Humans: run the free evaluation. Agents: copy the curl example below (requires curl and Node.js). No card, subscription, payment, or pre-existing credentials.
Safety rules: party anchors and identifiers must match exactly; a near-match never establishes identity. Scheduled, pending, in-transit, attempted, partial, suspended, expired, rejected, and draft states do not establish completed events.
Caller-declared relationships are retained as provenance, not proof. A pinned local NLI model compares each full evidence statement with its proposition; source_span is only a verified quote locator. A support finding requires a strict entailment score of at least 0.9 against the complete evidence statement; its supporting_span covers that statement so the predicate fact cannot be clipped away by the caller's locator. Caller labels contribute zero. Scores are not calibrated certainty. Model-only opposition without a scoped subject-and-attribute conflict remains unresolved, not a proven contradiction. Deterministic vetoes for explicit identity/scope conflicts, different approving actors, negative or incomplete states, amount qualifiers, validity windows, and action dependencies run after scoring. Uncertain model results remain unresolved. Assessments are advisory, do not authenticate sources, and do not authorize execution.
relation_assessment: caller_declared records provenance, alongside semantic_predicate_entailment_with_deterministic_vetoes_v5. The model proposes semantic relationships; it does not independently establish source authenticity or factual truth. Typed party anchors, identifiers, status/event states, quantities and validity dates are checked as safety vetoes. Different explicit parties or identifiers, different approvers for the same approval, explicit negative/incomplete states, incompatible amount arithmetic, and expired validity cannot be overridden by model scores. Mere absence of a literal word is not a veto; semantic support may be expressed in field/value records or paraphrase. Optional proposition.statement is strongly recommended; an opaque ID without an interpretable claim returns a named gap, never caller-label fallback. Findings preserve caller_declared_relation, engine_assessed_relation, relation_discrepancy, uncertainty, and named_gap. supporting_span gives offsets into the submitted statement, not authenticated source proof. Source authenticity is unverified; ALLOW never authorizes execution. Present-tense currency uses an explicit observation time where supplied; freshness limits apply only when a proposition or action supplies one. Chronology requires same-subject scope and a decisive current state; a later timestamp alone does not supersede an earlier record. Authority remains a separate action requirement.
GET /v1/status and GET /api/status report the active free synthetic evaluator, engine_version and build_sha; legacy services may remain paused. Save a support ticket instead of emailing an unverified mailbox. POST /v1/support-tickets accepts {"message":"A website or API issue, not case evidence","traffic":"real"} without credentials. It returns a ticket_id after database save and attempts a notification to SUPPORT_NOTIFY_EMAIL through the configured Resend sender. A 503 can include saved:true and the ticket_id when email notification fails; the site tells you plainly. A successful provider acceptance is not proof of inbox delivery. If an evaluation reference is included, the operator email links to the protected saved result without placing its retrieval token in a URL. Message: 10–2,000 characters; request: 4 KiB. Shared POST-attempt caps in production: 60/IP/hour, 300/IP/day; development only: 600/IP/hour, 3,000/IP/day. The shared global attempt ceiling remains 20,000/day. Support messages, ticket IDs, traffic and timestamps are stored for 30 days, with bounded request-driven cleanup; no public ticket listing or lookup exists.
Synthetic evidence only. Do not enter real people's names or identifying information, including real addresses, case records, personal data or PHI. Application storage retains the decision, generated findings, evidence/source/track IDs, engine/build versions, expiry, journey ID and hashed request/access/idempotency metadata—not submitted statements or quoted source spans. These narratives are also excluded from application request logs, error responses and evaluation analytics. IDs themselves are retained, so keep them synthetic. Contact-like identifiers are rejected; this is not a comprehensive detector of all personal data. This describes application behavior, not infrastructure backup deletion.
Each new saved evaluation includes a version-1 receipt, exportable with GET /v1/evaluations/{id}/receipt using the same private Authorization: Bearer retrieval capability. It stores a canonical input SHA-256, engine/build identity, decision and findings, evidence/source/document IDs, scope and date references, statement hashes, exact source_span offsets, and optional pin status—not raw statements or quote text. The input-hash canonicalization sorts object keys by UTF-16 code-unit order, preserves array order, normalizes strings to NFC and line endings to LF without trimming other whitespace, then hashes JSON.stringify output. Statement pins hash exact UTF-8 statement bytes. Offsets are UTF-16 code units into the exact submitted statement; multiple exact occurrences are listed. A matching statement pin checks only the submitted text against its declared SHA-256. Missing pins are UNVERIFIED_NO_PIN; a mismatch is an integrity warning, not a semantic contradiction. Semantic assessment remains an engine judgment with uncertainty. Neither quote/hash checks nor a pin authenticate an external source, establish truth, guarantee legal admissibility, certify compliance, or authorize execution. The receipt expires and is deleted with its evaluation, by default after 24 hours and at most 168 hours; early DELETE removes both. Keep your original synthetic input if you need to independently verify the offsets or hashes.
Analytical findings per proposition: contradiction, chronological_evolution, separate_proceedings, scope_mismatch, unsupported_recommendation, insufficient_evidence. These feed a separate action decision: ALLOW / DENY / HUMAN_REQUIRED. ALLOW means the submitted action passed that evaluation; it does NOT authorize dispatch or any external execution. A verdict never proves the underlying facts. Source authenticity remains unverified.
Internal prepare_verification actions may pass despite named factual gaps; rely_on_claim and spend require their declared premises to be supported. Declare changing_account_fact propositions explicitly: an attributable source, observation/document date and matching synthetic account identity are required. An optional maximum observation age of 1–3600 seconds may be supplied on the proposition or action; no default age is assumed. Missing required metadata or exceeding a supplied age limit means verify first, never buy credits. Directory contact failures are treated neutrally: requesting substantiation may pass; relying on unverified availability may require a human. No mandatory staff-confirmation rule is assumed.
For scheduled actions, proposed_action.action_at accepts YYYY-MM-DD or an ISO timestamp with an offset. Explicit validity and expiry evidence is compared to this action date; without it, the evaluation date is used. An action date is not authority or permission.
DENY applies to explicit in-scope opposition or unresolved supporting and opposing evidence in a required premise of a reliance action; unsupported or stale premises generally require HUMAN_REQUIRED instead. A written refusal to provide the claimed service is distinct from a written channel restriction or unanswered contact. Missing identity metadata alone neither proves an entity mismatch nor erases an explicit submitted service refusal. Internal preparation may be ALLOW with gaps still named; no verdict authorizes external action. These labels describe submitted evidence, not independently established facts.
Optional propositions[].statement and subject_id fields make the intended synthetic assertion and subject explicit. Scope comparisons can also use bounded claim/case/order/invoice references in descriptive IDs or the submitted objective/action. evidence_findings attributes excluded scope evidence and opposing evidence separately. Explicitly dated mutable-status records for the same submitted subject, proposition and track can form chronological_evolution without sequence_group; arbitrary date order never supersedes unrelated facts. effective_evidence_ids identifies the current comparison. Relative deadline claims in an action require a cited cutoff basis: a review permission alone is not a deadline. Additional action_findings can name a missing premise without relabelling supported controls.
Saved findings include generated reasons, evidence/source/track IDs, normalized subject-reference tokens, normalized source dates with extraction offsets, and normalized time assertions where applicable. Original submitted statements and source_span strings are not saved as narrative fields. This is not a claim that no information derived from text is retained. Optional developer-interest responses are separately stored; do not submit identifying information there. The default evaluation retention is 24 hours, configurable up to 168 hours. Expired results are inaccessible; physical cleanup is bounded and request-driven, not guaranteed instantaneous erasure from every infrastructure backup.
Additional first-class findings are written_refusal and failed_contact. Use the matching source_type for a reported written communication-channel decision versus an unanswered, hung-up or dropped attempt with no communicated decision. A written channel restriction is not proof of inability to provide a service; failed contact proves neither a written refusal nor availability. contact_findings retain evidence/source/track IDs and supplied event chronology. These describe submitted reports, not authenticated external facts.
Use proposition kind event_time for one event-time claim on its action track. Explicit 12-hour AM/PM and 24-hour HH:mm clock values are compared; time_assertions retain bounded normalized values, display labels and source IDs, not verbatim submitted statements. Conflicting same-event sources remain UNRESOLVED with both sourced assertions visible in reasons. Different tracks, dates and explicit sequence metadata are considered; no source is silently selected. Verify the applicable date/timezone and current source before relying on an unresolved time. No result authorizes execution.
POST requires a private random UUID v4 Idempotency-Key. Identical validated input and key replay the same record and capability; changed input gets 409. Keep the key private. POST returns an unguessable retrieval_token; GET and DELETE require it in Authorization: Bearer. Wrong, missing, expired and deleted records all return 404. There is no list endpoint. A UUID alone is not access. Idempotency lasts only until the record expires or is deleted; do not reuse an old key for a new run.
Use synthetic:true and fictional content only. Technical IDs such as SRC-01 and E1 are accepted (1–64 letters, digits, underscores or hyphens, starting with a letter). Ordinary insurer, agency, customer and checkout vocabulary, currency amounts, dates, clock times and procedure-code lists are not forbidden words or a realism score. References must resolve. source_span must be an exact substring of statement and is only a quote locator; evaluation assesses the full statement, not just that span. A synthetic declaration does not waive contact/private-data checks or prove that content is fictional.
Invalid input returns 422 with field and issues containing safe code/reason values: malformed JSON, over 16 KiB, field/count limits, unresolved references, span mismatch, or a specific privacy pattern. Email addresses, formatted phone numbers, SSN-like values, card-checksum patterns, explicitly labelled personal IDs and unclassified long digit runs remain restricted. Clearly labelled technical references and currency amounts are distinguished from unlabelled long numbers; this is not comprehensive PII detection. Errors do not echo submitted content. Idempotency-Key is named only when its own validation fails. Limits and privacy checks remain active for synthetic requests.
16 KiB JSON; 12 propositions; 30 evidence entries; objective/action 1000 characters, statements 2000, spans 1000. Production remains: 10 new evaluations/IP/hour, 100/IP/day, 5,000 globally/day; 60 POST attempts/IP/hour, 300/IP/day, 20,000 globally/day; 60 reads/mutations/IP/hour, 300/IP/day. Development only: 60 new evaluations/IP/hour, 600/IP/day; 600 attempts and reads/IP/hour, 3,000/IP/day; global caps unchanged. Current policy: production. Limits are database-backed, with UTC fixed windows and conservative Retry-After. Valid idempotent replay does not use a new evaluation slot. Reasoning tests run in-process without consuming HTTP quota.
Semantic assessment uses the pinned local NLI model identified in /v1/status; no model-provider call is made for an evaluation. The server warms this model before accepting requests; warmup failure leaves status available but evaluations fail closed. Deterministic safety vetoes remain authoritative over a candidate relationship. If the model is unavailable or the 10-second outer evaluation deadline is exceeded, the evaluation fails closed; there is no caller-label fallback. The service admits one evaluation per process, has a 1-second synchronous kernel ceiling, 2-second database statements, and a 10-second save transaction plus 3-second connection deadline. No external execution calls occur. Delivery within 60 seconds is a design target, not a guarantee. Retention remains 24 hours by default, selectable 1–168 hours; expired records are inaccessible. Physical cleanup is bounded and request-driven, not immediate erasure. DELETE removes the saved decision early. Journey/abuse cleanup thresholds remain 7 days/48 hours, not deletion guarantees.
BASE_URL="https://acclaimly.io"
# No pre-existing credentials or account. Generate a private random submission key.
KEY="$(node -p 'require("node:crypto").randomUUID()')"
RESULT="$(curl -fsS "$BASE_URL/v1/evaluations" -H 'Content-Type: application/json' -H "Idempotency-Key: $KEY" --data-binary @- <<'JSON'
{
"synthetic": true,
"objective": "Check whether fictional request SYN-REQUEST-A can advance",
"proposed_action": {
"description": "Rely on fictional approval",
"intent": "rely_on_claim",
"track_id": "SYN-TRACK-A",
"required_proposition_ids": [
"SYN-PROP-APPROVED"
]
},
"propositions": [
{
"id": "SYN-PROP-APPROVED",
"kind": "general",
"statement": "Fictional approval recorded"
}
],
"evidence": [
{
"id": "E901",
"proposition_id": "SYN-PROP-APPROVED",
"relation": "SUPPORTS",
"source_id": "SYN-SOURCE-A",
"source_type": "observation",
"source_locator": "synthetic://fixture/approval",
"statement": "Fictional approval recorded",
"source_span": "Fictional approval recorded",
"track_id": "SYN-TRACK-A"
}
]
}
JSON
)"
printf '%s\n' "$RESULT"
ID="$(printf '%s' "$RESULT" | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>process.stdout.write(JSON.parse(s).evaluation_id))')"
TOKEN="$(printf '%s' "$RESULT" | node -e 'let s="";process.stdin.on("data",d=>s+=d).on("end",()=>process.stdout.write(JSON.parse(s).retrieval_token))')"
# Retrieve twice. Do not publish KEY or TOKEN, or put either in a URL.
curl -fsS "$BASE_URL/v1/evaluations/$ID" -H "Authorization: Bearer $TOKEN"
printf '\n'
curl -fsS "$BASE_URL/v1/evaluations/$ID" -H "Authorization: Bearer $TOKEN"
printf '\n'
# Export the retention-bound, versioned receipt without placing TOKEN in a URL.
curl -fsS "$BASE_URL/v1/evaluations/$ID/receipt" -H "Authorization: Bearer $TOKEN" -o "acclaimly-receipt-$ID.json"
# Optional early deletion:
# curl -fsS -X DELETE "$BASE_URL/v1/evaluations/$ID" -H "Authorization: Bearer $TOKEN"
Invalid requests return 422 with a safe summary, affected field and up to 20 issues (field, code, reason). Evidence-specific issues also name the evidence_id when it passes technical-ID syntax and privacy checks; invalid or private-looking IDs are not echoed. Missing Idempotency-Key uses code missing_idempotency_key; an unmatched source_span uses source_span_mismatch and identifies evidence[index].source_span. Limits name the actual ceiling: for example, "propositions: Maximum 12 propositions." Idempotency-Key errors are reported separately. Submitted narrative values and unknown field names are not echoed.
The free no-signup preview remains free. The owner-approved API / independent-judgment starter credit pack costs $29.00 USD once: 100 credits, expiring after 30 days, one newly persisted synthetic evaluation per credit. POST /api/billing/credits/checkout creates a Stripe Checkout Session only after the exact price is verified and payment gates are enabled; GET /api/billing/credits/offer reports readiness. After verified payment, the buyer can view the account balance at GET /api/billing/credits and create a scoped key at POST /api/billing/credits/api-key. A newly persisted POST /v1/evaluations with that key consumes one credit; an exact idempotent replay consumes zero. No credit is charged for invalid or failed evaluations. Exhaustion returns HTTP 402 with code out_of_credits. Required settings are STRIPE_CREDIT_PACK_PRICE_ID, STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET and ADMIN_API_TOKEN; checkout fails closed if any are missing or configured terms conflict with the approved offer. Approval of terms is not proof of payment readiness. The raw API key is shown once; only its hash is stored. No API key is needed for the free preview. Credits do not authorize external execution.
Successful POST and GET responses include developer_interest with the question, status and submit_url. “Would you use this in an application? What would you evaluate, and how often?” is optional and dismissible. No account or email is required. Interest only: it promises no access. The separate owner-approved starter offer is $29 USD once for 100 credits with 30-day expiry. Checkout remains disabled until configuration, exact Stripe price identity and test fulfillment are verified. A scoped API key consumes one credit only for each newly persisted evaluation; identical idempotent replays do not debit. Anonymous free evaluations never require a credit key and remain free. Evaluation remains synthetic-only and advisory; no source authentication or external-action authorization is provided. This is separate from the content-free journey signal.
To submit, POST {"action":"submit","response":"Check fictional release evidence a few times per week."} to /v1/evaluations/{id}/developer-interest with Content-Type: application/json and the same Authorization: Bearer capability used for retrieval. To dismiss, POST {"action":"dismiss"} without response text. Both use the POST-attempt budget and read/mutation budget. Limit: 1,000 characters and 4 KiB JSON. Exact retries return the same timestamp; a different second response gets 409. No public response list exists.
The optional use-case text is deliberately stored separately with decision_id, created_at and traffic. Current anonymous evaluation traffic is labeled synthetic by the server; clients cannot label it real. Do not include names, contact details, addresses, account identifiers, credentials, private case data or PHI. Common contact/identifier patterns are rejected without echoing input; this is not a comprehensive personal-data detector. Text is not added to logs, analytics, evaluation findings or journey telemetry. Interest expires with the evaluation and is cascade-deleted when that decision is deleted or cleaned up; deletion remains request-driven, not guaranteed at the expiry instant.
One bounded journey record contains journey_id UUID, product, build_sha, evaluation_id when delivered, terminal_state (artifact_delivered, validation_error, timeout, save_failed), latency_ms, and useful/not_useful plus an optional allowlisted reason tag. No input narrative, IP, credentials, PII or PHI enters telemetry. Security rate counters are separate salted digests. POST /v1/evaluations/{id}/usability records this UI signal against its journey using the evaluation capability; it is not the future feedback resource.
The immutable evaluation UUID is the future attachment key for outcomes/feedback. Neither POST /v1/outcomes nor POST /v1/feedback is implemented. WDIS is the designated first consumer after this standalone proof; no other product integrations are included.
Evaluation submissions and optional feedback are not used as training data. This evaluator makes no model-provider calls. The usefulness question is optional and dismissible; its answer sends only journey_id and useful/not_useful with the protected evaluation request, never submitted evidence. Dismissing it sends nothing. Existing bounded journey metadata remains separate from submitted content.
Exact OpenAPI request/response contract · Build identity and ceilings