Batch 21 ownership decisions: operator recommendations

Decision and authority

Keep all eight original runs held. Prefer bounded ownership reconciliation for ACE and Agent Skills; propose separately authorized fresh replacement lineages for the six filesystem incidents. These are recommendations, not releases, execution plans or integration approvals. A replacement lineage means a new, separately identified production history and independently audited artifact. It does not mean resetting the old run or declaring its actions harmless.

The purpose is to decide whether more retained evidence can answer the actual ownership question, or whether new work should be considered without pretending to settle history. The parent owns scope, scheduling and integration decisions. Only this file is authored. No delegation, acquisition, installation, target execution, worker-prompt/log inspection, owner/mailbox contact, reservation release, cleanup, source/method/config/state mutation, reset, advance, judge, integration or commit is authorized by this document or performed for it.

Basis: the complete batch plan and machine bindings, plus the retained findings linked below. The machine record fixes exact source identities, old run/worktree paths, attempt snapshots and publication maps. This is not a fresh semantic audit or an expanded historical compliance audit. The one native attempt inspected was ACE's open reconciliation metadata; no pinned prompt or worker trace was opened.

What evidence can and cannot decide

A bounded read-only reconciliation can compare retained assignment/release records, native attempt metadata and existing audit decisions. It can establish what is recorded and identify contradictions. It cannot grant ownership, terminate a worker, release a reservation or prove that no writer can return. Process absence is not evidence of release. No current owner identity or release is invented where the records do not establish one.

For the six filesystem incidents, retained audits already establish the actions and their connection to the old publication. They do not retain a sufficient pre-write image of the affected paths. Reading those reports again, checking current absence, or producing correct new content cannot recover overwritten bytes or prove prior ownership. Those facts are irrecoverable from this retained evidence, not necessarily from every possible external backup. An already retained, independently attributable pre-incident snapshot could change the answer under a separately authorized, bounded evidence review. No such snapshot is established here. Do not expand into worker prompts/traces to search for one.

Two different decisions must stay separate:

  • Judging the old publication: preserve the established unauthorized acts, unresolved collateral effects and any historical validation/grounding limits. Semantic quality and exact-byte binding do not establish historical compliance. Unknown loss is not an acceptable premise for clearing the old hold. Nothing here recommends accepting it.
  • Creating a new artifact: a separate commission could authorize a new owned, isolated workspace and independently audited production history. It would need an explicit source identity/boundary, supported method, authorized write scope, normal verification, exact-byte binding and a fresh independent audit covering content and operational scope. A future claim of historical compliance must remain limited to evidence actually audited. New content does not certify old behavior, release old owners or resolve the original incident.

For the six proposed replacements, prefer a fresh source-first lineage rather than copying accepted old members and calling the result independently produced. The retained reports remain diagnostic evidence for the parent, not analytical answers to pass to fresh source-first authors. This costs more than narrow re-verification, but gives a genuinely separate artifact. It is a request for a future authority decision, not permission to prepare or execute that route now. Any future workspace ownership must be established independently; the old worktree and foreign paths remain untouched and held.

Case One recommended operator action Can bounded retained-evidence reading settle the stop?
ACE Leave held; prefer a bounded ownership/status reconciliation before choosing recovery or replacement. It can confirm the open attempt, not establish release from the records inspected.
Agent Skills for Context Engineering Leave reserved and held; prefer an owner-coordinated decision separating semantic repair from historical compliance. It can distinguish the issues, not release the reservation or retroactively validate the old worker.
AI-Context-OS Consider a separately authorized fresh replacement lineage; retain the old incident hold. No: the retained write/move/cleanup results do not reconstruct prior file or ancestor state.
NAMS Consider a separately authorized fresh replacement lineage; retain the old incident hold. No: the empty write and removal do not prove pre-write absence or ownership.
LLM-Wiki-v3 Consider a separately authorized fresh replacement lineage; retain the coordinator's old incident hold. No: repaired profile content does not recover foreign prior state.
llmwiki-marimo Consider a separately authorized fresh replacement lineage; retain the old incident hold. No: the foreign answers file's prior contents are not retained.
Siftly Consider a separately authorized fresh replacement lineage; retain the old metadata incident hold. No: present absence cannot show whether run-root metadata was overwritten.
Sift-KG Consider a separately authorized fresh replacement lineage; retain the old incident hold. No: successful directory removal does not establish the overwritten file's prior contents.

1. ACE — reconcile the unfinished assignment, not process presence

Recommendation: leave held and prefer bounded ownership/status reconciliation. There is no old publication to judge. Substantial accepted reports survive, so resolving the existing assignment could avoid rebuilding useful work. The consequential question is whether the prior worker still has writing authority and what supported recovery, if any, the responsible owner permits.

The native metadata at .commonplace/worktrees/ace-a22ddd59ed57/kb/agentic-system-analyses/state/AAS-2026-10-12-ace-a22ddd59ed57-01/state/attempts/000019-reconciliation.json records state: open, pinned report inputs and no returned outputs. It does not say whether the worker completed elsewhere, failed or never ran. The coordinator API failure is not a terminal worker record. No release is established by the batch-03 recovery boundary or the retry decision.

A bounded read can identify an existing attributable terminal/release decision if one is supplied later. Without it, the missing worker outcome and continuing write authority cannot be inferred from metadata or process absence. This is a missing status/authority record, not established collateral loss.

Alternative/tradeoff: the parent could seek separate authorization for a new, independently owned replacement lineage while preserving the open predecessor. That duplicates substantial analysis and still does not release the old worker. Prefer reconciliation first; if its evidence is unavailable, leave the old run held rather than fabricate completion or silently take over its workspace.

2. Agent Skills for Context Engineering — three separate questions

Recommendation: leave reserved and held; prefer an owner-coordinated decision that treats repair, grounding and historical validation separately. Contact or coordination is not authorized in this preparation. The checklist's exact row preserves reservation to the reported owner; neither this document nor a later publication identifies a new authority to release it. The audit worker named in the inventory is not evidence that the producing reservation was transferred to that auditor.

The exact-publication audit establishes a current content defect: the claim makes a preference against publishing unresolved conflicts into a prohibition. The path-confusion retrospective separately retains incomplete independent grounding of inherited routes and the old worker's altered artifact-path validation failure. The maintenance decision resolved the suspected shared producer defect, not this worker's obligation or reservation.

Read-only reconciliation can distinguish these findings and bind the later publication; it cannot correct the claim, release the owner or supply a missing historical pass. The retained historical worker record contains no successful supplied-path check. A new check would establish only a new result. Missing underlying source bodies could support a new source assessment under separate authority; they would not prove what the historical worker read or performed. No broader historical audit is proposed here.

Alternative/tradeoff: after explicit ownership coordination, a separately authorized fresh lineage could avoid inheriting partially grounded claims. It costs a new analysis and independent audit instead of a narrow correction and verification. Prefer the owner route because a concrete, bounded semantic defect is already identified. If historical compliance cannot be certified, retain that limit explicitly; do not rewrite the old history as compliant. New independently audited content could be judged on its own evidence, not as retroactive repair.

3. AI-Context-OS — replacement does not undo a move or cleanup

Recommendation: propose a separately authorized fresh replacement lineage, with the original publication and incident still held. The old provider failure is not the current stop. The synthesis worker wrote outside its worktree, moved that output to the correct destination, and attempted ancestor cleanup. The operational diagnosis, case 4 binds the moved bytes to the published synthesis and identifies the unresolved prior contents of both foreign and correct destinations and ancestor ownership.

Exact old worktree: /media/zby/2b5e98b6-baf0-4e97-90d2-03321ea23372/outputs/commonplace-worktrees/batch-17-ai-context-os. The foreign file appends -4a0e34822f96-01 to that worktree path, then /state/handouts/000038-synthesis/outputs/synthesis.md. Cleanup's nonzero result reports the shared commonplace-worktrees directory was not empty; it does not mean the preceding move and removals failed.

Bounded reading can explain that result but cannot inventory the old filesystem or recover overwritten content. No harmlessness claim follows from content continuity. A fresh lineage is proposed because more content work in this history cannot answer the unresolved effects.

Alternative/tradeoff: leave the target without a new artifact while awaiting an attributable retained pre-incident snapshot and ownership evidence. This avoids duplicated synthesis but may leave the target incomplete indefinitely. Narrow new synthesis plus verification could reassess text, not settle collateral effects; it is not the recommended substitute for a separate production history.

4. NAMS — the violating worker was the verifier

Recommendation: propose a separately authorized fresh replacement lineage; keep the old report-verifier incident held. The operational diagnosis, case 1 identifies 000022-report-verification, not a semantic author defect. It wrote empty content to /home/zby/llm/commonplace/.commonplace/worktrees/nams-605887a7f7a0-01/none, then removed that file and sibling directory. The coordinator hold remains controlling.

An empty successful write could truncate an existing nonempty file. Successful removal does not establish new creation, prior ownership or absence of loss. Those facts are not recoverable from the retained results. The old verifier's no-blocker content judgment and matching publication bytes do not answer them.

Alternative/tradeoff: retain the useful old publication unchanged and held, without commissioning replacement. If only a new semantic opinion were needed, separate report re-verification would be cheaper than rebuilding. It would not clear this authority stop or constitute a fresh independent production lineage. A new artifact is worth considering only if corpus coverage justifies that cost; no acceptance of unknown loss is proposed.

5. LLM-Wiki-v3 — do not revive the superseded positive recommendation

Recommendation: propose a separately authorized fresh replacement lineage; retain the coordinator's old hold. The operational diagnosis, case 2 attributes the incident to 000024-memory-profile. The foreign path was /home/zby/llm/commonplace/.commonplace/worktrees/llm-wiki-v3-3bd68241f463-01/kb/nonexistent. The worker wrote empty content and removed the file, foreign kb and sibling root. Earlier profile content repair is separate and already reflected in the published qualification.

The coordinator decision deliberately holds this case despite the original auditor's conditional positive. Bounded reading settles that precedence and action attribution; it cannot recover the foreign file's prior bytes or confer ownership. Neither a good profile nor an empty directory at removal proves harmlessness.

Alternative/tradeoff: leave the existing artifact held and pursue no new production. A separately commissioned profile reassessment would be narrower and cheaper, but addresses content rather than cross-owner effects. Replacing only profile authorship would still inherit the old run's held history. Prefer a separately identified artifact if new coverage is wanted; preserve the incident rather than republish under the assumption that the old positive now applies.

6. llmwiki-marimo — corrected profile, unresolved foreign answers file

Recommendation: propose a separately authorized fresh replacement lineage; keep the old synthesis incident held. The operational diagnosis, case 3 identifies 000032-synthesis, after the separately repaired profile refusal. The foreign file was /home/zby/llm/commonplace/.commonplace/worktrees/llmwiki-marimo-56b2c84667fe-01/kb/agentic-system-analyses/state/AAS-2026-10-14-llmwiki-marimo-56b2c84667fe-01/state/handouts/000032-synthesis/outputs/answers.md. It was written empty and removed. The retained sequence does not establish ancestor-directory removal; do not import that finding from the other cases.

Bounded reading can distinguish the repaired classification from the later incident, but cannot establish pre-write existence/content, foreign ownership or the remaining directory state at that time. The coordinator hold is not waived by the correct profile or synthesis verification.

Alternative/tradeoff: retain the old useful account held and wait for actual pre-state evidence. A new synthesis and synthesis verification would be cheaper than full independent production, but would not answer the foreign answers-file question or erase the held history. Prefer a separate artifact if the parent wants new coverage rather than another opinion on the unchanged text.

7. Siftly — small identity payload, consequential prior-state gap

Recommendation: propose a separately authorized fresh replacement lineage; retain the original metadata incident hold. The exact follow-up finds useful, publicly qualified content but establishes an unauthorized write and removal by 000008-epistemic at /home/zby/llm/commonplace/.commonplace/worktrees/siftly-71a1181fbc3d/kb/agentic-system-analyses/state/AAS-2026-10-15-siftly-71a1181fbc3d-01/worker-identity.json. Its supplied identity destination was inside its handout, not the run root.

This is not the sibling-directory pattern. The payload's small size and later correct identity write do not show whether pre-existing run metadata was lost. Retained evidence establishes neither target/source mutation nor zero collateral effect. Reading current absence or publication success cannot settle that gap. A new, independently owned lineage would avoid relying on that run's metadata history, without certifying what happened to the old file.

Alternative/tradeoff: preserve the qualified old publication and hold without new analysis. Replacing only epistemic authorship and its ordinary verification would cost less, but stays attached to the old unresolved metadata history. The missed operational acceptance decision belongs to the parent/coordinator; source-facing report verification was not assigned to audit worker traces.

8. Sift-KG — legitimate report and foreign empty file are different evidence

Recommendation: propose a separately authorized fresh replacement lineage; retain the old workflow incident hold. The exact follow-up establishes 000009-workflow wrote empty content to /home/zby/llm/commonplace/.commonplace/worktrees/sift-kg-dca68bc45a2c-01/state/handouts/000009-workflow/outputs/report.md, then removed that file and five containing directories. The later correct report is the installed workflow member; the empty foreign file is not that report.

Successful directory removal bounds what remained at removal time. It does not show that the file was absent before the empty write or recover prior content. Bounded reading can establish those distinctions, not authority over the foreign path or harmlessness. The public content's existing qualifications and no newly established semantic defect do not justify an old-run clearance.

Alternative/tradeoff: keep the old artifact held and defer coverage rather than pay for independent production. A new workflow member and report verification would be a narrower content exercise, not collateral recovery. Prefer the separate lineage if coverage is commissioned; do not treat removal of an apparently empty tree as proof that accepting unknown loss is safe.

Parent handoff and stop conditions

The actionable choice is between retaining these holds without new production and separately commissioning the recommended reconciliation or replacement route. Neither option grants mutation/execution authority here. Partial progress is sufficient: an unresolved owner or unavailable pre-state record means the old hold stays, not that investigation must expand until it manufactures closure.

Any future authorization would need to distinguish new artifact acceptance from old-incident resolution, establish current ownership without process-absence inference, and bind the intended evidence and write scope. Conflicting ownership, new evidence of contamination, or a request to accept unknown collateral loss returns to the parent for decision; none is silently settled here. The original run, publication, audit and incident evidence remain preserved regardless of whether a new artifact is eventually accepted.

Validation

commonplace-validate kb/work/agentic-system-reruns/batch-21-ownership-decisions.md returned VALIDATION SUCCESS: one file analysed, one text file, zero failures and zero warnings. git diff --check also passed. Only the authorized new file appeared in the final status check. Deterministic validation cannot certify these recommendations, historical compliance or ownership release.