Analyse-agentic-system amendments

Current purpose

Evolve the analysis method to chart the design space of agentic systems and find solutions useful to Commonplace at reasonable cost. A public database is a secondary goal. The operator supplied this direction on 2026-10-05; the durable design brief holds the goals and evaluation priorities.

The workshop opened on 2026-10-02 to investigate shared record ID errors. Those investigations and completed repair plans are now historical evidence, not the current execution plan. The earlier construction workshop records the method's initial development.

Read first

  1. Analysis design brief — operator intent; read before design or implementation work.
  2. Versioned corrections candidate — the design reasoning for correcting report text instead of keeping amendment overlays. Its first part is implemented; its publication step and its evidence questions are not settled.

  3. Live method and method maintenance — current execution contracts and the route for authorized method edits.

  4. Separate analysis types from instructions — operator-requested plan for content contracts, intermediate results and verifier inputs. Implemented on 2026-10-06 as far as the method itself goes; see the state entry below for what remains.

Read history only when a design question needs its observed cases or an implementation audit needs its original acceptance criteria.

State and open choices

As of the 2026-10-05 direction:

  • The live method remains authoritative. Since ADR 108 it corrects reports in place: verifier blockers go to the declaring analyst. No accepted result or retained set changed.
  • The open design question is now empirical: whether report correction reduces total work and correction failures compared with the stopped amendment-overlay runs. Round budgets, declined requests and declared dependencies between findings remain open to evidence.
  • The report-correction implementation plan scopes a first change from that candidate: verifier findings go to the declaring analyst, for all three reports, and reconciliation is narrowed to connecting reports. The operator accepted it on 2026-10-05 and it is implemented under ADR 108; its result section lists what was built. The October 6 Dynamic Cheatsheet runs exercised it; the Sol/Luna audit records correctness differences, model identities and elapsed runtimes. Those observations do not establish a causal method improvement. The implementation does not include the publication policy.
  • The type/instruction separation is implemented (commits 0e60d18d4 through ea942625e). Every accepted output now has a type: the boundary, the three analyst reports, the reconciliation, the three verifications and the synthesis, each written whole by its worker and checked at acceptance against the run's identity. The record contract states the correction answers' form; the record verifier loads the boundary contract it judges against. The run-state and overview types hold checkable properties only. The job instructions (reconcile, verify, runtime, memory, epistemic) are written as situation, mission and boundaries, and no longer restate criteria the loaded contracts carry. Not done: the three shared contracts still sit under instructions/ (layout only); a few cross-references to other jobs remain in contracts; and the plan's shared-contract review boundary was examined only against the gate/freshness review system, which the method does not use, so it stands as a decision for the operator if retained sets are ever reviewed through that system.
  • The analysts are scheduled from one table (ANALYST_SPECS) with <member>-<n> job names; a correcting analyst reads its previous report and a code-cut request packet, not the other reports.
  • The publication policy is implemented (ADR 109): each verification now classifies findings as blockers or limits, limits travel into the overview's Limitations with their records, and a limit on a profile value is admissible only where the value expresses the uncertainty. Its proposal is archived. No run has exercised it; the stopped Sol run's profile stage is the first case it would have changed.
  • The typed-output decision is recorded (ADR 110). The mission-form sweep covers all job instructions since 2026-10-06.
  • The withheld-instruction review ran twice on Luna: 14 of 17 operative criteria were recovered or partly recovered by at least one review, none was found to live only in an instruction, and the second review raised a defensible classification defect the run's verifier had passed.
  • The classification revision (ADR 107, method b95a2bb79) was tested by the four runs listed in the report-correction plan. In the complete Dynamic Cheatsheet run, write_agency and lineage kept supported findings beside named unresolved parts, as intended. The two stopped runs failed on correction propagation, which this workshop owns. Open item: the cause of the blocked run AAS-2026-10-05-dynamic-cheatsheet-56ba8ab235b8-01 has not been examined.
  • The analysis collection split owns its separate design and implementation plan. Its workshop framing is historical relative to the shipped collection; inspect the live method before inferring what remains unimplemented.
  • The older Sol follow-up plan stays visible because its full disposition has not been established. Its status section distinguishes delivered or superseded interfaces from items needing review. It is not a current implementation commission.

The accepted second Dynamic Cheatsheet run remains unchanged. Its duplicate identity and overstated initial-inspection findings were accepted limitations; reopen their repair if the same errors recur, not merely because the audit exists. Earlier deferred audit items remain deferred.

Authority and next decision

This workshop supports investigation and design. Neither its framing, the design brief nor the correction candidate authorizes method implementation, a new model run, recovery of a stopped run or edits to retained analyses. Keep separately owned work separate. Existing retained sets are frozen comparison evidence; do not revise an open run's pinned method to resume it.

The next decision is whether to authorize the trial the implementation plan describes. The candidate's evidence questions remain questions, not an authorized trial plan. Return a bounded proposal for operator decision before a further method change or a live trial.

Closure

Close when the current design question has a recorded decision, or the operator chooses to stop. Promote any adopted design through the live method and affected consumers; retain an unadopted design only if it earns a library proposal. Resolve or explicitly transfer the Sol plan's remaining obligations.

Extract independently useful audit evidence into durable retained reports when another consumer needs it; leave change history in git. Then delete this workshop, including its local history, and remove its active-list entry. The design brief already has a durable home and must not be deleted with the workshop.

Complete file listing (generated at build time)