recheck rejected the changed input; current conclusions must be withheld.

Two additional temporary copies retained the original canonical records but introduced profile contradictions. Their public result hashes were updated so that the semantic check, rather than a digest error, was exercised:

Synthetic change Structural reader Audit diagnosis and correction route
Replace the known storage set with only files Accepted Classification defect: omits SQLite parts in OBJ-1, OBJ-2 and OBJ-4 while retaining file-backed OBJ-3. The existing complete-set rule covers this; send the omission to a new producer run.
Upgrade aggregate read-back basis from afforded to wired, claiming wired task-model pull Accepted Classification defect: contradicts RTE-11 and ABS-1. Wired push in RTE-8/RTE-13 cannot upgrade the pull route. The existing weakest-basis rule covers this; send the contradiction to a new producer run.

These are synthetic diagnostic cases, not revised claims about Maka. Their defect dispositions were assigned by reading support, not detected by code. Both route to source inspection under the existing method; neither justifies editing a production profile in place.

The temporary exercise report is /tmp/commonplace-c10-yweicqwg/checks.json. The observations needed for acceptance are recorded here; that temporary file is not a durable evidence dependency. Real review/result bytes and all recorded method hashes matched at the final check.

Acceptance and remaining scope

The changed instruction, this record and the queue passed commonplace-validate with no failures or warnings; git diff --check passed. No executable implementation changed, so this Markdown migration exercised the reader directly rather than adding tests or repeating pytest. The preceding committed pilot passed all 723 tests.

C10 is done for its active procedure and bounded acceptance. The wider corpus has not been audited: all-generated selection still blocks on Pond. Quote grounding is the next independent procedure, C11; mandatory legacy publication remains for C12/C14 to resolve.

The subsequent oh-my-pi session-report audit exercises the diagnostic on a second published result and establishes two semantic defects despite structural success. That follow-up preserves this initial Maka trial's evidence boundary and records the required source regeneration separately.

The later producer-check exercise applies the clarified method independently to both inputs. It finds task-horizon and learning-branch accounting gaps in Maka that this initial bounded audit did not establish. Affected learning claims now require regeneration or supported uncertainty; the original structural-reader checks still pass.