# B2 paired critique/revision audit Development tooling with unfrozen defaults, not empirical evidence. The separate [adapter](b2_contract.py) supports `b2-critique-revision/1.0`. It does not run models, judge semantic leakage, repair inputs, deliver critiques, select candidates, or generate task answers. All outputs remain coordinator-only, outside adaptation and artifact-quality packets. Directory names do not enforce confidentiality. ## Explicit inventory Copy [manifest-b2.template.json](manifest-b2.template.json) and fill its null placeholders using already sealed archives. Use the existing `prepare` and `summarize` commands in the [main guide](README.md). Every run in every supplied archive remains assigned, regardless of status, resources, quality or coverage. Mix contracts/scopes using their existing row declarations; do not replace runs. This adapter uses existing manifest **1.1** and its scope-bearing run schema. Only the new B2 `continuation_input` variant adds `pairs`; existing matched, B1, initial-input, packet and human-form schemas remain unchanged. Declare: - `request_slots`: explicitly classify every adaptation **worker** attempt as initial or continuation, using the existing partition convention. Noninitial worker attempts are B2 revisions. Every adaptation **critic** attempt must also be explicitly designated in a pair. The archive's role labels are validated, but stages and pairs are never inferred from text, filenames or request order. - Each `pairs` entry has a run-scoped unique `pair_id`, `critique_attempts`, `revision_attempts`, and `revision_coverage`. An attempt may appear in only one pair/stage. Each nonempty stage has one original request plus its explicitly linked transport retries. Retry parents must remain in the same pair/stage. Failed, incomplete, cancelled and unknown attempts are retained, not filtered. - Each critique designation names `attempt_id`, `response_state`, `response_record_id` and `response_sha256`. `available` requires a hash-bound archived **operational_record** linked to that critic attempt and run. Supply the exact UTF-8 prose response body, separately from provider envelopes/traces; the adapter performs no extraction, stripping or summarization. It cannot be a delivered `artifact`. A captured empty string remains available bytes, not an invented missing response or a verified useful critique. - `not_produced` declares that no response was produced; `unavailable` declares unknown/missing response evidence. Both require null record/hash. These are coordinator declarations, never deductions from request status. A revision cannot link a declared unproduced response. An unavailable linked response leaves exact-response validation unknown, even if the revision contains text. - Each revision designation names its `attempt_id` and `critique_attempt_id` within that same pair. All revision transport retries retain that link. Available response records may be shared within one pair's critic retries, but cannot be reused across pairs/runs. Separate identical-text response records do not prove independence; semantic provenance remains a human question. - `revision_coverage` is `attempted` with nonempty revision attempts, `not_attempted` with an explicitly known empty list, or `unknown` with an empty list whose coverage is unavailable. Empty lists do not automatically establish that no revision occurred. Every designated critic/revision attempt needs its own chronological `packets` slot, including null record/hash for missing inputs. A linked critique's explicitly supplied sequence must precede its revision; sequence order is checked for consistency, not used to infer stages or pairings. Unknown runtime logging completeness still requires `packet_inventory_complete: false`. Initial-only scope does not declare or validate B2 pairs. For combined scope, add the existing common-initial declaration and its packet slots. For continuation-only scope, leave `initial_input` null; common initial requests remain explicitly accounted for in worker stage declarations but outside validation. ## Structural checks and evidence The exact literal [critique](../baseline-prompts/b2-critique.txt) and [revision](../baseline-prompts/b2-revision.txt) templates are pinned respectively to SHA-256 `94211f9a78516ead0837d2e6a0d8960f9ff15b891c6a7bb63c67b61acb2412bb` and `e4004fce79bb54ba198730dc2190a9d743917ce9a49b1ff40f6b968d80ef5962`. Extra prose/blocks, altered wrappers, duplicate JSON keys, invalid UTF-8, unresolved placeholders, wrong types/IDs, and unsupported contracts/versions are rejected. Both inputs must contain the full immutable public-packet JSON string matching the registry hash, current artifact, at most two earlier artifacts, and at most twelve prior directive records with the existing unique-ID and permitted-field rules. Compact/indented JSON is accepted. Raw B2 bytes are checked first; a non-exported B1 projection reuses its compatible bounded-context validator. Its internal wrapper size is not a B2 resource estimate: the original B2 input byte cap is independently enforced. No repaired or synthesized request is exported. All supplied contexts within a pair, including transport retries, must be identical after JSON decoding, including IDs, full text, array order and CR/LF. Each revision has exactly one additional `CURRENT_CRITIQUE_JSON` **string**, whose UTF-8 bytes must equal its linked archived critique response. Critique requests have no such field. No response is inferred from revision text, selected by quality, promoted to a delivered artifact, or supplied to another pair's revision. Reused historical feedback disguised inside artifact/history strings may still pass structural checks; humans must determine leakage and whether directives are permitted. The adapter verifies only declared records, not hidden provider/system/chat context, true receipt, honest stage labels, token counts, critique correctness, ordinary-prose semantics, complete logging or the reason a retry occurred. It is not a reproduction or equal-budget claim. Existing resource accounting separately counts every failed, discarded and retry request; no costs are recalculated here. ## Coverage and human judgments Existing blank independent/adjudicated forms and exact hash-bound packet evidence are reused. Reviewers assess request-input packets; the response is not a new deliverable or quality rating. Copies are labeled `b2_critique_input` and `b2_revision_input`. Raw response records remain in the coordinator archive; their exact content is already present in a linked revision when available. `b2_pair_coverage` reports pair/attempt denominators, original/retry and outcome counts, verified links, missing inputs, unavailable versus declared-unproduced responses, known critiques without revision, and unknown revision coverage. Run/condition/series summaries retain all assigned-run denominators. A missing input/response or unknown revision history prevents complete evidence coverage and scoped compliance from a zero human judgment. A confirmed human violation still makes the run noncompliant despite other unknowns. No judgment is generated automatically; independent and adjudicated series remain separate. Known critique-only pairs can have complete audit evidence for the observed inputs, but are never counted as fully linked critique/revision pairs. A scoped compliant label does not imply a revision occurred, the B2 workflow finished, or the task succeeded. An available critique response is fallible internal feedback, not an external evaluator score or proof of support. Initial-only, continuation-only and combined coverage remain explicit; `whole_run_compliance_certified` stays false. Run local temporary artificial-fixture checks with: ```sh cd 'Japa two/observation-boundary' python3 -B -m unittest -v test_b2_contract.py ``` These fixtures are not actual study inputs, judgments, trajectories or findings.