Provenance and limits

Commission and inputs

The Commonplace operator commissioned the conceptual investigation on 2026-10-31, then directed that its practical follow-up be a watch/checklist: correct the next few cases and use them to formulate a more substantial hypothesis. The latter direction controls the current README. It does not assert that no confirming literature exists.

The initial research files were supplied in commonplace-epistemic-boundaries-workshop.zip, SHA-256 a26ebbfcc2f55056c96cb36bb153a2e7f290045917c454e7bb991bd6c11ed1f0. That is the identity of the generated workshop bundle, not a primary-source snapshot or experiment hash. The evidence register and initial framing are retained as research background; the findings and experimental proposals now make the observational priority explicit. This PR adds the checklist and case log. No new external literature review was performed while preparing the PR.

On 2026-11-01 the operator supplied a second, externally generated literature review (ChatGPT). It is retained verbatim in the external review file. Its seven new sources were first verified by landing page only and entered as E14 to E20. Later the same day the operator authorized ingesting them: five new ingests were created through cp-skill-ingest with a fresh drafting worker each, and the register entries were rewritten from those ingests. E19 (Williams, Lombrozo and Rehder 2013) has no open-access text and stays at abstract level. Its proposed sharper formulation is recorded as a candidate for the review point, not adopted.

Repository snapshot and conventions

Repository: https://github.com/zby/commonplace/

Baseline: e53fc1bb1041d54de99beda36302b6ab889ad536 (main, checked again through the GitHub connector when preparing the PR), commit timestamp 2026-10-31T08:38:56Z. The workshop and source contracts read during the initial investigation remain at that baseline. Plain Markdown is permitted for workshops; the Active Workshops index receives a single new entry. No operationalization, theory, source-ingest, or shipped-instruction files are changed.

Initial investigation reads included AGENTS.md, kb/work/COLLECTION.md, kb/sources/COLLECTION.md, the work index, the operationalization and structural-coherence workshop READMEs, all six starting theory files in the commission, and lightweight-search-control-does-not-license-adoption.md. Full existing ingests inspected were the semantic-leakage and rationale-faithfulness ingests; search confirmed the GSM-DC ingest. Those searches were not an exhaustive duplicate audit.

Research provenance and scope

The initial research record states that experimental framing preceded the targeted literature search. Search directions covered conceptual integration, relevance/distractors, goal conditioning, preference incentives, rigidity/correction, and separation. The evidence register records inspected versions, reading depth, primary links, evidence classes, and ingestion suitability.

The review was not systematic: there is no exhaustive screening count, registered search protocol, or prevalence estimate. Source cards distinguish inspected versions from uninspected revisions. No new immutable primary-source snapshots or ingest reports were created. Public availability was not treated as a license to redistribute full texts.

Episode provenance

The initial investigation relied primarily on the operator's retrospective handoff. The later O001 primary-evidence record improves that basis with selected exact conversation excerpts and immutable historical Git files from the operationalization workshop at commits 01bab9d (initial draft), 8f034e4 (first repair still including generality), 56b00de (executor-relative repair still listing generality), and 961eb45 (generality explicitly outside task-relative operationalization). The revised operationalization README corroborates the eventual distinction, not the sequence by itself.

The new record is manually excerpted from the conversation available when this change was made; it is not a full original chat export or a harness-authenticated raw transcript. The Git versions are independently retrievable but represent authored workshop artifacts, not the assistant's hidden deliberation. Retrospective model explanations remain hypotheses, not privileged access to earlier computation. All of the corrections belong to one episode, not independent occurrences. The candidate source passage is kept separate from observed run-level recurrence in the case log.

Execution and validation boundary

This change is an observational workshop, not an experiment or automated watch. No benchmark, pilot, training run, scheduler, new infrastructure, or recurring notification was executed or installed. No new source-capture checksum was fabricated. Proposed source or theory repairs remain proposals unless independently authorized within a later task.

When the workshop was prepared, the sandbox could not resolve github.com for a Git clone and commonplace-validate was not installed, so GitHub reads and writes used the connector and local checks covered only the authored Markdown, local links, file identities, and the scoped index edit. That limit no longer holds: on 2026-11-01 commonplace-validate ran on the workshop, the edited ingests, and the edited theory note from a local checkout. The Q001 repairs of 2026-11-01, the operator-requested re-ingest of the GSM-DC source the same day, the follow-up qualifications of the arithmetic-execution clause in two theory notes, and the five operator-requested source ingests for E15 to E18 and E20 are the only edits this workshop has made outside its own directory, each under explicit operator authorization. The ingests added no links into other KB notes; the link candidates their connect passes produced are recorded in the git-ignored connect cache, not acted on.