Analysis collection implementation plan

Commission

Written on 2026-10-03 at the operator's request, for the executor the operator launches on it. The launch is the authority to implement; this file alone starts nothing. The two design files authorize nothing by themselves.

It authorizes the profile replay of phase 2. It does not authorize an analysis run, regeneration of any system, recovery of a stopped run, or changes to the Sol run follow-up plan, which the operator revises after this work.

Intent

Analysis workers fail parts of their jobs. The operator's response is to simplify each job so that an analyst holds only the analysis. This work builds the structure for that. The operator will then regenerate every analysis under the simplified method, once.

Two changes serve the intent:

1. A self-contained analysis collection. Analysts read a small contract that addresses analysis only, supplied to them by path. Publication, comparison and method authoring cannot add rules to analyst input later.
2. A separate profile job. The memory analyst traces and records. A later job classifies the accepted records into the memory comparison profile.

When a choice is not settled below or in the design files, choose what leaves the analyst jobs with fewer concerns to hold, fewer references to resolve and fewer instructions to reconcile. Do not buy that by dropping a rule a worker needs, weakening source grounding, or removing independent review.

End state

The work is done when all of this holds:

• kb/agentic-system-analyses/ exists and is self-contained. It owns its contract, the analysis method, the member and run-state types, run state, retained/ and retained-archive/. A run needs no method file, type or contract from another collection.
• Every job packet supplies the collection contract as a read-first path, and its bytes participate in the job's dependencies.
• No analyst packet contains a rule about publication, comparison writing or method authoring, except those listed with a reason in the result record.
• Publication places the accepted set at retained/<system-slug>/, moves a superseded set unchanged to retained-archive/<run-id>/, and writes no second review. The list of current analyses is generated at site build. Comparison tools and the site build enumerate current sets through one function.
• The memory job no longer produces the comparison profile. A profile job and a verify-profile job produce and check a memory-profile.md member after record verification closes and before synthesis.
• Workers can apply every registered term their contracts use without opening a file the packet does not supply, and the worker rules say so.
• The old kb/agentic-systems/reports/ tree is byte for byte unchanged and excluded from validation and from the comparison population.
• Required tests and validation pass. Two decision records exist. A result record states what was done, measured and left.

The first regenerated analysis is the outcome check. It is not part of this work and needs its own commission.

Design authority

The design is settled in these files. Read them before acting. Where they and this plan disagree, this plan governs; report the disagreement.

• Collection design: ownership, publication, what moves, the contract's content, registered vocabulary, consumers and the Decisions section.
• Profile job design: position in the run, inputs, output, verification, what existing jobs lose, and the replay.
• Decision draft: source for the first decision record, with two TODO revisit conditions.
• Drafts of the analysis contract, the remaining contract of the old collection and method maintenance. These are starting points. Revise them freely toward the end state.
• Complexity measurements: the five measures, their counting rules and the baseline per role.

Everything under evidence/ records how the decisions were reached: the measurements, the rejected shortening alternative and the earlier reasoning. It is not design. Use it for baselines and for the decision records' context. Do not take design from the analyse-agentic-system-amendments workshop; its audits are background on observed failures.

The live code and method are evidence for how things work now. Inspect them before relying on any path, count or interface named in the workshop files; those were found by search on 2026-10-03 and may be stale.

Boundaries

These bind every route.

• Nothing is migrated. No retained set, archive or run state moves into the new collection. Do not edit, move or delete anything under kb/agentic-systems/reports/, except adding a validation exclusion.
• No run crosses the change. Check for open runs before editing. A run opened under the old method finishes there or stays as stopped evidence. It is never resumed under the new method.
• Analytical content is unchanged. Definitions, record grammar, evidence rules, acceptance gates and the profile's axes, values and per-value evidence stay as they are. This work moves and separates; it does not redefine. The profile job reads accepted records and may read the frozen source, but declares no records and adds no evidence.
• Independent review stays. Do not remove reconciliation, record verification or synthesis verification, and do not fold verify-profile into code validation.
• No worker exemption. Do not tell workers to skip a root rule. The one stated deviation is the vocabulary rule, and only after the sufficiency check below shows the contracts carry what workers need.
• Hand-authored reviews stay in kb/agentic-systems/reviews/. A review is retired only when its system is regenerated, which is outside this work.
• Repository rules apply: explicit staging, relocation-command results committed alone, no compatibility shim without a consumer, no commonplace-init in this checkout, bare commonplace-* commands.
• One owner at a time. Other sessions edit this repository. Check status before each commit and stage only this work's files.

Priorities and partial results

The two phases are ordered because the profile type and job instructions should be written once, in the new collection. Phase 1 is useful without phase 2. Phase 2 is not started on a failing phase 1.

Within phase 1, the order of value is: the collection with its contract supplied to packets; then publication into stable paths with the generated list; then the vocabulary sufficiency check; then the opening disposition.

If work stops early, leave the repository in a state where required checks pass and no committed state mixes old and new paths for one consumer. Record what is complete, what remains, and anything learned that changes the design. A partial result with that record is acceptable. An unreported gap is not.

A negative finding is a result. If a check below fails, report it with its evidence; do not adjust the check to pass.

Supported route

This is one workable route. Take another if it reaches the end state inside the boundaries, and say what you changed. Steps marked fixed have a binding reason.

Phase 1: collection, contract and publication

1. Fixed: inventory before mutation. Many consumers read these paths and type identities: workflow, set, analysis, publication, validation and matrix code, tests, scripts, the skill projections, site configuration, command and validation reference pages, landscape synthesis, taxonomy refresh and the transfer scan. Follow change a contract that several consumers read and keep the inventory with its dispositions. A missed consumer fails silently later.
2. Relocate the analysis method and the member types. Fixed: commit the relocation command's result alone, so file history survives the rename.
3. Promote the collection contract and supply it in every job packet. Make retained/ under the old tree excluded from validation, and retire the one generated review that pins an old set.
4. Change publication to the stable path per system with the archive move, and generate the list of current analyses at site build. Write the publication instruction; only the coordinator loads it.
5. Vocabulary: for each registered term the supplied contracts use technically, check that the contract states what a worker needs to apply it. Fill gaps inline in a few lines. Then state in the worker rules that the supplied contracts are the binding copy and the linked definition notes are background, and list those notes in the maintenance instruction as dependencies to recheck.
6. Establish what the boundary job uses from opening. Pass only that, or reword it as run metadata. Remove the memory job's mention of publication checks if nothing analytical depends on it.
7. Re-measure, verify, write the first decision record and the result record.

Return to the operator after phase 1 with the result record before starting phase 2.

Phase 2: profile job

1. Check the design against ADR 083 and ADR 093 clause by clause. Report any conflict before building.
2. Draft the profile type, the profile job instruction and the verify-profile instruction in the new collection.
3. Fixed: replay before changing the run. Rebuild the profile for the two retained Dynamic Cheatsheet sets from their members with the profile removed, and compare axis by axis as the profile job design describes. The operator authorized the replay on 2026-10-03 and directed that its model jobs run on Luna. It gates the build because a classifier that cannot recover the retained profile from records would make every regenerated analysis weaker.
4. Add the two jobs to the run, the sixth member to the set, and move the profile checks out of the memory, reconcile and record-verification jobs. Point the matrix reader and set validation at the new member.
5. Re-measure the memory, reconcile and verify packets and the two new ones. Verify, write the second decision record and complete the result record.

Tests that start model jobs

The replay is the only planned test before regeneration that starts model jobs. All other checks in this plan are deterministic. Run the replay's worker jobs on Luna, by the operator's direction, and record the exact model ID in the result record. If another test that needs model jobs proves necessary, run it on Luna too and say why it was needed. If Luna cannot be selected for a worker, stop and ask; do not substitute another model, because the replay result is compared across the two retained sets and must come from one model.

Confirm the model from each worker's own session record, not from the request. The configured Pi worker omits model and inherits the dispatching session's model, and this KB has a recorded case of workers requested as Luna that ran as another model: harness sub-agent model selection regression.

Left to the executor

Decide these from the code and the intent, and record the choice: file and function names; the slug function, if the existing run-ID slug is not usable as it stands; the form of the generated list page; how the validation exclusion is expressed; test structure; commit granularity beyond the fixed points; the wording of contracts and instructions; whether a draft is revised or rewritten; the order of steps that do not depend on each other.

Return to the operator

Stop and ask, with the evidence and your recommendation, when:

• the end state cannot be reached inside a boundary;
• a consumer needs old and new paths at once, or a compatibility shim;
• the contract cannot stay near 3 KB without omitting a rule a worker needs;
• a registered term cannot be made sufficient inline in a few lines;
• an analyst packet must keep a publication or comparison rule;
• the replay weakens many axes, or shows the records lack facts the profile needs;
• the design conflicts with an accepted ADR other than ADR 099;
• the work would change an analytical definition, add a workflow stage beyond the two profile jobs, or touch files another session is editing.

Otherwise proceed without asking.

Verification

• uv run pytest -q and uv run ruff check . pass.
• Targeted commonplace-validate passes for the new collection, the changed contracts and instructions, and the redirect map. The old reports/ tree is skipped and unchanged.
• Site boundary tests show that run state is not published and that retained members are.
• Publication tests cover a first publication, a replacement with the archive move, an interruption on each side of the move, and refusal of two sets of one source or a directory whose name does not match its set.
• A test moves a real retained set to an archive path at the same depth and confirms its relative links resolve and its hashes are unchanged. The design asserts this without having tested it.
• Packet tests show, for every job, the contract supplied by absolute path and included in dependencies, also through scripts/analyst_trial.py.
• Re-measurement: per analyst role, bytes of mandatory method input and the five complexity measures, against the recorded baseline, with the itemized list behind each count. If concerns, references or conflicts did not fall for analyst roles, say so plainly; that is the claim this work rests on.

A passing fixture shows interface behavior. It does not show that workers follow the method or that analyses improve.

Records

• Result record: result.md in this workshop, written as work proceeds: what was implemented, each executor choice with its reason, the consumer inventory, measurements, deviations from the supported route, and what remains.
• Decision records: two ADRs, numbered when written, after implementation. The first revises ADR 099 and is built from the decision draft; keep its literal TODO revisit markers. The second records the profile job. Each names its consumption paths and considered alternatives, as the ADR type requires.
• Commits: follow the repository's commit rules, with the Workshop:, Decision: and Model: trailers where they apply.