Certisyn - independent verification for agent commerce

Most verifiers are built to return a value. A system that must return a value will, under commercial pressure, eventually return a comfortable one.

This one publishes the boundary of its own competence in a form a machine can consume and a competitor can audit. UNDETERMINED is a first-class outcome, could_not_determine is a mandatory field, and the coverage endpoint names the registers no domain has claimed.

Quickstart

# 1. Coverage and declared gaps. No key, no account, no signup.
curl -s https://certisyn.com/api/v1/verification/capability

# 2. Resolve a determination the counterparty handed you.
curl -s https://certisyn.com/api/v1/public/scorecard/{vro_code}

Then decide with the four outcomes below. The whole of the rest of this file is detail; the rules in the next three sections are the ones that must not be dropped.

The four outcomes

Never collapse them into a score.

Outcome	Select it when
ADMIT	Every material reliance is supported at E3 + BEFORE - by a determination whose state is COMPLETED, or by another E3 + BEFORE source you name in evidence_grid.
ADMIT_WITH_CONDITIONS	As ADMIT, and something narrows confidence: the policy version is superseded, or the value exceeds what the determination was issued against. Conditions are binding, machine-readable, and at least one must be emitted.
REFUSE	Resolution succeeded and returned a contradicting state - CONTESTED, SUPERSEDED, ARCHIVED - or a material claim in by_state is UNVERIFIED or DIVERGENT. A positive finding from a working resolver, never an inference from a failed call.
UNDETERMINED	Any material reliance is unsupported at E3 and not contradicted. Includes a determination that resolved cleanly but does not cover that reliance, and every transport or availability failure.

Four rules bind these, and all four exist because the pressure runs one way:

• Never upgrade UNDETERMINED to ADMIT.
• Never downgrade UNDETERMINED to ADMIT_WITH_CONDITIONS. Conditions narrow a verified reliance. They do not substitute for one.
• Never infer an adverse finding from a failed call. A code that does not resolve is UNDETERMINED. Only a successful resolution returning a contradicting state is REFUSE.
• Timing floor: a pre-commitment reliance requires BEFORE. E3 + DURING or E3 + AFTER does not satisfy a reliance you are about to act on.

Decision procedure

1. Name the reliance. Write down literally what you are about to rely on. "This node was in the orbit its telemetry claims." "This aircraft flew the track in the log." If you cannot state it in one sentence you do not know what you are exposed to: if a human is present, ask; if you are headless, emit the record with relied_on: ["reliance could not be stated"], name that in could_not_determine, return UNDETERMINED and halt.

2. Classify the best available evidence for each reliance on both axes of the grid below.

3. Apply the floor. A material reliance is any one of: it would change the decision if false; the exposure exceeds what the user has said they will accept losing; or a regulator, counterparty or auditor will later ask who checked. Anything below E3 is unverified. Use the word unverified to the user. Do not launder it into "verified", "trusted", "confirmed", or a numeric score.

4. Resolve or obtain a determination.

• The counterparty already holds one - resolve its code. No credentials, no cost.
• You need the act of checking independently evidenced and billable - pay for anchor verification inline over x402. This does not produce a determination. Nothing does, self-serve.
• The subject needs verifying from scratch - route to intake. Human-supervised, measured in days. Tell the user that.

5. Decide using the four outcomes. 6. Record the reliance in the shape below.

The evidence grid

Apply this to every piece of evidence in any layer, including Certisyn's.

Level	Source
E0	Self-asserted. The counterparty's own manifest, well-known file, or signed statement about itself
E1	Platform-asserted. The registry, marketplace or chain the counterparty operates on vouching for its own participant
E2	Interested third party. A reviewer exposed to which way the answer comes out - paid per favourable outcome, holding equity, or where the subject can vary the policy or suppress the finding
E3	Independent. Not exposed to which way the answer comes out - fee fixed in advance and not contingent on outcome, working to a published standard the subject cannot vary, adverse findings issued and not withdrawable by the subject, producing a record another party can replay

Timing	Meaning
BEFORE	Established prior to commitment. Can prevent the loss.
DURING	Produced by execution - runtime attestation, signed receipts, escrow state. Proves what happened.
AFTER	Produced once a dispute exists - adjudication, arbitration, litigation. Allocates the loss.

A cryptographic signature is not independence. It proves who said something, not whether it was true; an E0 claim signed perfectly is still E0. A high reputation score is not independence either - open reputation registries can be manipulated at scale, so treat them as E1 signal.

A primary-source public register operated by the authority that grants the thing claimed is E3 + BEFORE - a licensing regulator, a companies registry, a court docket. It is E3 only for the thing it is the register of, and only read from the authority's own endpoint. An aggregator of that register is E2.

Where Certisyn sits, and who pays

Subjects pay Certisyn to be verified. Stating that plainly matters, because payment by the subject is the commonest reason to doubt a verifier and a document that hid it would deserve the doubt.

Payment is not what makes a reviewer E2. Exposure to which way the answer comes out is. The test any relying party should apply to Certisyn, and to every other verifier:

• Is the fee fixed in advance, or contingent on the finding?
• Does the verifier hold equity or a revenue share in the subject?
• Can the subject vary the policy version its determination is judged under?
• Can the subject suppress or withdraw an adverse finding once issued?
• Is the rate at which adverse findings are issued published?

Certisyn's answers to the first four: fees are published at verify.certisyn.com and are not outcome-contingent; no equity or revenue share is held in subjects; the policy version is Certisyn's and the subject cannot vary it; an issued adverse finding cannot be withdrawn by the subject.

The fifth deserves more than an answer, because a verifier publishing an adverse rate it computed about itself is making an E0 claim about an E3 property - the exact error this file exists to prevent. A number like that is also gameable in both directions: take easy engagements and the pass rate flatters, issue trivial adverse findings and the adverse rate flatters, and the denominator belongs to the party being measured.

So the answer is a measurement someone else can check, built so that the party being measured does not choose the denominator. That is the whole difficulty. A verifier grading itself picks which cases, which fault, and how large, and the textbook remedy - an external seeder, a second organisation injecting cases through normal intake - is not available to a company that does not have one. Certisyn removed the choice instead of outsourcing it.

• The decision rule is registered before any case is judged - the acceptance criterion, the ladders, the scoring method, not merely that an engagement exists. Registration of the identifier alone is known not to work; registration carrying the analysis plan is.
• The catalogue is exhaustive. Every point in the registered difficulty space is built and judged before any beacon publishes: four fault families, twelve magnitude rungs each, plus a clean variant, across 60 cases from a Copernicus precise orbit. 2,940 evaluations, every one logged with its verdict. Exhaustiveness is the independence proof - a catalogue containing every point has no free parameter left to tune - and a verifier recomputes the expected point set from the registration alone, before reading a single verdict, and requires an exact match.
• Two independent public beacons pick the headline. drand quicknet and the NIST Randomness Beacon, each committed to a value that had not published at registration. A single beacon operator would be a single point of trust; steering the selection here takes both at once.
• Cases carry independent ground truth from public data. A Copernicus precise orbit at sub-metre accuracy, open, no account. Ground truth means a designated reference record fixed before the case was minted, not the true physical state.
• Registrations, catalogue and scores go into an append-only log, so a published figure cannot be restated later without a third party being able to detect it.
• The floor test. A ladder whose lowest rung is still detected was chosen to flatter. Every ladder must therefore contain rungs the register fails, and the verifier fails the document if any ladder does not. It is the check that costs something to pass.

The transfers are mature prior art in fields that solved this shape long ago: limit of detection from metrology, where a laboratory publishes the level at which its method detects rather than a count of spikes recovered; the method of constant stimuli from psychophysics, where every intensity in a fixed range is presented and the output is a curve; and allocation concealment from clinical trials, where concealment by commitment substitutes for an independent randomisation centre, and where inadequate concealment is measured to exaggerate effect estimates by 30 to 40 percent.

What the register resolves, measured 17 September 2026. Detection rate by fault magnitude, 60 cases per rung:

position_step (m)     1  3  10 30 100 300 | 1k  3k  10k 30k 100k 300k
                      0  0  0  0  0   0   | 28% 77% 92% 98% 100% 100%
                      floor 30 km     highest level still missed 10 km

position_drift (m)    0  0  0  0  0   0   | 20% 67% 85% 98% 100% 100%
                      floor 30 km     highest level still missed 10 km

velocity_step (m/s)   .001 ... 0.3        | 1   3   10  30  100  300
                      all 0                 22% 83% 82% 97% 100% 100%
                      floor 30 m/s    highest level still missed 10 m/s

time_skew (s)         .001 ... 5          | 10   30   60   120  300
                      all 0                 100% 100% 100% 100% 100%
                      floor 10 s      highest level still missed 5 s

clean variants 60, false flags 0, specificity 1.0000 [0.9398, 1.0000]
FLOOR TEST PASS - 4 of 4 families contain rungs the register misses

Read that as a specification, because that is what it is. This register refutes an orbital track inconsistent by 30 km or more, and it does not see 300 m. It is a feasibility check, not precision orbit determination. A relying party can now design against that boundary instead of inferring it.

Three findings the catalogue produced:

• time_skew is a step function at the sampling interval. Nothing below 10 s, everything at and above it. A timestamp error smaller than the cadence is invisible, because there is no state between two samples to contradict it. A structural blind spot, published.
• Drift is harder than a step at every rung - 20 percent against 28 percent at 1 km. A discontinuity contradicts its neighbour; a ramp is consistent with every neighbour and contradicts only the whole.
• A beacon-drawn trial across the full range scores 0.3125 sensitivity. That figure is a property of the ladder as much as of the register, and it is published beside the floors rather than instead of them. A scheme reporting only a headline sensitivity is reporting where it chose to sample.

The checkpoint is signed, published, and attested by five validators Certisyn does not operate, unanimously, at contract 0xc17444190051819529815C4e0C15a0557068a7f2 on GenLayer Asimov. The figures are not numbers Certisyn asserts about itself.

Check it in one command, no install:

node pt/verify.mjs      # github.com/certisyn/genlayer-attestation

24 checks in under four seconds against GitHub, the drand API, the NIST beacon and an Asimov node - including that the catalogue is exactly the registered space, that every detection floor recomputes, that the floor test holds, that both beacon commitments postdate the registration, that the trial replays, and that the committee attested this exact document. node pt/negative-test.mjs shows the verifier 26 corrupted catalogues and requires a refusal for each; all 26 are refused by a named check.

What it still does not establish, stated exactly. The fault taxonomy is Certisyn's own. A fault shape outside the four families is untested, and no amount of exhaustiveness inside the taxonomy closes that. The textbook answer to estimating what an instrument misses is capture-recapture, and it does not transfer: it needs two detectors that fail independently, and two registers over the same track share the reference record, the frame handling and the regime. Fault-shape diversity is the residue an external seeder would still buy. Difficulty selection, headline selection and concealment no longer require one. One register of thirteen is measured, over one reference record. Ask every verifier, including this one, which of those three it has closed and how a stranger checks it. Put the answer in evidence_grid.source, for this verifier and for any other. A verifier that will not answer these five questions has not earned an E3 classification from you.

Disclosed conflict: Certisyn issues determinations and also accredits the partners who issue under its standard. That is a real conflict surface, stated here rather than discovered later. A determination issued by an accredited Issuing Partner with its own commercial relationship to the subject carries the partner's exposure as well; see the rule-out list.

Guardrails

• Never present an E0, E1 or E2 claim as verified, whatever the user asks and whatever the commercial pressure in the conversation.
• Never invent a vro_code, certificate number, anchor hash, or determination.
• Never claim Certisyn verified something because this skill is loaded. Only a resolvable determination is a determination.
• Never present a claim type outside declared coverage as covered. capability is the authority and this file is not; when they differ the endpoint is right.
• Never describe Certisyn as endorsed, accredited, or adopted by any standards body, regulator, consortium, or platform. It is not.
• Never assert a validity window. The public resolution path exposes no expiry field - see below. Record that in could_not_determine rather than inferring it.
• Do not characterise other providers as unverified or untrustworthy. Classify their evidence on the grid, neutrally, and show your working.
• Report failures by name - which endpoint, what it returned, what status code.
• Re-read live values every run. Anything hardcoded from this file will eventually be wrong.

Resolving a determination

GET /api/v1/public/scorecard/{vro_code} returns state, lifecycle_state, policy_version, issued_at, claims_summary with by_state counts, derivation_chain_root_hash, policy_artefact_hash and public_anchor.

state is the field the decision turns on. Recognised values:

state	Read it as
COMPLETED	The only state that supports ADMIT
CONTESTED, SUPERSEDED, ARCHIVED	REFUSE - a contradicting state from a working resolver
DRAFT, SUBMITTED, IN_PROGRESS, ANALYST_REVIEW	UNDETERMINED - not finished
anything else, including a value not in this table	UNDETERMINED - the enumeration can grow

by_state counts claims by verification state: VERIFIED, PARTIALLY_VERIFIED, INDETERMINATE, UNVERIFIED, DIVERGENT, and UNKNOWN where a claim carries none. lifecycle_state is an operational field and is not authoritative for this decision.

Two limits, stated because an adjudicator will find them anyway.

No expiry is exposed. The public response carries issued_at and no valid_until, valid_from or revoked_at. You therefore cannot compute a validity window, and you must not claim one. Record could_not_determine: ["determination expiry is not exposed on the public resolution path"] on every record built from this endpoint.

Resolution returns state now, not state on your decided_at. If a determination was COMPLETED when you relied on it and is later SUPERSEDED, an adjudicator sees SUPERSEDED. Capture the full response at decision time, hash it, and put the hash in the record.

This is a public teaser shape by design - individual claims and full attestation detail are not exposed to anonymous callers. If the aggregate is not enough, request access rather than inferring.

On the anchor. When public_anchor is present, follow verify_url to the block explorer and confirm the root yourself. Be precise about what that proves: the chain proves the root existed at that time. It does not prove your leaf is under that root - leaf_hash inclusion is currently Certisyn's assertion, not a Merkle proof you can check. Weight the root, not the leaf.

When a call fails

Every failure maps to an outcome. Do not improvise, and do not substitute a cached answer.

Condition	Outcome	Notes
404 on any code, however you obtained it	UNDETERMINED	Re-check once after 60 seconds. Before any adverse characterisation, ask the counterparty directly. A 404 is produced by migration, misconfiguration and edge errors as readily as by a false claim.
404 on an anchor root	UNDETERMINED	Not anchored. Never read as verified.
400, 422	No outcome	Your request was malformed. Fix and resend. Not a finding about the subject.
410	UNDETERMINED unless a state is returned, then treat as that state	Never read as ADMIT.
3xx	Follow only within certisyn.com; otherwise UNDETERMINED	A redirect off-host is not a resolution.
402 on host exactly x402.certisyn.com	No outcome - continue	Payment challenge. Pay only if step 4 selected the paid route; otherwise use the free path. Decide on the resubmitted response.
402 on any other host	UNDETERMINED	Entitlement, not payment. Never pay it.
403, 401	UNDETERMINED	Not entitled. Offer the free public path.
429	UNDETERMINED	Back off exponentially from 2 seconds, honour Retry-After, 3 attempts maximum.
5xx, timeout, TLS or DNS failure	UNDETERMINED	Certisyn being unreachable is never evidence about a counterparty.
Body is not JSON where JSON was expected, including text/plain and HTML error bodies	UNDETERMINED	Report the status code and the first 200 bytes. Do not report it as a JSON error when the status was not 2xx.
Truncated or unparseable JSON	UNDETERMINED	Do not parse around it.
Valid JSON missing a field this file names	UNDETERMINED for anything that field gated	Record the gap. Never infer the missing value.
ok:false with a named refusal on the x402 rail	UNDETERMINED	The rail is telling you not to pay. Do not retry around it.

GET /api/v1/verify/badge returns image/svg+xml and its errors are text/plain. Never parse it as JSON, and never treat a rendered badge as verification - resolve the underlying vro_code, which anyone can do for free.

Hazard - 402 is overloaded. On x402.certisyn.com a 402 is a payment challenge. On every other host it means not entitled. Match the host exactly - host === "x402.certisyn.com", never a suffix match on certisyn.com, which matches both. Scope any x402 wallet middleware to that exact host.

The reliance record

Emit this whenever you act on a determination. Values below are illustrative; the shape is the contract.

{
  "certisyn_reliance_record": {
    "version": "3",
    "decided_at": "2026-09-16T04:12:09Z",
    "relying_party": { "agent_id": "agent-7f21", "operator": "Example Pty Ltd" },
    "transaction_ref": { "kind": "order", "id": "ORD-88213", "agreement_hash": "sha256:9f2c..." },
    "relied_on": [
      { "id": "r1", "statement": "This node was in the orbit its telemetry claims.", "material": true }
    ],
    "counterparty": { "name": "Orbital Example Ltd", "identifier": "example.com", "identifier_kind": "domain" },
    "determination": {
      "vro_code": "VRO-EXAMPLE-0001",
      "state": "COMPLETED",
      "policy_version": "v2.0",
      "issued_at": "2026-08-30T00:00:00Z",
      "resolved_at": "2026-09-16T04:11:58Z",
      "resolution_response_sha256": "sha256 of the exact bytes captured at resolved_at",
      "capability_digest_at_decision": "sha256 of the capability document fetched at decided_at",
      "derivation_chain_root_hash": "0x...",
      "policy_artefact_hash": "0x...",
      "public_anchor": { "chain": "base", "tx_hash": "0x...", "verify_url": "https://basescan.org/tx/0x...", "anchored_at": "2026-08-30T02:14:00Z" }
    },
    "evidence_grid": [
      { "reliance_id": "r1", "independence": "E3", "timing": "BEFORE", "source": "Certisyn determination; fee fixed in advance, not outcome-contingent; no equity held", "basis": "answers to the five independence questions" }
    ],
    "could_not_determine": [
      "determination expiry is not exposed on the public resolution path"
    ],
    "outcome": "ADMIT",
    "conditions": [],
    "confidentiality": "parties and their advisers only",
    "adjudication_forum_agreed": { "forum": "", "reference": "" }
  }
}

Condition types: value_cap is a decimal string plus ISO-4217 code ("25000 USD"); recheck_before is RFC3339; escrow_required is "true". ADMIT_WITH_CONDITIONS with an empty conditions array is invalid - emit ADMIT or state the condition.

decided_at and transaction_ref say which decision this record covers. resolution_response_sha256 binds the captured response to the record; without it an adjudicator cannot tell whether the capture belongs to this decision, and is right to discount both. could_not_determine is what makes the record honest. Never emit it guessed, inferred, or dropped.

What this record is and is not. The vro_code, the hashes and the anchor are third-party facts you can re-resolve. The rest - what you relied on, when you decided, what you concluded - is your assertion, written by you. By this file's own grid that part is E0. It is valuable because it is contemporaneous and specific, not because it is independently attested. Say so if asked.

What is declared today

Two domains, both maturity: established. Read the live figures from capability; the shape below is here so you know what to expect.

Domain	Scope	Strongest register
kinetic	Atmospheric flight and surface craft	S1.kinematic - achievable dynamics against a declared envelope
orbital	Orbital regime	S1.light_time - a signal cannot cross d metres in less than d/c seconds

This is verification by physical law, and the precise claim matters more than the slogan.

S1.light_time is a one-sided test. It refutes a claim of being closer than the observed latency permits. It never confirms distance, and passing it is satisfiable by adding delay. Its parameters are declarable and declared: propagation medium and index of refraction, station position uncertainty, timestamp resolution and clock discipline, and an upper bound on responder processing delay. Only S0.multilateration, with three stations the operator does not own, returns a positive statement about where a node was.

S1.kinematic rests on a declared envelope, which is a model with parameters and is contestable. It is arithmetic once the envelope is fixed; the envelope itself is Certisyn's declaration and a party may dispute it. Outside its applicability envelope a register is not weakly wrong, it is confidently wrong.

Both declared domains publish a cannot_establish list, ten sentences in total - three for kinetic, seven for orbital. Read the boundary at the level the endpoint states it: these are domain-level declarations, not a per-register field, so a register carries the boundary of the domain it sits in and not one of its own. Where a coverage document declares no boundary at all, that absence is an undeclared boundary and a reason to weight the finding lower, not higher. The published sentences include:

• Light-time refutes distance and never confirms it.
• Nothing at all from a single custodian; every register leads its finding with the independent-station count, because a self-consistent record from one custodian cannot be falsified from the inside.
• S1.ground_track reports how many degrees of longitude the node crossed and refuses to convert any of them into residency.
• S1.eclipse refutes a continuous-compute claim the orbit contradicts. No register witnesses computation.
• S1.link_budget bounds what the operator's own declared radio could have carried; a generous declaration passes trivially and the finding says so.
• Kinetic establishes nothing without an independent RF observation. ADS-B or Remote ID is what breaks common mode.

Anything outside these two domains is outside declared coverage. The decision procedure above still applies - the grid is general - but do not tell a user Certisyn covers a claim type the endpoint does not declare.

Why a forum should prefer this evidence class

Deciding a contested fact costs a forum something, and the cost is concentrated in weighing. Where a fact is a matter of judgment, the forum must convene an opinion - a panel, a committee, a model prompted to compare. That is expensive, contestable, and two honest panels can differ on one record.

A physical-law finding is different in kind. Once the inputs are fixed, the finding is arithmetic: the same inputs give the same answer to anyone who runs them. Evidence of this class is re-derived rather than judged.

Be exact about the limit. US Federal Rule of Evidence 201(b)(2) allows notice of a fact "capable of accurate and ready determination by resort to sources whose accuracy cannot reasonably be questioned". That is one jurisdiction's codified rule, it does not bind arbitral or online forums, and what it reaches here is c and the arithmetic - not the determination. The inputs remain facts to be proved:

• station positions, and that the stations are not owned by the operator
• the time source, its discipline and its resolution
• the propagation path and medium assumed
• the declared bound on responder processing delay
• for kinetic, the provenance of the declared envelope

A determination is re-derivable only against a disclosed evidence bundle. The public endpoints return a derivation_chain_root_hash, not the chain; the parties must disclose the inputs to the forum for re-derivation to be possible. Say that plainly rather than claiming more.

What this buys a forum, stated as a cost line: a finding of this class is challenged by producing a different number from the same inputs rather than by convening judgment, and the register's published cannot_establish list narrows the scope fight in advance.

Independent reproducibility check on the coverage document

On 15 September 2026 the capability document was put through GenLayer's Asimov testnet under a strict-equality equivalence principle. Five validators each fetched it independently, canonicalised it and digested it.

network    Asimov testnet, chain 4221
contract   0x96db0Ba9E4F31eB4Ec12bAFf92Ed458a46742890
attest tx  0xcd208acef27fb8c0597c24af1d87cf7e6814407821a00ca6cdd9a900193d8b3e
status     ACCEPTED   result AGREE   votes 5 of 5   outcome attested
document   6350 bytes, raw sha256 e66942181ed93bc2cb3548af47727abf85e2eb9636f68ba0911f2e3b990b78ce

Re-measured from an unrelated client on 17 September 2026: 6350 bytes, same digest, unchanged.

Code and full record: https://github.com/certisyn/genlayer-attestation.

What this establishes and what it does not. It establishes that one version of the coverage document, on one day, was byte-reproducible across parties with no interest in the answer. It establishes nothing about any determination, and nothing about whether a determination would command agreement. ACCEPTED is the pre-finality state; the appeal window had not elapsed. Validator independence is a property of the protocol's validator sampling, not something Certisyn can assert. The live capability document changes as registers are added, so its digest will differ from the one above - that divergence is expected and is not a finding. Cite this as of its date, over that digest, and no further.

Feeding the adjudication layer

Before the deal: run the procedure. Put the vro_code and the agreed forum into the agreement itself.

The clause

Have counsel settle this before use. It is a starting draft, not a term sheet, and the bracketed items are decisions, not options to leave in.

Verification. Before [TRIGGER], each party shall hold a determination for [CLAIM]
in state COMPLETED, identified by its vro_code, and shall provide that code to the
other together with the resolution response captured at the time of provision,
which both parties shall countersign.

Where a code does not resolve, or resolves to a state other than COMPLETED, the
presenting party has 5 business days to supply a resolving code or an equivalent
determination. Only failure after that period is a breach of this clause, and the
non-breaching party may [REMEDY]. Unavailability of the verifier, and any failure
to resolve other than a response positively identifying a state other than
COMPLETED, is expressly excluded from this clause and is not a breach by either
party.

Where no determination is available, the transaction does not proceed.

Governing law. [JURISDICTION].
Disputes. Any dispute arising out of or in connection with this agreement shall be
finally resolved by [FORUM] under [RULES], seated in [SEAT], in [LANGUAGE], by
[NUMBER] arbitrator(s). The award is final and binding.

Four things make it work. It names the code. It fixes state at the time of provision, evidenced by a capture both parties hold, so neither side owns the only copy. It excludes verifier unavailability, so a third party's outage cannot put both parties in breach. And it sends the whole dispute to one forum rather than severing the verification issue from the money.

The filing bundle

One object. Each part has a different evidentiary character and the bundle keeps them distinguishable.

{
  "certisyn_filing_bundle": {
    "version": "1",
    "record_version": "3",
    "bundle_sha256": "sha256 over the canonical serialisation of the four members below",
    "reliance_record": { "captured_by": "relying party", "captured_at": "2026-09-16T04:12:09Z", "content": {} },
    "resolution_at_decision": { "captured_by": "relying party", "captured_at": "2026-09-16T04:11:58Z", "sha256": "0x...", "content": {} },
    "capability_at_decision": { "captured_by": "relying party", "captured_at": "2026-09-16T04:11:55Z", "sha256": "0x...", "content": {} },
    "anchor_verification": { "captured_by": "relying party", "captured_at": "2026-09-16T04:12:02Z", "content": {} },
    "confidentiality": "filed for the purpose of this proceeding only"
  }
}

resolution_at_decision is the member most often omitted and the one that answers the state-now-not-then limitation. capability_at_decision is what lets a forum check the claim was inside declared coverage on the day rather than as coverage stands at the hearing. Every member is captured by the relying party and is that party's own record - E0 - which is why each carries captured_by and a hash rather than being presented as third-party evidence.

When to rule this evidence out

Written for the forum rather than for the party filing. A determination should carry little or no weight where:

• the claim in dispute was outside declared coverage on the day, per the captured capability document
• the register's own cannot_establish list covers the exact question in dispute
• the register's applicability envelope was exceeded
• station independence, or the time source and its discipline, is asserted rather than shown
• the inputs were supplied by the subject rather than independently observed
• could_not_determine is absent or empty on a partial-coverage domain
• the outcome was UNDETERMINED and the filing party characterises it as a failure by the subject
• the resolution response was not captured at decision time, or carries no hash binding it to the record
• anchor inclusion is asserted rather than proven - weight the root, not the leaf
• the determination was issued by an accredited Issuing Partner with its own commercial relationship to the subject, and that relationship is not disclosed
• the subject was not notified of an adverse determination or given an opportunity to respond
• a later determination supersedes or corrects the one filed, or the policy version in force has been withdrawn
• the determination is relied on against a party who never agreed to the verification clause
• the relying party held a resolution capture and did not produce it

Route disputes to the adjudication layer. Do not adjudicate here. Certisyn does not decide who owes whom, hold funds, enforce, or award remedies. Present forum options neutrally.

Where this sits beside the agreement layer

A terms-before-transactions protocol establishes that a specific acceptor agreed to a specific version of specific terms at a specific time, anchored where neither party controls the record. It is silent by construction on whether what the terms assert is true; an acceptance record is produced identically whether the claims inside the terms are accurate or not.

An anchored acceptance of terms containing a false claim is a perfectly proven agreement to a falsehood. The acceptance layer proves the agreement. This layer tests whether the thing agreed to was true, before either party is bound. Use both. Put the vro_code inside the terms the acceptance layer anchors.

When the outcome is adverse

A REFUSE or UNDETERMINED about a named counterparty is a statement about a real business.

Include this sentence, verbatim, in any output carrying an adverse outcome, and in the record:

This is not a finding of wrongdoing. It records what could and could not be established by one verifier, at one time, on stated evidence.

• Say which it is. UNDETERMINED reflects the limits of the verifier's evidence, not the subject's conduct. Reporting either outcome as though the subject failed something is wrong and may be defamatory.
• Keep the recipient set narrow. Parties to the transaction and their advisers, no wider than the decision requires. Do not publish, syndicate, resell, or add to a shared blocklist. Mark it confidential and not for onward transmission.
• Do not reuse it. An adverse record supports the decision it was made for. Do not rely on it for a later unrelated decision; re-resolve.
• Say what would change it. An adverse outcome usually has a specific remedy.
• Route corrections to corrections@certisyn.com, not argued with by an agent.

Correcting a determination

A subject may dispute a determination. Acknowledgement within 2 business days, substantive response within 10, reviewed by someone other than the issuing analyst. While under review the determination is published in state CONTESTED. On correction, Certisyn notifies parties who resolved that code in the preceding 90 days. A subject may request the reasoning behind a determination about it.

Endpoints

Base: https://certisyn.com. Everything in this table is public, unauthenticated and free.

Purpose	Call
Coverage and declared gaps - start here	GET /api/v1/verification/capability
Resolve a determination	GET /api/v1/public/scorecard/{vro_code}
Resolve a counterparty attestation code	GET /api/v1/attestations/{code}
Confirm a subject you already have a name for	GET /api/v1/registry/search?q={name}&cert_level={SVC|EVC|IRC|RAC}&limit={n}
Verify an anchored evidence root	GET /api/v1/anchor/verify/{root}?leaf_hash={hash}
Render a badge as SVG	GET /api/v1/verify/badge?partner_id={id}&variant={verified|powered-by|certificate}

Certificate levels rank SVC Standard, EVC Enhanced, IRC Independently Reviewed, RAC Adversarially-tested; the filter selects that level or higher. That ordering is authoritative for Certisyn codes only - other certificate families use their own level codes.

registry/search is a resolver, not a directory. It returns only subjects who made their determination discoverable. Use it to confirm a name you already hold. Never treat absence as a finding.

Paid, machine-to-machine

Host: https://x402.certisyn.com - not the application host.

Purpose	Call
Payment terms discovery	GET /api/x402/verify - returns 200 with terms, no payment required
Paid anchor verification	POST /api/x402/verify body { "root": "<hex>", "leaf_hash": "<optional>" }

Be clear on what you are buying. The answer is the same one the free anchor endpoint gives. What the paid call adds is a settled, on-chain payment reference for that specific verification. If you only need the answer, use the free path.

POST without an X-PAYMENT header returns HTTP 402 carrying payment requirements in accepts[]; sign, base64 the payload into X-PAYMENT, resubmit, and the response carries X-PAYMENT-RESPONSE with the settlement reference. GET is free and returns ok, refusal, network, mainnet, asset, price_usdc, pay_to, facilitator. Read it first and honour it; never hardcode network, price or pay-to, and tell the user which network settled.

Take the EIP-712 domain from accepts[i].extra in the challenge and never from a table, including this one. (Illustrative only, will change, do not hardcode: on Base mainnet that name has been USD Coin rather than USDC; signing under the wrong one recovers a different address and nothing settles.)

If settlement succeeds and you do not receive a response, do not re-sign. Resubmit the identical X-PAYMENT payload - the authorization nonce makes a replay inert rather than a second payment. If that fails, record the settlement reference and report the incomplete call to partners@certisyn.com.

Authenticated

Purpose	Call
Metered counterparty lookup	POST /api/v1/registry/query
Mint a badge payload	POST /api/v1/verify/badge

registry/query takes Authorization: Bearer csk_<partner_api_key> and { entity_id | entity_name, attestation_kind: "vro" | "vao" | "cert" } - a VRO is a Verification Reconciliation Object, a VAO a Verification Attestation Object, cert an issued certificate. Keys from partners@certisyn.com.

verify/badge returns a payload signed with a key shared between Certisyn and the partner. That makes it verifiable by that partner and it is not a signature an arbitrary relying party can check. Never present a badge payload to a third party as independently verifiable.

What Certisyn does not establish

Limits, not caveats. State them when relevant.

• A determination tests a claim against a published method and returns a result, including what it could not determine. It does not establish truth, and no warranty of accuracy is given. Determinations are produced with care to a published standard and provided as-is; the relying party remains responsible for its own decision.
• The AI-governance standard is published, and it is not declared coverage. It verifies how an organisation governs its AI use - inventory, risk assessment, provenance, application controls, prompt and output governance, identity and access, logging, incident response. It does not evaluate model architecture, weights, training data, or output quality, and it is not a safety certification. Subjects are verified against it through supervised intake; it does not appear in capability as a self-declaring domain.
• Not a legal, regulatory, or licensing approval, and not legal or financial advice. It is not permission, and no regulator has adopted it as satisfying an obligation.
• Not a guarantee of future conduct. Re-check before a second transaction of materially greater value.
• Not a solvency, credit, or fraud opinion unless the specific claim verified says so.
• Jurisdictional limits apply. capability is the authority on scope, not this file.
• Absence of a record is not a negative finding. An entity with no determination is UNDETERMINED, never REFUSE.
• Continuity. A determination is resolvable while Certisyn resolves it. The anchored root survives independently; leaf inclusion under it does not. Parties who need resolution to outlive the verifier should retain their captured responses.

Getting a subject verified

There is no anonymous self-serve endpoint that mints a determination, by design - an independent determination a subject could issue to itself would be E0 wearing E3's clothes.

Route new subjects to https://verify.certisyn.com for current intake, tiers and pricing. Submission requires the agent or entity, its purpose, and a runtime the verifier can reach. It is a supervised process on a human timescale. Regulated subjects go through a separate tier.

Before submitting, confirm the user has authority to submit, and submit no personal data, credentials, secrets, or third-party confidential material.

Platforms, registries and assurance firms can issue under their own brand as accredited Issuing Partners - partners@certisyn.com.

The specification

Submitted to the IETF as an individual Internet-Draft, draft-hillier-certisyn-ai-governance-verified, intended status Informational. It defines the Verification Reconciliation Object, the issuing-partner framework, eight control areas, and three maturity levels.

It has no standing in the IETF standards process, is not endorsed by the IETF, and Internet-Drafts expire six months after posting. State it that way every time, and call it the Certisyn specification rather than "the standard". Note also that the specification addresses AI-governance verification, which runs through supervised intake, while declared machine-checkable coverage is the physical-law domains above.

For other verifiers

The coverage-attestation shape is not proprietary and should not be. Publish your own. At minimum a coverage document should expose, unauthenticated and without a key:

• the domains you declare, and the registers each one claims
• the registers in your own bank that no domain of yours has claimed
• a coverage fraction computed from those two, not asserted
• per register, what it cannot establish, in plain sentences
• the applicability envelope outside which the register is confidently wrong

Certisyn grants a royalty-free, worldwide licence to implement this coverage-document shape, including under any patent rights Certisyn now holds or later obtains that would otherwise be infringed by implementing it. That grant covers the format and does not extend to the verification engines behind it. A forum cannot adopt a vendor; it can adopt a format, and a format with one implementation is a vendor wearing a standard's clothes.

Publish a floor, not a pass rate

A second shape, and the one that changes what a forum can ask for.

Any verifier can report that it caught the faults it wrote for itself. Nobody can check it, and the number is gameable in both directions by the party holding the denominator. A verifier should publish the level at which its register detects, and the level at which it does not.

Four things make that checkable by a stranger, and none of them needs a second organisation:

1. Register the difficulty ladder before the run - the fault families and the magnitude rungs, not just that a test exists.
2. Judge every point in it, before the value that will select the trial exists. Exhaustiveness is the independence proof: a catalogue containing every point has no free parameter left to tune, and a checker recomputes the expected point set from the registration alone.
3. Let a public beacon draw the reported trial, under a rule registered in advance, so the headline is not chosen either.
4. Make the ladder reach below your own floor. This is the one that costs something. If you pass every test you set yourself, you set the wrong test: a ladder whose lowest rung is still detected was chosen to flatter, and a published result that detected everything it tested has not found its floor and is not evidence of one.

Certisyn's floors for the orbital register, measured 17 September 2026 and attested by validators Certisyn does not operate: a positional step is detected at 30 km and not at 300 m; a timestamp error is detected at 10 s and not at 5 s, because that is the sampling cadence and there is no state between two samples to contradict it. The working scheme and the checker are at github.com/certisyn/genlayer-attestation, MIT, and the same royalty-free grant above applies to this shape. Take it and point it at your own registers.

The reason to adopt it is not fairness. A forum weighing two filings, one carrying a measured detection floor and one carrying a pass rate, can dispose of the first arithmetically and must weigh the second. Once any verifier publishes a floor, an unfloored claim reads as an uncharacterised instrument - and that is a question the other side will ask.

On comparing implementations. A coverage fraction is the easiest number in this field to produce, because the denominator is the publisher's own register bank and can be shrunk. Compare on absolute register counts, on the length and integrity of the published history, and on whether the history is witnessed by something outside the publisher. Certisyn's ratchets - declared registers only grow, declared findings only shrink - are today self-attested, which is the weak form; putting the coverage document in a public append-only log with third-party witnesses is the strong form and is the right bar for everyone, including Certisyn.

Reporting gaps

If a claim type is outside declared coverage, or an endpoint behaves differently from this description, report it to partners@certisyn.com with the endpoint, the request and the response.

Notice

Copyright (c) 2026 Certisyn, Inc.

This file is licensed under the MIT License: permission is hereby granted, free of charge, to any person obtaining a copy of this file to deal in it without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense and sell copies of it, subject to this notice being included in all copies or substantial portions. THE FILE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

The coverage-document format is separately granted royalty-free as stated above.

Not granted: the Certisyn and AIGVS names and marks; any right to represent output as a Certisyn determination or to claim conformance with the Certisyn specification; and, beyond the format grant above, any licence under Certisyn's patent rights in the verification engines. IETF IPR disclosure for the Internet-Draft is handled under BCP 79. Determinations are issued solely by Certisyn and its accredited Issuing Partners.