{"criteria_id":"run-artefact-correspondence/v1","criteria_revision":"1.0.0","family":"run-artefact-correspondence/v1","family_note":"Criteria id run-artefact-correspondence/v1 is the coordinator's proposal in the M-49 brief, recorded as an interpretation in REGISTRY.md. Source: BUILD-QUEUE Q-26 + agent-fabrication-quality RESEARCH section 6 C1 (Class A tool-result + Class B completion-claim, one service, two artefact classes).","ops":["check"],"ops_not_offered":[],"corpus":{"id":"run-artefact-correspondence-corpus/v1","sha256":"ff2d77e8a9c16bf2a2987b688aa954f138e6b97ea0fb902a44631bf0b93173ef"},"corpus_pin_definition":"The corpus pin is the SHA-256 of the corpus's canonical serialization (the corpus object excluding its sha256 field), computed and verifiable by the published harness (harness.mjs). Expected outputs inside corpus cases are derived from the criteria text by an independent implementation in gen-corpus.mjs (never the engine's code path), and every case is a synthetic pair authored in gen-corpus.mjs, provenance-labelled in corpus.json.","price":{"usdc_units_per_call":20000,"label":"$0.02","basis":"per report check (one submitted transcript + report claims + evidence)"},"free_tier":{"per_day":100,"min_interval_ms":5000,"on_exhaustion":"402"},"input_contract":{"body":"One JSON request body carrying transcript, report, and evidence, plus caller? on the free pre-check only.","caps":{"body_bytes":262144,"claims_per_class":200},"over_cap":"An over-cap body is a free pre-gate 413 refusal naming the cap; an over-cap claim count is a named free refusal. No partial or truncated verdict exists in this service.","validation":"The submission is validated FREE, before any payment: format acceptance, per-format shape, claim typing, and the published caps. A malformed submission fails free and never consumes a payment. Grading runs only after payment verification (or on a verified-marker resume with no second charge)."},"input_contract_text":"Input contract: the request body is one JSON object { transcript: { format, content }, report: { claims_a, claims_b }, evidence } plus caller? (free pre-check only). transcript.format is one of the accepted formats: letta-af (a Letta AgentFile .af JSON object; the engine reads its message history), anthropic-messages (a JSON object { messages: [...] } or a bare message array in the Anthropic Messages shape), run-transcript-jsonl (a string of one JSON object per line, each { type: tool_call | tool_result | text, ... } — the service's published generic envelope). The transcript text graded against is extracted as: every text part, every tool-return / tool-result content, and the JSON serialisation of every tool-call input. Class A claims: { kind: execution, tool } | { kind: figure, value } | { kind: quote, text }. Class B claims: { field, claim_type: identifier | count | byte_size | hash, value }; evidence is one flat JSON object and field names a top-level key. Caps: request body up to 262,144 bytes (the binding cap for a transcript; an over-cap body is a free 413 refusal naming the cap, before any payment, and no partial verdict exists); up to 200 claims per class (over-cap is a named free refusal).","malformed_rule":"Malformed submissions (named FREE refusals, before any payment): body not a plain object (invalid_request_body); transcript missing or not an object (transcript_required); format not one of the accepted formats (transcript_format_not_accepted); content not in the format's shape (transcript_content_invalid); a JSONL line that is not JSON (transcript_line_not_json) or whose type is not tool_call / tool_result / text (transcript_record_type_invalid); report missing or not an object (report_required); evidence missing or not a plain object (evidence_required); claims_a / claims_b present but not arrays (claims_not_array); more than 200 claims in a class (claims_cap); a Class A claim whose kind is not execution / figure / quote — including untyped prose claims (claim_kind_not_typed); an execution claim without a tool string (claim_tool_required); a figure claim without a value (claim_value_required); a quote claim without a text string (claim_text_required); a Class B claim without a field string (claim_field_required); a Class B claim_type outside the published set (claim_type_not_typed); a Class B value of the wrong type for its claim_type (claim_value_type_invalid). A malformed submission fails free and never consumes a payment. Error surfaces carry static strings plus numeric facts only — never any part of a submission.","grading_rule":"Grading rules. Class A, against the submitted transcript: an execution claim is PRESENT when a tool-call record with the exact claimed tool name exists; FOUND-AS-VARIANT under N5 (tool-name normalisation) citing N5; otherwise ABSENT. A figure or quote claim is PRESENT when its string appears exactly in the extracted transcript text; otherwise the cumulative normalisation stages are tried in order (stage 1 = N1; stage 2 = N1+N2; stage 3 = +N3; stage 4 = +N4; stage 5 = +N6) and the first stage that finds it grades FOUND-AS-VARIANT citing every rule applied up to and including that stage; found at no stage, it is ABSENT. Class B, against the submitted evidence: the named field absent from the evidence grades ABSENT; present and equal under the claim type's published equality grades PRESENT-AND-MATCHING; present and unequal grades PRESENT-BUT-DIFFERS with the field named (the evidence value is never echoed). Typed equality: identifier = exact string equality; count and byte_size = numeric equality; hash = hex equality, case-insensitive. Per-claim grades are returned by claim index (plus the field name for Class B), with counts per type. Factors are printed; there is no composite score, grade, or band anywhere in this service.","normalisation_rules":{"N1":"case-fold (Unicode lowercase)","N2":"whitespace collapse (every run of whitespace becomes one space; trim)","N3":"quote/punctuation normalisation (curly quotes become straight; surrounding quote marks are stripped)","N4":"numeric canonicalisation (thousands separators and currency symbols removed; trailing decimal zeros dropped — 1,500.00 becomes 1500)","N5":"tool-name normalisation (case-fold; hyphen and underscore treated as equivalent) — execution claims only","N6":"Unicode NFC normalisation plus line-ending normalisation (CRLF / CR become LF)"},"accepted_formats":["letta-af","anthropic-messages","run-transcript-jsonl"],"accepted_formats_evidence":"letta-af: Letta AgentFile (.af) — a purpose-built export artifact (ADE Export Agent / REST export / SDK exportFile) packaging complete message history, per the X6 format survey on disk (lane-1 sections 2, 9.1, 9.4; confirmed in that pass's VERIFY.md). anthropic-messages: the operator-held request/response JSON in the Anthropic Messages shape (tool_use / tool_result blocks paired by tool_use_id), per the same survey. run-transcript-jsonl: this service's published generic envelope, the catch-all for harnesses whose record the operator serialises themselves (survey section 9.2 common core; fleet generic-schema precedent). CrewAI and elizaOS have no native whole-transcript class on the record and are not accepted as native formats.","claim_types_b":["identifier","count","byte_size","hash"],"not_assessable":["Whether the submitted transcript is authentic (a doctored log passes its own check).","Whether the submitted transcript is complete (an excerpt grades as the whole submitted artefact, and is never disclosed as more).","Whether a present tool result was used correctly downstream.","The quality of the work, or whether the task was done well — or done at all beyond the submitted evidence.","Whether an absent claimed execution happened through a channel not in the submitted transcript.","The provenance of the submitted evidence (evidence fabricated to match the report grades as matching — correspondence, not provenance).","Claims submitted as untyped prose (untyped = ungradeable; refused free at validation, never interpreted).","Whether the agent's interpretation of a present result is semantically correct.","Anything about runs, tools, or artefacts not submitted."],"boundary_line":"Issue #6 boundary: this service grades the document question — correspondence between the submitted report and the submitted transcript and evidence. It never observes a run, issues or validates no runtime receipts, and answers nothing about whether an execution actually happened in the world — only whether it is present in the transcript as submitted. A claim that could only be settled by runtime observation is NOT-ASSESSABLE here, not graded.","wording_fence":"Wording fence: no surface of this service (criteria, catalog, report, pre-check, errors, metrics) may state or imply that a task or run was verified complete, that work was certified done, or that a report is true. The only sentence in this criteria that names the prohibited claim does so to prohibit it: no report from this service may be presented, quoted, or relied on as a \"verified complete\" verdict — presenting one that way triggers the stop-and-review tripwire in the kill line.","disclosure_sentence":"Submitting a transcript to any third party, this service included, gives its contents to a new holder however briefly; the check is processed in memory only and never stored, but it does not undo the disclosure, and an operator who suspects a live credential in a transcript should rotate it at its issuer rather than submit the transcript.","verdict_grammar":{"line":"RUN-ARTEFACT CORRESPONDENCE — A present: <n> · A variant: <n> · A absent: <n> · B matching: <n> · B differs: <n> · B absent: <n> — <check-end UTC>","vocabulary":"The report's facts are the published closed set: per-claim Class A grades (PRESENT / ABSENT / FOUND-AS-VARIANT with the normalisation rules cited), per-claim Class B grades (PRESENT-AND-MATCHING / PRESENT-BUT-DIFFERS with the field named / ABSENT), and counts per type. There is no composite score, grade, or band anywhere in this service, and no field states or implies that a task or run is complete."},"free_precheck":"The free pre-check (POST /precheck) validates the submission under the published contract and returns assessability ONLY: the common signed fields, whether the submission is assessable (or the free refusal naming the failure), the accepted formats, the transcript's tool-call count, the claim counts per class, and the NOT-ASSESSABLE list — with NO grading run and NO grade returned. Free allowance: 100 per day per caller wallet, minimum interval 5,000 ms, 402 on exhaustion.","custody":{"construction":"In-memory processing within a single request lifetime; submissions persist nowhere, on any path including errors (the O5 no-retention disciplines, binding from the design's first page — transcripts carry secrets).","ledger_closed_fields":["payer wallet","timestamp/day","price (USDC units) with the CAD fair-market-value block (rate, rate date, basis, cad_pending, refund)","payment tx id + Solana fee (lamports)","criteria version","verdict line","op","grade counts per type (a_present, a_variant, a_absent, b_matching, b_differs, b_absent)","house flag"],"ledger_note":"The ledger list above is closed: adding a field is a criteria/handling version change. The ledger NEVER carries any transcript content, any claim value, any evidence value, or any hash of a submitted value.","errors":"Error handling is closed-form: static strings plus numeric facts only (lengths against caps, refusal codes). No error path interpolates any part of a submission."},"forbidden":"Forbidden in every surface of this service (report, pre-check, errors, ledger, metrics): any composite score or grade; any submitted content, claim value, or evidence value; any hash of a submitted value; any statement or implication that a task or run is complete or that work was certified done; any runtime observation, receipt issuance, or live contact with any endpoint or store.","honesty_line":"Every grade in the report is computed from the submitted transcript and the submitted evidence under the published rules and nothing else: the service observes no run, contacts no endpoint, executes nothing, and infers nothing about any artefact beyond the correspondence of the submitted pair. A grade the rules did not produce is never guessed or filled in.","retention":"Submissions are processed in memory only and are never stored — not in KV, not in logs, not in the payments ledger, not in metrics, and not in error records. The report carries per-claim grades by index, field names for Class B, and counts only — never transcript content, never claim values, and never evidence values, and no hash of any submitted value. The payments ledger records only the closed field list published in the criteria. Because nothing is kept, a paid retry re-submits the transcript, the claims, and the evidence.","payment_includes":"$0.02 per report check. The payment includes exactly one correspondence check of the one submission you send (a transcript in one accepted format, a report's typed claims, and one evidence artefact, inside a request body of up to 262,144 bytes, up to 200 claims per class) and one signed correspondence report under run-artefact-correspondence/v1 stating: per-claim grades for Class A tool-result claims (PRESENT / ABSENT / FOUND-AS-VARIANT, each variant citing the normalisation rules that matched it) graded against the submitted transcript, and for Class B typed completion claims (PRESENT-AND-MATCHING / PRESENT-BUT-DIFFERS with the field named / ABSENT) graded against the submitted evidence; counts per type; and the NOT-ASSESSABLE list. No report states or implies that a task or run is complete. A free pre-check validates the submission's shape before any payment and returns assessability only — no grade is computed free. Submissions are processed in memory only and are never stored; the payments ledger records only the payer, the time, the price, the grade counts, and the verdict line.","kill_line":"Kill line (M-49 / Q-26, under the operator's 2026-11-03 build-bank-hold doctrine): build, bank, and hold; judged at the fleet's periodic reviews. There is NO demand kill line for this item. Day-0 = first outside paid settle (accounting only; house proofs never count). The Q-26 format-survey tripwire was settled at design (DESIGN.md section 1: PASS, narrowly — one native export format on the record plus one operator-held JSON class plus the published generic envelope). Tripwires, all stop-and-review: (1) refunds >5% of paid calls in any 30-day window; (2) free use >10x paid use; (3) a wrong verdict from undisclosed stale or partial data (stop-serving under the affected criteria version; the error enters the corpus first as a regression case, corrected rules publish as a new criteria version, the corpus re-passes at 100% before serving resumes; issued reports are never amended); (4) a correspondence report circulating as a \"verified complete\" claim; (5) any retained surface carrying submitted content or a hash of a submitted value; (6) any composite score on any surface; (7) the only demand found being for the runtime-receipt or model-judged forms (those forms are barred and are never built here).","never":"This service never fetches anything; never executes anything; never observes a run; never stores a submission; never prints submitted content or a hash of a submitted value; never states or implies a task or run is complete; never computes a composite score.","service":"artefact-correspondence","version":"1.0.0","endpoints":{"criteria":"GET /criteria (free)","precheck":"POST /precheck (free, input validation + assessability only: no grading, no grade)","check":"POST /check (paid, $0.02 per report check)","metrics":"GET /metrics","not_offered":"Any other POST path answers NOT COVERED. No best-effort result is produced."}}