{"criteria_id":"memory-provenance-audit/v1","corpus":{"id":"memory-provenance-audit-corpus/v1","sha256":"b026a11c1433ae9b2502d4d2c83d4b762531a7ef49c1b5e4cfc13621ce8cb131"},"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).","price":{"usdc_units_per_block":50000,"label":"$0.05","basis":"per block","receipt_usdc_units":5000000,"receipt_label":"$5.00"},"block":{"claims_per_block":10000,"bytes_per_block":10000000,"formula":"blocks = max(ceil(claims / 10000), ceil(bytes / 10000000)), minimum 1","minimum_blocks":1,"submission_ceiling":"none","byte_count":"the raw submitted payload bytes as received (for a chunked submission, the sum of the part bytes as received), printed on the verdict"},"free_tier":{"per_day":100,"min_interval_ms":5000,"resets":"the daily count resets at 00:00 UTC","keyed_on":"the caller wallet value passed as ?caller=<wallet> (self-declared for the free pre-check; the paid audit needs no caller — the payer is read from the payment)","on_exhaustion":"402 — when the day's 100 free pre-checks are used up, or a pre-check arrives less than 5 seconds after the previous one, /precheck returns a 402 payment challenge for the paid audit, with the reason stated in the response's free_tier note, instead of a pre-check result; the pre-check itself is never charged. /precheck-chunked returns the same 402 status with an error body naming the reason.","includes":"format recognition + record count + assessability preview (which checks will run and which will be NOT ASSESSABLE, with reasons) + the exact block quote"},"formats":["mem0-json","zep-graphiti-edges","generic-jsonl","hash-only (any format: claim text replaced by text hashes, plus structural fields — the hashes and structural fields can still identify individual records)"],"generic_schema":{"name":"generic claim JSONL — one JSON object per claim-event","fields":["claim_id","text | text_hash","source","learned_at","valid_from?","valid_to?","supersedes?","reason_code?","reason_text?","evidence_hash?","state?"],"note":"The same event shape may also be submitted as a JSON array or as a JSON object carrying the records under records, claims, memories, edges, or facts, optionally with a top-level history, events, or rewrite_log array and a top-level evidence_spans array."},"mem0_note":"Mem0 JSON export records carry {id, memory, hash, metadata, created_at, updated_at, user_id | agent_id | run_id} and may carry replaced_by, lifecycle_state, categories, synthesized, and input. Mem0's hash field is a hash of the memory TEXT itself: by rule it never counts toward evidence addressability (check C4); it is counted separately as self_text_hash_excluded. Degrees are graded, not assumed zero: a populated replaced_by link is real supersession evidence.","zep_note":"Zep / Graphiti edge exports carry temporal edges with valid_at / invalid_at validity windows plus created_at / expired_at stamps and episode (source) references. Edges are facts by construction for check C6 typing; currentness by validity window (invalid_at absent) is an operative-version rule for check C7.","checks":[{"id":"C1","property":"P2","name":"Supersession completeness","text":"Of the content-changed records (updated_at later than created_at), the percentage carrying a supersession link (a supersedes pointer, a replaced_by pointer, being the target of another record's supersedes pointer, or a history record for that claim). The complement count is reported as silent_overwrites: updated after creation with no link and no history record. NOT ASSESSABLE when no record exposes the created_at/updated_at pair and no history is submitted."},{"id":"C2","property":"P3","name":"Adjudication presence","text":"Of all supersession edges, the percentage whose superseding record (or whose history event) carries a reason at all (reason_code or reason_text non-empty; Graphiti/Zep invalidation reason fields count). NOT ASSESSABLE when the submission contains no supersessions."},{"id":"C3","property":"P6","name":"Adjudication structure","text":"Of the adjudicated supersessions (the C2 numerator), the percentage carrying a structured reason code (reason_code non-empty), as opposed to free text only. NOT ASSESSABLE when no supersession carries any reason."},{"id":"C4","property":"P5","name":"Evidence addressability","text":"Per claim: an evidence pointer that is a content hash resolving against a submitted evidence span counts as addressable; a content hash with no matching submitted span counts as unresolved-hash; a bare row ID or URL pointer counts as bare; no pointer counts as none. A record's own text hash (including Mem0's hash field) is excluded by rule and counted as self_text_hash_excluded, never as evidence. The factor reports addressable_pct = resolving / total claims, plus every category count."},{"id":"C5","property":"P4","name":"Bitemporal consistency","text":"Whether separate valid-time and learned-time axes are present, the percentage of claims carrying both axes, and the count of overlapping-validity contradiction pairs: two versions of the same claim (same claim identity or same text/text_hash) whose validity windows overlap with no supersession link between them."},{"id":"C6","property":"P1","name":"Claim/fact separation","text":"The percentage of records typed as claim/opinion vs fact (an explicit type, kind, record_type, or claim_type field; Zep/Graphiti edges are facts by construction), and the percentage of claims carrying a source attribution (source, episodes / source_episodes, or a submitted input / provenance source)."},{"id":"C7","property":"P7","name":"Operative-version rule","text":"Present or absent, with the detected rule named: an explicit operative_version_rule / version_rule field; a current/head flag on records; a lifecycle or state field determining current records; validity-window currentness (Zep/Graphiti); or a supersession chain whose heads (records superseded by no other) are determinable (head-of-chain). Stated as a factor — the audit never judges which version is true. An explicit rule is echoed verbatim, control characters stripped, up to 200 characters; a longer explicit rule is truncated to its first 200 characters and the factor additionally carries rule_truncated: true, the full sanitised rule's character length (rule_length), and the first 16 hex characters of its SHA-256 (rule_sha256_prefix)."},{"id":"C8","property":"P8","name":"Contested-state handling","text":"Where the export exposes a state field (state, lifecycle_state, or status): the distinct states present, their counts, and whether any contested state exists (pending, quarantine, quarantined, contested, disputed, or review) as opposed to instant-overwrite-only. NOT ASSESSABLE where the export exposes no state field."},{"id":"C9","property":"chain","name":"Chain integrity","text":"Runs only where a history or log is submitted (a top-level history, events, or rewrite_log array, or claim records themselves carrying event_hash / record_hash plus prev_hash fields — a plain text hash, including Mem0's hash field, is never a chain hash). In submitted order: gaps = events whose prev_hash does not match the previous event's hash; out_of_order = events whose learned_at is earlier than the previous event's; missing_hash = events with no chain hash."}],"verdict_line":{"values":["PROVENANCE RECORD: COMPLETE","PROVENANCE RECORD: PARTIAL","PROVENANCE RECORD: ABSENT","PROVENANCE RECORD: NOT ASSESSABLE"],"rule":"NOT ASSESSABLE when there are zero parseable claim records or the format is unrecognised. ABSENT when at least one percentage factor is assessable and every assessable percentage factor is 0, no supersession link exists, and no operative-version rule is present. COMPLETE when every check C1–C8 is assessable, there are no silent overwrites, there are no overlapping-validity contradiction pairs, the chain (where it ran) has no gaps, no out-of-order events, and no events missing a chain hash, every assessable percentage factor is 100, an operative-version rule is present, and a contested state exists. PARTIAL in every other case, including a submission whose records expose no assessable percentage factor but do not meet the ABSENT conditions. There is no composite score."},"verdict_display":"The verdict line is never displayed by this service without its coverage block and the stated exclusion in the same object; the line alone certifies nothing.","hash_only":{"text":"In hash-only mode the submitter replaces claim text with text hashes plus structural fields, for any format; a submission in which no record carries claim text is treated as hash-only whether or not the mode was requested. Coverage discloses exactly what is reduced: evidence spans are matched by hash only (span content is never content-verified), and overlapping-validity identity is by text_hash only. All other checks run on the structural fields as submitted. Hash-only keeps claim text out of the submission; it does not make a submission anonymous — the text hashes and the structural fields can still identify individual records, and may still be personal information where the export concerns people. For an export that may contain personal information about other people, hash-only is the mode designed to keep that claim text out of the submission entirely.","reduced":["evidence span content verification (hash match only)","overlap identity by text_hash only"]},"chunked_submission":{"text":"Payloads above the platform request-body limit use the chunked submission path. POST /session accepts a manifest {format, part_count, per-part SHA-256 hashes, declared totals (claims and bytes)} and opens a session in a session-scoped namespace with a TTL of 1 hour; the declared totals are informational — the free pre-check (POST /precheck-chunked) and the audit are computed from the assembled parts exactly as received. Parts upload against the session in any order. The audit becomes payable and runnable only when every part is present and each part's SHA-256 matches the manifest; an incomplete manifest means no audit and no charge; a part whose hash does not match is rejected and the session stays open; on session TTL expiry the parts are deleted. Each part's TTL is the session's remaining lifetime from the manifest's opened_at, so no part expires later than the session or its manifest. Parts are deleted immediately when the audit completes. Parts are never logged. Single-request submissions bypass this path entirely and are processed in-request only.","session_ttl_seconds":3600},"not_assessable_rules":["A check that cannot be computed from the fields a submission exposes is reported NOT ASSESSABLE, with the reason stated — never as a silent zero.","Missing updated_at (or the created_at/updated_at pair) with no history record: check C1 (supersession completeness) is NOT ASSESSABLE.","No supersessions in the submission: checks C2 and C3 (adjudication) are NOT ASSESSABLE — there is nothing to adjudicate-grade.","No adjudicated supersession: check C3 (adjudication structure) is NOT ASSESSABLE.","No state field exposed by the export (no state, lifecycle_state, or status on any record): check C8 (contested-state handling) is NOT ASSESSABLE.","No history or log submitted: the chain-integrity check does not run and the coverage block says so.","A submission with zero parseable claim records, or a submission whose format is not recognised, grades overall NOT ASSESSABLE — it is never partially graded and never silently truncated."],"scope_block":"This verdict grades the revision record of the submitted export exactly as submitted, bound by the submission SHA-256 printed in the coverage block. A verdict line of COMPLETE means only that the export as submitted met every condition of the published verdict-line rule on the checks that ran — it is not a statement about the whole store the export was taken from, about any claim's truth, or about safety. It does not verify the truth of any claim in the export, it does not certify the whole store from which the export was taken (an operator could submit a curated subset and the audit cannot detect that), and it is not a safety or correctness certification. Factors are computed only over what was submitted and say so: every verdict carries its coverage block first. COMPLETE describes only the internal completeness of the submitted export's own revision record under these mechanical checks. No check compares the export against the store it was taken from or against any external source, so a curated or fabricated export that is internally consistent grades exactly like a faithful one.","custody":"Single-request submissions are processed in-request only: the submitted content is never written to storage and never logged. Chunked submissions (the path for payloads above the platform request-body limit) are different in one stated way: the parts are held transiently in a session-scoped namespace under a short session TTL (1 hour), the audit becomes payable and runnable only when every part in the manifest is present and each part's SHA-256 matches the manifest, the parts are deleted immediately when the audit completes and are deleted by TTL if the session expires, no part outlives the session (every part's TTL is the session's remaining lifetime from the manifest's opened_at, so no part expires later than the session or its manifest), the parts are never logged, and the parts are never used for anything else. Retained after any audit is one payment-ledger line on the published closed field list (the payment ledger is the service's business record and is retained without expiry): the payer wallet, the payment transaction, the price, and the coverage counts (format, claims, bytes, blocks, verdict line, checks run, and the count of checks not assessable), plus the CAD valuation fields on that list. The submission SHA-256 and the full factor values are returned to you in the verdict and are not stored by the service. Keep your signed verdict: the anchored receipt requires the submission SHA-256 and the verdict hash from you, because the service does not hold them. Three operational records are also kept, none containing submitted content: payment and replay markers keyed by the payment transaction (delivery state, payer wallet, and network fee), kept for 7 days so a paid call can be resumed or an owed receipt write completed without a second charge; free-allowance counters keyed by the caller wallet, kept for 2 days; and a list of distinct payer wallets, kept for the service's public distinct-payer count — a payer wallet identifier is kept for 12 months after that payer's last payment and then expires, and the distinct-payer count itself is an aggregate number kept indefinitely — together with aggregate daily and total counters (numbers only). No claim text, no identifier of any record in your export, and no other submitted content is retained in any log, ledger, metric, or stored record, beyond the chunked session parts while a session is open, as stated above.","exclusion":"This verdict grades the revision record of the submitted export. It does not verify the truth of any claim, and it is not a safety or correctness certification.","privacy_note":"Privacy note: this service is operated from Quebec, Canada. A submitted export may contain personal information about people other than the submitter (memory records can carry user_id, agent_id, or run_id and facts learned about people). Submit only what you hold the right to submit; hash-only mode keeps claim text out of the submission. The custody block states everything the service holds and retains. Privacy questions, and access, rectification, or deletion requests under Quebec's Law 25 or PIPEDA, go to the person in charge of the protection of personal information for this service: privacy@exchangerebel.com.","payment_includes":"$0.05 per block. One block is up to 10,000 claim records AND up to 10 MB of submitted payload, whichever limit is reached first; blocks = max(ceil(claims / 10,000), ceil(bytes / 10,000,000)), minimum 1 block, and there is no submission ceiling. The payment includes: one Memory Provenance Audit of one whole submitted memory export — the eight published provenance checks of memory-provenance-audit/v1, plus chain integrity where a history or log is submitted, computed over the export exactly as submitted — returned as a verdict signed with this service's own Ed25519 service key: the coverage block first, the verdict line, all factor values, the submission SHA-256, the as-of time, and the stated exclusion. The exact block count and the exact total price are quoted at the free pre-check before any payment; one audit is settled in one payment of blocks x $0.05. By submitting the export you warrant that you hold the right to submit it and to have it processed as described here, including any personal information it contains. Delivery is the response to the paid call. Payments settle on Solana before delivery. If a verified payment does not produce a verdict in that response, re-send the same request with the same X-Payment-Tx header within 7 days and the service delivers from the verified payment with no second charge; a payment that has been delivered against cannot be used again. If we still fail to deliver, we refund the payment. You only pay for delivered answers. A verdict that has been delivered is final: it is not refunded because the buyer disputes or disagrees with the verdict. Single-request submissions are processed in-request only: the submitted content is never written to storage and never logged. Chunked submissions (the path for payloads above the platform request-body limit) are different in one stated way: the parts are held transiently in a session-scoped namespace under a short session TTL (1 hour), the audit becomes payable and runnable only when every part in the manifest is present and each part's SHA-256 matches the manifest, the parts are deleted immediately when the audit completes and are deleted by TTL if the session expires, no part outlives the session (every part's TTL is the session's remaining lifetime from the manifest's opened_at, so no part expires later than the session or its manifest), the parts are never logged, and the parts are never used for anything else. Retained after any audit is one payment-ledger line on the published closed field list (the payment ledger is the service's business record and is retained without expiry): the payer wallet, the payment transaction, the price, and the coverage counts (format, claims, bytes, blocks, verdict line, checks run, and the count of checks not assessable), plus the CAD valuation fields on that list. The submission SHA-256 and the full factor values are returned to you in the verdict and are not stored by the service. Keep your signed verdict: the anchored receipt requires the submission SHA-256 and the verdict hash from you, because the service does not hold them. Three operational records are also kept, none containing submitted content: payment and replay markers keyed by the payment transaction (delivery state, payer wallet, and network fee), kept for 7 days so a paid call can be resumed or an owed receipt write completed without a second charge; free-allowance counters keyed by the caller wallet, kept for 2 days; and a list of distinct payer wallets, kept for the service's public distinct-payer count — a payer wallet identifier is kept for 12 months after that payer's last payment and then expires, and the distinct-payer count itself is an aggregate number kept indefinitely — together with aggregate daily and total counters (numbers only). No claim text, no identifier of any record in your export, and no other submitted content is retained in any log, ledger, metric, or stored record, beyond the chunked session parts while a session is open, as stated above. This verdict grades the revision record of the submitted export. It does not verify the truth of any claim, and it is not a safety or correctness certification.","receipt_includes":"$5.00 buys one anchored receipt for one Memory Provenance Audit verdict: you supply the submission SHA-256 and the verdict hash from a verdict this service issued, and the service writes those two hashes — and nothing else from the verdict — to Solana in a single memo transaction by the service's own wallet (the service absorbs the anchor write cost inside the $5.00). The memo is a public Solana transaction and is permanent, as all Solana transactions are. Anyone who already holds the verdict can recompute its hash and confirm, from the anchor transaction's on-chain time and without calling this service, that this exact pair of hashes had been recorded by then; the anchor does not make the verdict's contents public and proves nothing about the export beyond the two hashes. The service does not retain verdicts, so it cannot check the two hashes you supply against an issued verdict: the receipt anchors the pair you supply, and its accuracy as a receipt for a particular verdict depends on you supplying that verdict's true hashes. If the on-chain write does not confirm after your payment settles, the write is owed, not lost: retry POST /receipt with the same X-Payment-Tx header and it completes with no second charge (the owed state is tracked for 7 days). If the owed write cannot be completed, we refund the $5.00 payment. You only pay for delivered answers. A receipt whose anchor write has completed is final and is not refunded. A receipt involves no submission content: no part of any export is sent, held, or stored for a receipt, and the custody block in the criteria document governs the underlying audit. It does not re-run, upgrade, or extend the underlying audit.","retention":"Retained after any audit = one payment-ledger line on the closed field list (payer wallet, payment transaction, price, coverage counts, and the CAD valuation fields on that list), plus the operational records stated in the custody block (7-day payment and replay markers, 2-day free-allowance counters, the distinct-payer identifier list, and aggregate counters). Payer wallet identifiers expire 12 months after that payer's last payment; the aggregate distinct-payer count and the payment ledger are retained indefinitely as business records. The submission SHA-256 and the full factor values are delivered in the verdict and are not stored. No claim text, no identifier of any record in the submitted export, and no submitted content is retained in any log, ledger, metric, or stored record. Chunk parts are the single stated exception while a session is open, under the chunked-submission mechanism above.","ledger_closed_fields":["payer","ts","date","rung","price_usdc_units","payment_tx","solana_fee_lamports","criteria_version","counts (audit rungs only: format, claims, bytes, blocks, verdict_line, checks_run, checks_not_assessable_count — counts and flags only; null on receipt rungs)","house","cad_rate_usdcad (added at append)","cad_rate_date (added at append)","cad_basis (added at append)","cad_fmv (added at append)","cad_pending (added at append)","refund (added at append, null unless refunded)","anchor_tx (receipt rung only, when the anchor write completes)","our_write_fee_lamports (receipt rung only)","anchor_status (receipt rungs only: write_owed on the owed line, stubbed or complete on the settled line)","anchored-receipt-completion rung (closes an owed anchored receipt at price_usdc_units 0 — a completion line, not a second payment line)"],"never":"This service never verifies the truth of any claim; never issues a composite score; never certifies a whole store from a submitted export; never silently truncates or partially grades a submission (an incomplete chunked manifest is no audit and no charge); never counts a record's own text hash as evidence; never logs submitted claim text, and never stores it except as chunked session parts under the custody block above; and the grading engine itself fetches nothing from any external register or source — payment verification reads the Solana chain, and the ledger's CAD valuation reads the Bank of Canada's daily rate after grading; neither affects any factor or the verdict line.","kill_line":"Kill line: NONE — build, bank, hold (ruled 2026-11-03). The service's internal stop-serving tripwires are operating rules, not part of this offer; what a payment buys, what is delivered, and what happens if delivery fails are stated in the payment-inclusion lines above."}