Release-candidate documentation. Check the deployed revision before integrating. New pilot routes remain gated and paid execution is off by default.

GriffynX Trust · Private pilot

Inspect the request.
Understand the evidence.

A structured API for evidence-led onboarding checks, with a readable report, explicit coverage and the reasons an answer remains unresolved.

Open the gated playground Request pilot access

Evidence before a decision

Source statements, supported attribution and analytical conclusions have different meanings. A report keeps those distinctions visible. High, Medium or Low is shown only when its named evidence policy is satisfied; otherwise the result is Not assessed.

The /v1 fields below are a compatibility reference. A legacy score or interval is not a calibrated probability or the new report policy.

Private access, bounded execution

Approved existing tenant keys must be explicitly allowed into the pilot. Preview validates the values you enter and performs no research. Live requests need a separate server switch, an authorized purpose and positive configured per-key/day and aggregate spending limits. Browser controls cannot enable paid calls.

The playground keeps your key only in page memory. It never puts provider credentials in the browser or puts your key in a URL.

Choose the right product

Discover offers the full Individual report flow. Company quick checks are gated; deeper due diligence is on request. Trace is Coming soon.

Authentication

A bearer token on every request. Once you have an account, you mint and revoke your own keys through the control plane. Each is shown once and stored only as a salted hash: we cannot recover a key for you, and neither can anyone who reaches our database.

# Every request
curl https://griffynx-api.wonderfulisland-732a0a9b.centralindia.azurecontainerapps.io/v1/inquiries \
  -H "Authorization: Bearer gx_live_..." \
  -H "Content-Type: application/json"

Scopes are separate on purpose. A key that opens inquiries cannot mint another key, publish a policy or read the audit trail. Those need an admin key, which should not be the one deployed to your application servers.

Test keys carry the gx_test_ prefix and reach the same surface against non-production data.

Your first call

This is the whole integration for most teams. One request, one response, and the two fields that matter are the score and the interval around it.

# Request
POST /v1/inquiries
{
  "subject": {
    "presentations": [
      { "type": "EMAIL", "value": "applicant@example.com" },
      { "type": "PHONE", "value": "+919876543210" }
    ]
  },
  "intent":  "underwriting_screen",
  "purpose": "credit_assessment",
  "stake": {
    "kind": "credit",
    "amount": 250000,
    "currency": "INR",
    "reversible": false,
    "deadline_ms": 3000
  }
}
# Response
{
  "inquiry_id": "6f0a6d58-...",
  "status": "concluded",
  "assessment": {
    "score": 62,
    "low":   51,      // the interval is not decoration
    "high":  73,
    "preliminary": false,
    "methodology_version": "m1-graph",
    "label": null,   // your policy fills this, not us
    "evidence": {
      "signals": 14,
      "failed_signals": 2,
      "contradictions": 0,
      "by_trust_band": { "high": 6, "medium": 7, "low": 1 }
    }
  }
}

Read the interval, not just the score. 62 with an interval of 51–73 is a usable answer. 62 with an interval of 15–85 is the system telling you it does not know, and treating the two the same way is the most expensive mistake an integration can make. The width is a function of how much evidence answered, and the response tells you that too.

Core concepts

Five things you declare or receive. Getting these right is most of a good integration.

ConceptWhat it isWhy it exists
intent What you are about to do: onboard, underwrite, authorise, expand a case. Sets how hard the engine works and when it is allowed to stop.
purpose The lawful basis for asking. Mandatory on every call, without exception. It is written into the audit trail. When a regulator asks why you held this data, the answer is a record rather than a recollection.
stake What is at risk, how much, whether it is reversible, and how long you can wait. Fixes how narrow the interval has to be before an answer is worth acting on. A reversible listing tolerates doubt a wire transfer does not.
assessment Score, interval, methodology version, as-of, and an evidence summary. What the evidence supports. Never a decision, never a recommendation.
label Your own band applied to that score, under a policy version you published. Your risk appetite is not ours to guess. Changing a label changes an outcome; changing a score would change history.

By domain

Seven industries, one surface. Each section below is the same API called with a different declaration, which is the whole design: what differs between a lender and an investigator is the question being asked, not the evidence or the endpoint.

Why there is no /v1/banking. A vertical endpoint would mean a vertical model, and a vertical model cannot see the pattern that crosses them. The device applying for credit this morning listed on a marketplace last week; the wallet receiving funds tonight was named in a dating scam in March. One surface over one graph is what makes those the same question.

Banking and lending

Whether to open an account or extend credit to somebody whose file is thin.

What gets missed today

A bureau tells you about a person's borrowing. It cannot tell you that the email address on the application is three weeks old, that the device has applied under four other names this month, or that the phone number was issued after the employment history it claims to support. None of that is on a credit file, and all of it is knowable before you decide.

What we add

Correlation across identifiers the applicant supplied separately. A clean record on each one and a damning pattern between them is the case this exists for.

Send outcomes. Repaid, defaulted, reversed: it is the only way an assessment learns your book rather than the average book.

Payments and fintech

Whether to authorise a transfer to a payee seen for the first time, inside the authorisation window.

What gets missed today

Rules catch what they were written for. A first-time payee at an unusual hour is either a fraud or a Tuesday, and the difference is not in the transaction. It is in what the payee's identifiers have been doing elsewhere.

What we add

A preliminary answer inside your deadline, refined afterwards rather than withheld. You get something usable in the window and something better after it.

Built for India first. UPI identifiers and DLT-registered SMS headers are first-class artifact types, not a generic string field with a note attached.

Crypto and digital assets

Whether to accept a deposit from, or release funds to, a wallet address.

What gets missed today

An address is either sanctioned or it is not, and every exchange checks that. The interesting question comes before it: has this address, or the cluster around it, been named in scam reporting, drained a contract, or appeared beside identifiers you have already refused.

What we add

Sanctions screening and community scam reporting on the same call, with the address treated as an artifact that can correlate to an email, a handle or a device rather than as a string in isolation.

Ten tools on wallet addresses, four of them at no cost, including sanctions lists and scam-address reporting.

Dating and social platforms

Whether a new profile is a person, and whether it is the person it says.

What gets missed today

Romance fraud does not fail a document check, because the documents are real and belong to somebody. What gives it away is the shape of the account: a handle claimed across nine platforms in a week, photographs that predate the profile, a phone number in one country and a device in another.

What we add

The broadest coverage we have. Eighteen tools on social handles, because this is where identity is thinnest and the correlation between an account, its age and its other appearances does the most work.

The assessment is about the account, not the human being behind it. We report what public evidence supports, and we do not infer protected characteristics.

Marketplaces and platforms

Whether to let a seller list, a driver drive, or a host host.

What gets missed today

Everything on the application is true. The seller is real, the address exists, the documents check out. What you cannot see is that the identifiers reuse a network already behind three suspended accounts, because that is a fact about your own platform's history joined to public evidence.

What we add

Your suspensions are testimony. Send them as observations and they weigh on future assessments, which no external vendor can do for you, because they do not have them.

Monitors matter here. An account that was fine at onboarding and is not fine now is the common case, and asking again on a schedule is the expensive way to find out.

Investigations

What happened, who was involved, and what you can put in front of a regulator or a court.

What gets missed today

The subject is a network, not a person. Several analysts touch it, the scope grows as it goes, and at the end somebody has to show their working to somebody who was not there.

What we add

Cases carry an authorisation record, a budget, a legal hold and an append-only event log. Every expansion is one deliberate, costed hop rather than an open-ended crawl, and the export is an evidentiary bundle rather than a screenshot.

Requires an investigative tenant. Every case needs an authorisation on the record before it opens, which is a constraint we are not willing to make optional.

Compliance and AML

Whether you can show, later, that you looked and what you found.

What gets missed today

Screening tools answer the question you asked on the day you asked it. What they rarely produce is the thing an examiner actually wants: which rule ran, on which version, against which evidence, and why the answer was what it was at that moment rather than today.

What we add

Every call writes one audit row carrying the purpose you declared, and every assessment names the methodology version that produced it. Your risk bands are a policy you publish and version here, so a label from March can be reconstructed in October without anybody relying on memory.

Scores never change retroactively. A published assessment is a fact about what the evidence supported at a moment; when your appetite changes you publish a new policy and the label changes, which is a different thing and is recorded as one.

Reference

Generated from the named source contract. This reference does not certify what is deployed. Endpoints marked contract have a reference shapes; availability depends on the deployed build and tenant. They are documented because integrating against a shape that will not change is worth more than a surprise later.

Trust private pilot

Release-candidate request/response playground. Approved tenant keys and server switches are required; paid execution is disabled by default.

2 endpoints · /v2/trust

POST /v2/trust/checks release candidate · gated

Trust Check

Validate, reserve and run one canonical check; live mode is separately disabled.

200 TrustResponse
GET /v2/trust/schema release candidate · gated

Trust Schema

Describe the actual implemented pilot request body after authorization.

200 no body

Report core

Authenticated report projection for a completed inquiry. Legacy evidence does not inherit canonical attribution or source-lineage guarantees.

1 endpoint · /v2/reports

GET /v2/reports/{inquiry_id} release candidate · gated

Report Core

Expose report-owned meaning for an already completed authorized inquiry.

200 no body
Parameters
inquiry_idpathrequired
profilequeryrequired
formatqueryoptional

Inquiries

Ask about a subject and get an assessment. This is the endpoint most integrations use and, for many, the only one they ever need.

3 endpoints · /v1/inquiries

POST /v1/inquiries live

Open Inquiry

Open an inquiry (command OpenInquiry). mode=sync waits for the answer within stake.deadline_ms. mode=async returns 202 the moment the inquiry is admitted and runs it in the background, for a check too deep to hold a connection for (S52): poll GET /v1/inquiries/{id} for reasoning, then the answer, which is kept a day. Every refusal (a purpose, a limit, an unreadable material) still comes back at once, exactly as it does for sync. stream is not served and is refused here rather than silently treated as sync. Rejected with I13 when purpose is missing or not allowed by policy. Every outcome, including a refusal, writes one audit row; its id comes back on the X-Griffynx-Audit-Id header so a caller can cite the call they made.

body InquiryRequest200 InquiryResponse
GET /v1/inquiries/{inquiry_id} reference

Get Inquiry

The inquiry as it stands: reasoning while it runs, then its answer, kept a day (S52). An inquiry that is not yours, does not exist or is more than a day old is not found, the same answer for all three so this cannot be used to learn whether somebody else's exists (I19).

200 InquiryResponse
Parameters
inquiry_idpathrequired
GET /v1/inquiries/{inquiry_id}/events reference

Inquiry Events

SSE: AssessmentIssued, AssessmentRevised, InquiryConcluded. Replayable with after.

200 no body
Parameters
inquiry_idpathrequired
afterqueryoptional

Observations

Your own testimony about a subject. What you know that nobody else does enters here, carrying its provenance, and is weighed accordingly.

2 endpoints · /v1/observations

POST /v1/observations reference

Submit Observation

Submit testimony (command SubmitObservation). Enters with its provenance class; never sets a score.

body ObservationIn201 ObservationOut
POST /v1/observations/batch reference

Submit Observations

Submit many observations at once.

body ObservationIn[]201 ObservationOut[]

Outcomes

What actually happened after you decided. The single most valuable thing you can send us, and the only way an assessment gets better at your book specifically.

2 endpoints · /v1/outcomes

POST /v1/outcomes reference

Submit Outcome

Report what happened after a decision (command SubmitOutcome). Feeds calibration (E14).

body OutcomeIn201 ObservationOut
POST /v1/outcomes/batch reference

Submit Outcomes

Report many outcomes at once (a month of repayments, a day of chargebacks).

body OutcomeIn[]201 ObservationOut[]

Material

Documents, images, audio and video, declared before they are sent. Extraction produces bindings; the bytes themselves can be processed and never stored.

5 endpoints · /v1/material

POST /v1/material reference

Create Material

Declare material (command IngestMaterial). A text material carries its text and is held at once; its id is the sha256 of the text, which is what an inquiry names in material_ids. Anything else declares the sha256 of the bytes that will follow and is told where to put them (PUT the bytes to upload_url). retention_mode is a promise (D35): the default, transient, is released as soon as the inquiry that reads it concludes. Only text is read today; an image or a document is stored and an inquiry that binds one is refused with that reason. C2 material is refused.

body MaterialCreate201 MaterialOut
GET /v1/material/{material_id} reference

Get Material

The material's status. A material that is not yours is not found.

200 MaterialOut
Parameters
material_idpathrequired
DELETE /v1/material/{material_id} reference

Delete Material

Let go of the content now. The record that it was held, and for what, stays.

204 no body
Parameters
material_idpathrequired
PUT /v1/material/{material_id}/content reference

Put Material Content

The bytes of a declared material. They must hash to material_id.

200 MaterialOut
Parameters
material_idpathrequired
GET /v1/material/{material_id}/extraction reference

Get Extraction

What is stored about the material's extraction; what a message named is in the inquiry's findings.

200 ExtractionOut
Parameters
material_idpathrequired

Cases

An investigation with more than one subject, more than one analyst, and a record of who did what. For investigative tenants; every case carries an authorisation.

14 endpoints · /v1/cases

POST /v1/cases reference

Open Case

Open a case (command OpenCase). Requires an investigative tenant (KYB) and an authorization record (I13).

body CaseCreate201 CaseOut
GET /v1/cases/{case_id} reference

Get Case

The case and its current state, derived from its event log.

200 CaseOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/assertions reference

Assert In Case

An analyst's judgement as evidence (command AssertInCase, E9).

body CaseAssertionIn201 ObservationOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/close reference

Close Case

Close the case (command CloseCase).

200 CaseOut
Parameters
case_idpathrequired
findingqueryoptional
GET /v1/cases/{case_id}/events reference

Case Events

SSE: CaseChanged, AssessmentRevised for expansions, MonitorTriggered for case monitors.

200 no body
Parameters
case_idpathrequired
afterqueryoptional
POST /v1/cases/{case_id}/expand reference

Expand Case

One hop of attribution (command ExpandCase), budgeted. With human_gate the candidates are shown first and nothing is spent until one is chosen.

body CaseExpandIn202 InquiryResponse
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/export reference

Export Case

Produce the evidentiary bundle (command ExportCase). Requires sign-off.

202 CaseExportOut
Parameters
case_idpathrequired
formatqueryoptional
GET /v1/cases/{case_id}/graph reference

Case Graph

The case graph: nodes and edges with types, weights and provenance classes.

200 CaseGraphOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/hold reference

Hold Case

Place legal hold (command HoldCase).

body CaseHoldIn200 CaseOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/intake reference

Intake Case

Bring many things into a case at once (command IntakeBatch, D43). A bounded manifest of presentations and declared material; bytes follow by hash. Each item answers for itself (accepted, duplicate, rejected); the manifest becomes one case event. Nothing is spent: gathering is expand. For open-ended intake open a session with case_id instead.

body IntakeManifest201 IntakeOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/members reference

Add Case Members

Add members (command AddCaseMembers).

body CaseMemberIn200 CaseOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/release reference

Release Case

Release legal hold (command ReleaseCase).

body CaseHoldIn200 CaseOut
Parameters
case_idpathrequired
POST /v1/cases/{case_id}/review reference

Review Case

Move the review state (command ReviewCase).

body CaseReviewIn200 CaseOut
Parameters
case_idpathrequired
GET /v1/cases/{case_id}/timeline reference

Case Timeline

Everything that happened in the case, in observed order.

200 EventEnvelope[]
Parameters
case_idpathrequired

Monitors

A standing watch on a subject. You are told when the answer changes, rather than asking again on a schedule.

4 endpoints · /v1/monitors

POST /v1/monitors reference

Open Monitor

A standing watch (command OpenMonitor).

body MonitorCreate201 MonitorOut
GET /v1/monitors/{monitor_id} reference

Get Monitor

The monitor and its last assessment.

200 MonitorOut
Parameters
monitor_idpathrequired
DELETE /v1/monitors/{monitor_id} reference

Close Monitor

End the watch (command CloseMonitor).

204 no body
Parameters
monitor_idpathrequired
GET /v1/monitors/{monitor_id}/events reference

Monitor Events

SSE: MonitorTriggered.

200 no body
Parameters
monitor_idpathrequired
afterqueryoptional

Sessions

A live connection for open-ended intake: frames in, events out, over a WebSocket.

2 endpoints · /v1/sessions

POST /v1/sessions reference

Open Session

Open a live session (command OpenSession); returns the WebSocket URL for frames and events.

body SessionOpen201 SessionOut
POST /v1/sessions/{session_id}/close reference

Close Session

Close a session (command CloseSession); the retention mode is applied.

202 no body
Parameters
session_idpathrequired
reasonqueryoptional

Entities

A subject we have resolved across more than one identifier, and its assessments in order, which is reputation over time.

2 endpoints · /v1/entities

GET /v1/entities/{entity_id} reference

Get Entity

A resolved subject as the tenant may see it.

200 EntityOut
Parameters
entity_idpathrequired
GET /v1/entities/{entity_id}/ledger reference

Get Ledger

The entity's assessments in order: reputation over time.

200 LedgerEntryOut[]
Parameters
entity_idpathrequired

Registry

What your tenant may present, ask and declare. Read it at startup rather than hard-coding a list that will drift.

3 endpoints · /v1/registry

GET /v1/registry/artifact-types reference

Artifact Types

Artifact types this tenant may present, from the loaded packs, with the platform list for a handle.

200 RegistryEntryOut[]
GET /v1/registry/intents reference

Intents

Intents this tenant may use.

200 RegistryEntryOut[]
GET /v1/registry/purposes reference

Purposes

Purposes this tenant's policy allows.

200 RegistryEntryOut[]

Control plane

Keys, policy, webhooks and your audit trail. Separate scope; a key that can open inquiries cannot mint another key.

8 endpoints · /v1/admin

GET /v1/admin/audit reference

Audit

The tenant's audit trail: every call and analyst action with its cost.

200 AuditEntryOut[]
Parameters
afterqueryoptional
limitqueryoptional
POST /v1/admin/keys reference

Mint Key

Mint a key (command MintKey): shown once, stored hashed (I11).

body KeyCreate201 KeyOut
GET /v1/admin/policies reference

List Policies

Every policy version this tenant has published, current first.

200 PolicyOut[]
POST /v1/admin/policies reference

Publish Policy

Publish a policy version (command PublishPolicy). Labels change; scores never do. Empty bands means the loaded packs' default bands. The new version is the tenant's current one; the versions before it stay, because an assessment already labelled names the version that labelled it.

body PolicyIn201 PolicyOut
POST /v1/admin/trials reference

Provision Trial

Provision an X-Ray trial for one address and return its key, once. The only route that mints a credential for somebody who is not in the room, so three things bound it and all three are here rather than in the caller. It needs the trials scope, which one key in one deployment holds. It takes one field, so nothing about the trial's shape can be asked for. And every value it writes — tier, budget, limits, purposes, lifetime — is a constant in control.trials, which means a stolen provisioning key can create trials and cannot create anything better than a trial. Whoever may hold a trial is decided before this is called, by the whitelist in the surface that calls it. This route deliberately does not know that list: it is a policy about people, it changes often, and putting it here would make every change to it a deploy of the gateway.

body TrialCreate201 TrialOut
POST /v1/admin/webhooks reference

Register Webhook

Register a webhook (command RegisterWebhook).

body WebhookIn201 WebhookOut
DELETE /v1/admin/keys/{key_id} reference

Revoke Key

Revoke a key (command RevokeKey).

204 no body
Parameters
key_idpathrequired
GET /v1/admin/trials/{tenant_id}/usage reference

Trial Usage

What one trial has left: calls today, budget spent, days remaining. Read by the playground to draw its own usage strip, so the numbers a prospect sees are the numbers the gateway is actually enforcing rather than a second count kept by the page and drifting from it.

200 TrialUsageOut
Parameters
tenant_idpathrequired

Health

Liveness, for your monitoring. `/healthz` additionally reports the running build, the tool count and whether the service can reach its own database.

2 endpoints · /v1/health

GET /healthz live

Healthz

Liveness and the two things worth knowing before a deploy is believed. Open, like the provider service's own health endpoint: it reveals how many tools were discovered and whether the gateway's schema answers, and nothing about any tenant.

200 no body
GET /v1/health live

Health

Liveness. The only endpoint with a body today.

200 no body

Objects

The shapes you send and receive. Every one is closed: an unrecognised field is an error rather than something silently ignored.

InquiryRequest Ask about a subject at a stake, for a purpose, with an intent
subjectrequiredSubject
intentrequiredstringA registered intent id, e.g. pre_transaction, onboarding_screen.
purposerequiredstringA registered purpose id. Required on every call (I13).
stakerequiredStake
fraud_intakeFraudIntakeV1
contextObservationIn[]The caller's own testimony, attached inline: at most 20, each about something this inquiry presents, each under this inquiry's purpose, and tenant_supplied or self_attested (with a consent reference). It is written into the inquiry's evidence as the caller's word and never lends a tool's trust. A subject's name is carried as an observation whose summary is Name: <text>.
material_idsstring[]Bound supported text material; images/PDF/OCR are not readable evidence. Binding unreadable or mixed text/non-text materials is refused. Declaration/storage alone does not run a check.
consent_refstringThe caller's own record of the reader's consent, for example the id of the consent event. Required for a purpose that enumerates someone's own accounts (self_exposure_audit); otherwise optional. It is an opaque reference, never the consent text.
tier_ceilingstringHighest provider tier this inquiry may use.
render"none" | "template" | "model"Whether to write the findings up as a report. none is the default because a caller that only wants the numbers should not pay for prose. A tenant may be clamped below what it asks for; the reply says which renderer ran.
render_langstringLanguage tag for the report, e.g. en, hi.
mode"sync" | "async" | "stream"
idempotency_keystring
Subject What the inquiry is about: presentations, an entity we already know, or a claim
presentationsPresentation[]
entity_idstringA subject we have resolved before.
claimClaim
Presentation One presented identifier: a registered artifact type and its value
typerequiredstringA registered artifact type id (see GET /v1/registry/artifact-types).
valuerequiredstringThe presented value; normalised by the kernel.
Stake What is at risk and how much time there is; fixes the tolerable interval and the budget
kindrequired"payment" | "credit" | "account" | "listing" | "hire" | "belief" | "other"
amountnumber
currencystring
reversiblebooleanCan the truster undo the act after the fact?
deadline_msintegerTime the caller can wait for a first answer.
Assessment The public projection of a kernel Assessment
assessment_idrequiredstring
inquiry_idrequiredstring
scorerequiredinteger
lowrequiredinteger
highrequiredinteger
preliminaryrequiredbooleanTrue when the stopping rule was not yet met (pre_transaction only).
as_ofrequiredstring
methodology_versionrequiredstring
labelrequiredLabel
evidencerequiredEvidenceSummary
entity_idrequiredstring
EvidenceSummary What the tenant may see of the chain of evidence: counts and bands, never raw or provider names
signalsrequiredinteger
failed_signalsrequiredinteger
contradictionsrequiredinteger
by_provenance_classrequiredobject
by_trust_bandrequiredobject
oldest_observed_atrequiredstring
newest_observed_atrequiredstring
Finding One thing that was established about the subject
finding_idrequiredstring
node_idstringThe node in graph this fact is about, when the check drew one. Opaque and stable within this response only.
aboutrequiredstringThe caller's own presented value this concerns, when it was theirs.
about_typerequiredstringThe registered artifact type of about.
kindstringWhat sort of check produced this, e.g. account_age, domain_popularity. A renderer uses it to say what a finding of this sort means; it names the question asked, never who was asked.
statementrequiredstringWhat was found, in words, with every source name removed.
disposition"supports" | "weakens" | "neutral"Which way this points for somebody deciding whether to trust the subject, as the source itself declared it. Never parsed from statement, never a score, and never an input to the assessment. It is here so a tenant's policy can form a label (D05).
provenance_classrequiredstringHow this came to be known, e.g. third_party, computed.
trust_bandrequired"high" | "medium" | "low"The banded trust of its source.
observed_atrequiredstring
data_classrequiredstringC0, C1 or C2. What handling this fact requires.
attributesobjectWhat the source said, as structure, when it said something that has one. For a statement a public page makes: predicate (role, affiliation, founded, located, event, ownership, dispute, description, public_statement, review), quote (the page's own words, verbatim), source_url and page_date. For a profile: the platform, name, bio, location and when the account was created. Absent for most findings. It is what was said, never what it is worth.
Connection An edge the engine drew between two things, and why it drew it
connection_idrequiredstringStable id for this edge, so a report can cite it.
kindrequired"declared" | "resolved" | "contradiction" | "temporal" | "shared" | "similar" | "asserted"How the edge came to exist, so a reader can weigh it: declared (a source reported it), resolved (evidence says these are one subject), contradiction (two sources disagree), temporal (a date a source asserted), shared (both reach the same third thing), similar (their values are close), asserted (a person said so).
relationrequiredstringThe registered relationship type, e.g. shares_registrant.
from_labelrequiredstringThe caller's own value where it was theirs, else its type.
from_typerequiredstring
to_labelrequiredstring
to_typerequiredstring
weightrequirednumberThe kernel's computed strength for this edge.
reasonrequiredstringWhy the edge holds, with every source name removed.
Report One inquiry, written for the person who asked
inquiry_idrequiredstring
assessment_idrequiredstring
langstring
rendered_byrequired"template" | "model"Which renderer produced this. A consumer should not otherwise be able to tell.
prompt_versionstringThe prompt that wrote it, when a model did. Null for a template.
depthReportDepthHow deep the report went and what a deeper one adds. Set on the individual report.
sectionsrequiredReportSection[]
template_sectionsstring[]Sections the template wrote because the model's did not survive its checks (a paragraph citing something it was never given, a section that did not validate, or one it left out). Empty for a report the model wrote whole and for a template report.
sharerequiredShareCardOut
generated_atrequiredstring
ObservationIn Testimony the caller supplies about a subject, entering with its provenance class
subjectrequiredSubject
provenance_classrequired"tenant_supplied" | "self_attested" | "analyst_asserted"
summaryrequiredstring
observed_atrequiredstring
observed_range_startstring
observed_range_endstring
purposerequiredstring
data_class"C0" | "C1" | "C2"
consent_refstringRequired for self_attested.
case_idstringRequired for analyst_asserted.
OutcomeIn What actually happened after a decision
assessment_idrequiredstring
decision_refrequiredstringThe tenant's own id for the decision they took.
kindrequired"repaid" | "defaulted" | "reversed" | "settled" | "confirmed" | "exposed" | "closed_finding" | "other"
observed_atrequiredstring
detailstring

Errors

A refusal is an answer, and it names the rule that produced it. Every call, refused or not, writes one audit row whose id comes back on X-Griffynx-Audit-Id.

StatusMeansWhat to do
200Concluded, with an assessment.Read the interval.
401No key, or a revoked one.Check the bearer token.
422Refused. The body names the rule. Do not retry unchanged. A refusal for purpose or intent means your tenant is not permitted to ask that, and asking again will be refused identically.
501A published contract with no handler yet. The shape is fixed; build against it and it will start answering.
502The engine could not complete. Retry once. The body carries the audit id for the attempt.

Versioning and deployment status

The reference is generated from a named source contract. It is not deployment evidence. Read /healthz for the revision and its JSON status: HTTP 200 alone can still mean degraded.

Collection wire 2.0.0 remains compatible. The additive semantic vocabulary is 1.0.0 in SDK 2.1.0. Kernel a25 and report-core 1.0.0rc2 form the candidate core. Historical /v1 methodology remains separately versioned.

Unknown source rights, unsupported capabilities and pending independent human acceptance remain explicit release gates.