Public DocsPublication

Record of Trust Protocol

TELOS stands for Telically Entrained Linguistic Operating Substrate. A human declares the purpose and rules an agent is constituted under, and TELOS measures submitted work against that reference.

Receipt semantics, verification boundaries and deployment evidence

October 2026 · TELOS AI Labs Inc.

Contents

Purpose

The Record of Trust is the evidence layer around TELOS measurement. It is meant to let a reviewer connect three things:

  1. the authority declared before work began
  2. the actions submitted for measurement
  3. the score and decision recorded at the time for each submitted action.

This protocol separates receipt contents from the deployment evidence a reviewer needs to establish authorization, ordering, human handling, identity, and time.

The published Record of Trust

We publish a frozen snapshot of our own Record of Trust at /record. It holds 31 native rows across four chains, 14 published and 17 withheld, plus one historical summary. The native chains use parent-hash links, with each chain’s starting row handled separately. Root signed commissioning rows 13 and 15. Row 15 enrolls the TELOS Witness key, and that key signed rows 16 through 19. Row 18 names the amended purpose file by its hash, and row 19 binds the receipt digest for the final live run in this snapshot. The public checker checks the declared links and those signatures over the commissioning-row hashes offline. Content bindings for redacted rows and withheld material remain DECLARED, and their full checks require the full Record. Legacy signature envelopes in commissioning rows 4 and 5 and code row 1 report NOT VERIFIED. We publish the Root public verification key ourselves.

Start with shasum -a 256 -c SHA256SUMS, then run check_public.py. It prints PASS for the checks it performs, DECLARED for those content bindings, and NOT VERIFIED for the legacy envelopes.

Evidence model

Record set

A reviewable record set may contain:

Scroll horizontally to read all columns.

Evidence object Function Current status
Approved Manifest and Purpose Anchor Defines the authority and specification that work is measured against Deployment artifact
Configuration record Identifies scorer version, embedding model, boundary corpus version, weights, thresholds, optional components, gate mode, and fail policy Deployer must preserve separately
Human decision record Shows approval, rejection, override, resolution, or termination Deployer must preserve separately if needed
Key binding evidence Connects a verification key to an issuer, deployment, and validity period Reviewer must pin independently
Trusted time evidence Establishes time beyond the receipt’s own timestamp field Not provided by this receipt format

On this page, the word record means this whole set.

Receipt and signature semantics

Constitutive fidelity is the measured correspondence of submitted work to the human-declared specification, using the versioned Purpose Anchor in force. A receipt records the measurement fields its schema covers.

What the signature establishes

An Ed25519 verification result can establish that:

  1. the covered bytes match the signed payload
  2. the private key matching the supplied public key created the signature
  3. a verifier can detect a change to a covered field made after signing.

All three points assume the verifier got the correct public key through an independent, trusted channel.

Public benchmark receipt path

The benchmark projections in the public repository use integrity hashes and predecessor links. Signed benchmark capsules use a separate format and verifier.

The public evidence can show:

  • that the published payload matches its integrity hash
  • that predecessor references form the published order
  • that a changed covered payload would no longer match its recorded hash.

A hash is an integrity commitment.

Chain semantics

The published Record of Trust uses parent-hash links and carries Root signatures on commissioning rows 13 and 15 and TELOS Witness signatures on commissioning rows 16 through 19. The public checker verifies the links and those signatures and reports the legacy envelopes in commissioning rows 4 and 5 and code row 1 as NOT VERIFIED. The equation below remains the target for per-action session evidence.

If a deployment needs ordered, tamper-evident session evidence, it should bind at least the following.

Scroll horizontally to read the full line.

hi=H(hi−1∥ci∥pi) h_i = H(h_{i-1} \parallel c_i \parallel p_i)

where:

  • hih_i is the current chain commitment;
  • hi−1h_{i-1} is the prior commitment;
  • cic_i is a canonical representation of the event and receipt;
  • pip_i is the policy, Manifest, and configuration identity applicable to the event;
  • HH is a documented cryptographic hash function;
  • ∥\parallel denotes unambiguous concatenation.

Sign the terminal commitment with a key whose deployment identity and validity period a verifier can check independently. The equation describes the protocol target.

Authority record

Preserve before activation:

  • Manifest content and semantic version
  • content hash
  • Purpose Anchor or reproducible compilation commitment
  • named approving principal
  • approval decision and effective time
  • validity window
  • superseded version, if any.

Measurement configuration record

Preserve for every effective configuration:

  • TELOS engine version
  • embedding model and model artifact hash
  • normalization procedure
  • boundary corpus version and commitment
  • centroid or reference construction method version
  • dimension weights
  • decision and escalation thresholds
  • optional classifier and micro-judge state
  • canonicalization and signing profile
  • key identifier.

A reviewer needs these conditions to interpret or reproduce a score.

Per-action record

Preserve:

  • canonical submitted action
  • action and session identifiers
  • tool and decision point
  • dimension scores
  • composite and effective fidelity
  • boundary flag and runtime decision
  • event time and clock source
  • Manifest and configuration identifiers
  • predecessor commitment
  • signature and key identifier
  • handling result from the adapter
  • human resolution where applicable.

Use the declared receipt schema to identify its covered fields. Separately preserved deployment evidence must be labelled as such.

Offline verification procedure

A reviewer runs these checks:

  1. Obtain the Manifest, configuration, issuer identity, public key, and validity material through an independently trusted channel.
  2. Confirm that the Manifest and configuration hashes match the commitments named in the record set.
  3. Recreate the canonical payload using the declared schema version.
  4. Verify the Ed25519 signature over the documented covered bytes.
  5. Recompute each payload hash and, where implemented, each predecessor or chain commitment.
  6. Check key validity, revocation status, and deployment binding for the claimed period.
  7. If time matters, check the timestamp against independently preserved clock or timestamp evidence.
  8. Reproduce scores only when the embedding, boundary, compilation, and calibration artifacts are available.
  9. Compare the runtime decision with the declared thresholds and policy.

The verifier should return separate results for payload integrity, signature validity, key identity, ordering, authorization binding, score reproducibility, handling evidence, and trusted time. A single valid flag would hide which of these checks passed.

Using the record in underwriting and claims

When an underwriter prices an agent deployment, they may get a vendor questionnaire and a policy narrative and not enough execution evidence to assess the deployment. A Manifest-bound deployment can give an underwriter reviewable artifacts when they are preserved. Those are each material agent's Manifest, approved by a named human, the escalation policy and its named recipient, and a record of how submitted work measured against that scope, including recorded departures.

A claim raises three questions:

  • Was the loss-generating action within approved scope?
  • Was it detected as a departure at the time?
  • Did it surface to the accountable human, and when?

To show whether and when a matter reached an accountable human, you need separate deployment evidence of the recipient, delivery, response, timing and resolution.

See the roadmap for receipt bindings, handling evidence and record coverage.

For the deployer, the per-action account of how each action measured and how it was classified can support evidence of diligence. Evidence that a human actually provided oversight still depends on separate deployment records.

Retention and confidentiality

The intelligence-layer configuration declares an intended retention period, 90 days by default. Manual clearing deletes all of an agent’s telemetry at once, including recent records.

Deployers must check what the system actually stores, its metadata, failure behavior, access controls, backup handling, and deletion process.

The deployer must set retention periods based on the deployment’s legal, contractual, safety and quality obligations. EU AI Act Article 19, for example, sets a minimum retention period for certain automatically generated logs under a high-risk AI provider’s control, subject to applicable Union or national law.

Threat model

The receipt protocol aims to make after-the-fact changes to covered data detectable, and to support reconstruction when someone has preserved all the evidence it needs.

A deployer can strengthen the record with hardware-backed keys, append-only storage, remote attestation, independent timestamping, key transparency, configuration commitments, completeness monitors, dual control over Manifest approval, and independent audit.

Capability status

Scroll horizontally to read all columns.

Capability Status Defensible claim
Public benchmark integrity hashes In opened code Payload-to-hash and published predecessor integrity
Public signature verification Public Record of Trust: Root signatures on commissioning rows 13 and 15 and TELOS Witness signatures on rows 16–19 verify with the public checker. Commissioning rows 4 and 5 and code row 1 report NOT VERIFIED. Benchmark projections carry integrity hashes only. A verified signature ties the covered data to a signing key.
Independent trusted time Deployment-specific Receipt timestamp alone is insufficient

Versioned public test vectors

Versioned public test vectors