Agentic analysis output documents: closure, 2026-10-02

Type: types/note.md

The operator commissioned closure of the output-documents workshop on 2026-10-02. The replacement of the monolithic result is implemented and has bounded producer and consumer acceptance. This report retains the evidence needed to assess that conclusion after the workshop is deleted. It supports future review of the output design; it does not establish improved analysis quality or full-corpus refresh completion.

Adopted design

The analysis set contract owns directory membership and shared consistency checks. Each member has its own type. The current complete set contains overview, runtime, memory, epistemic and reconciliation reports. Blocked and out-of-scope sets contain only the overview. Reviews and run state pin ARTIFACT.yaml; it pins the reports. Publication and comparison readers use the shared artifact checker. ADR 095 records the directory model and the separation of shared checks from checks requiring a frozen source or local run state.

The original candidate had four reports. The fifth separates reconciliation from synthesis, implemented in 0ad34f5d. The later collection ownership migration is documented in its retained acceptance record. The retired result type and compatibility readers were removed. Historical results remain frozen in the collection archive; they are not current comparison inputs.

Types specify document contents; instructions specify production and judging. The workshop's cleanup reduced the five governing files from 15,701 to 12,701 whitespace-separated words before the subsequent independent review. The review brought the total to 12,795. These are dated measurements of that method, not word counts of today's reorganized job instructions.

Retained trials

Four exact records are retained under captures/. Their original locations, source commits and SHA-256 values are in capture provenance. Their relative links belong to their original checkout context and are preserved as capture data, not live navigation. The validation-ignore marker applies only to these captures.

  • Cleanup regression: three frozen source analyses validated, published and passed independent handoff and bounded consumers. It found omitted profile values despite surviving prose and changes in functional scope. Structural success did not establish semantic completeness or comparable breadth.
  • Fixture split: two existing results split with zero duplicate declarations, zero unresolved references and unchanged quote unions. The dropped memory sections survived verbatim in the memory reports. Measured set sizes were 19,505 and 15,100 words, respectively 87% and 84% of the original retained result plus local report. The largest members were 67% and 51% of the old result. Smaller projections after reducing amendment copies were estimates, not another measured trial.
  • Fresh batch acceptance: three whole-system sets published, passed independent handoff and full-set validation, and supplied the same selected population to matrix, table and statistics. The memory member remained largest even for Agent-S; this trial did not demonstrate a runtime-dominated output. Body totals were 16,622, 14,098 and 17,979 words. Specialist correction turns were 1, 2 and 2.
  • Batch trace audit: confirmed zero failed analysis validations or prepares, while retaining recovered command errors, truncated reads and incomplete instruction delivery. Final passes do not establish error-free execution.

The last two records came from branch refresh-batch-01-rerun at c854800b; its two batch commits remain absent from main at closure. Retaining the records neither merges their sets nor advances the main inventory. The corpus-refresh workshop owns that integration decision.

Closure checks

The inspected method and evidence commit is 1bdd2929650bad5a5ecb9d0fb6b2f23b6f87c00a. Closure checks used the committed review and retained set for AAS-2026-10-01-instinctual-memory-07, excluding concurrent uncommitted review replacements. The review pins manifest SHA-256 e409b30ce075f68a71fbdf29719dd9d2816e639d687a8d0bfc0fc7a164bade8e.

The selected existing test modules passed: 211 tests. They cover directory membership and hashes, set identity, cross-member references, analysis publication, comparison readers, member contracts and site publication. Missing members, changed bytes, differing run/source identities and unresolved references have rejection cases. No new tests or analysis runs were added. Explicit full validation of the retained five-report set passed without warnings.

A temporary export of the committed kb/ and scripts/ trees supplied the same selected review to build_systems_matrix.py, render_systems_table.py and analyze_matrix.py. All exited zero, producing one code-grounded row and no doc-grounded rows. The profile had ten known axes; all listed value bases were wired. This is a one-row reader trial, not a corpus distribution.

The recorded reader trial produced this bounded numerical summary: write_agency retains both automatic and manual values, and read_back_direction retains both pull and push values. Each is counted once within one selected row, rather than as another system. The former cites MEM-RTE-1, RTE-3, RTE-4, RTE-5, RTE-8 and RTE-9; the latter cites RTE-1, RTE-2, RTE-5 and RTE-8. These are profile classifications, not measured behavioral benefit. The evidence-bearing set and review were later retired during analysis-offload closure; this compressed reader observation remains, and Git retains the original records. The comparison procedure and landscape synthesis skill both require manifest-pinned evidence of this form. The retired trial set is no longer selected as a live input.

The transfer scan checks complete run state, complete substantive disposition, manifest identity and hashes before and after reading. Its completion and source assumptions were inspected; closure did not commission a new Commonplace transfer scan.

Limits and disposition

The adopted contract checks declared membership, hashes, identities, declarations and references. It does not prove semantic support, complete profile classification or absence of duplicated prose. Exact derivation of the finalized specialist body and a set-wide quote minimum were deliberately dropped during implementation. The original proposal's stronger checks are not implied by this closure.

The fixtures support preserved evidence and reduced duplicated text on two historical cases. Fresh trials establish operation of the split output and its consumers. There is no matched estimate of quality improvement, time savings or general context benefit. Coordination still requires semantic reconciliation and corrections; the trace audit bounds the acceptance claim.

The output-documents workshop is closed. Its draft contracts and superseded handoffs are deleted; Git retains their history. Corpus regeneration, historical consumer migrations and ongoing analysis-method repair remain with their existing workshops. The unmerged batch is not a blocker for the implemented output contract. No workshop redirect is added.