Wiki · Research
Follow-on extraction — shared comparison spine
Shelf
Research

Follow-on extraction — shared comparison spine

{{Subject}}

1. Snapshot and claim boundary

FieldValue
Track{{durable/local product · evaluation/optimization · interoperability protocol}}
Subject role{{product/runtime · evaluator · optimizer · normative protocol · reference implementation}}
Repository or normative source{{owner/repository and URL}}
Pinned commit{{full 40-character SHA and tree permalink}}
Version represented{{release/specification version if established; otherwise “commit only”}}
Date read{{YYYY-MM-DD}}
Licence{{SPDX id or “not established”, plus pinned licence-file link}}
Read boundary{{directories, packages, specifications, tests, and commands inspected}}
Explicitly excluded{{hosted/current behaviour, packages, transports, integrations, or claims not inspected}}
Measured scale{{result — command, date, checkout/ref, and path scope}}

Snapshot each additional repository or normative authority in a separate table. Never use one repository’s pin to imply that a second moving source was frozen.

2. Question and claim register

State the work item’s selection question, the end-to-end mechanism being traced, and the claim that would falsify the selection hypothesis.

ClaimLabelPinned or reproducible evidenceScope and limit
{{one material claim}}{{Normative / Implemented / Executed / Measured / Documented / Opinion}}{{permalink or command}}{{what this does and does not establish}}

Use the register for the extraction’s load-bearing claims, not every sentence. Prose still keeps its evidence adjacent.

3. State and identity

Trace identifiers, ownership, mutable state, persistence, and resume semantics. Keep these layers separate even when the subject calls several of them “memory” or “session”:

LayerOwner and identityWrite/read pathDurability and resumeEvidence boundary
Conversation or request history
Recovery/checkpoint state
Documentary intent or task description
Repository/workspace state
Agent-managed long-term memory

Use “not applicable” with a reason where a layer is outside the subject. “Not found” additionally requires a bounded absence search.

4. External interaction boundary

Follow one external interaction from declaration or negotiation through validation, execution, result/error return, cancellation, retry, and recovery. Name process, filesystem, network, credential, and trust boundaries that the source actually establishes; label host/application policy separately.

5. Human authority

Trace interruption, steering, approval, rejection, editing, and attribution. Distinguish durable pending state from authenticated authority, authorization policy, user interface, and retained decision evidence. If the subject supplies only some of these, name the missing owner.

6. Durable evidence and recovery

Inventory only artifacts that can reconstruct or assess a run: events, trajectories, checkpoints, logs, task inputs, tool results, artifacts, scores, aggregates, and provenance. For each, state:

  • when it is written and what becomes durable together;
  • identity, ordering, retention, and replay/deduplication semantics;
  • whether it proves execution history, recovery state, evaluation evidence, or accountability; and
  • what an interrupted outside-world effect leaves ambiguous.

An event stream is not automatically an audit log, and a checkpoint is not documentary intent.

7. Operating cost and limits

Record applicable model/tool calls, retries, concurrency, storage, executor/sandbox requirements, external services, and explicit termination budgets. Performance claims require an Executed or Measured record with inputs and environment. A passing test is not a benchmark.

8. Strongest counter-evidence

Present the strongest pinned fact, executable result, or bounded absence that weakens the work item’s selection hypothesis. State whether it narrows the claim, contradicts it, shows an optional path was mistaken for a default, or demonstrates that the subject is not commensurable with its track.

9. Catalogue consequence

Do not edit the catalogue in a subject extraction.

Pattern idVerdictEvidenceReason and claim boundary
{{existing id or candidate: proposed-id}}{{take / take-as-warning / leave / not commensurable}}{{pinned link or labelled opinion}}{{what WI-028 should adjudicate}}

End with the strongest warning or rejected analogy, not only practices that agree with rungs.