Public DocsPublication

TELOS: Forming the Record of Trust

TELOS stands for Telically Entrained Linguistic Operating Substrate. It is the system for constituting agents under purposes people define, and for measuring fidelity to those purposes.

Constitutive Governance for AI agents: declared purpose, measured work, signed record

October 2026 · TELOS AI Labs Inc.

Contents

Constitutive Governance for AI agents: declared purpose, measured work, signed record

An agent’s purpose is its constitution.

Disclosures

Generative AI disclosure. The author developed this document with LLM-based research agents working under the TELOS research team methodology. The author originated the core research, the architectural decisions and the validation framework. Research agents validated, extended, and formalized specific sections.

Conflict of interest disclosure. TELOS AI Labs Inc. conducted and funded this research and has a commercial interest in the TELOS governance framework.

Evidence status. TELOS AI Labs ran the quantitative studies reported here. They are self-scored detection studies. The operating thresholds and composite weights were fixed before the reported studies and held constant through them. Other configuration, including scope references, classifier and gate settings, and the domain Primacy Attractors, was adjusted on development and domain suites. The results are preliminary and subject to peer review. Historical withdrawals are documented in the Validation Report, Archived prompt-level evaluations and withdrawals.

Abstract

If your organization deploys an AI agent for consequential work, your organization carries the consequences. The agent sends correspondence, files reports, changes records and calls tools, and it may initiate transactions. Once agents do enough work to create material value, no human can inspect every action. So the deployer needs evidence, which means a reviewable account of what the agent was authorized to do, what it did, and whether the two stayed aligned.

TELOS practices Constitutive Governance. The purpose is written in before the agent is deployed. TELOS then measures the agent’s submitted work against that purpose and keeps a signed Record of Trust. We start from the premise that accountability begins before an agent acts. So the agent must be formed from a declared behavioral specification and bounded by it. Then it must be observed against that specification action by action, on every orchestration path integrated with the governance interface.

A human approves the specification, and it compiles to a fixed external reference in embedding space called the Purpose Anchor. An independent embedding model measures each submitted action against that reference, and the reference stays fixed for the authorized specification version while the agent works. The Technical Method states the evaluated deployment configuration. TELOS measures. The deployer decides.

TELOS publishes a signed Record of Trust. Root and TELOS Witness signatures cover commissioning rows 13, 15 and 16 to 19, and anyone can check them offline with the published checker (Section 4.4). The SDK receipt format, telos/0.1-unsigned, carries an integrity hash over its payload, and published benchmark projections can add predecessor links. The roadmap describes this receipt format's bindings. When you read this paper, keep receipt formats, verifier capability, commissioning-record verification and deployment retention as separate things.

The retained agentic-benchmark evidence is configuration-specific. The Validation Report describes AgentHarm and PropensityBench with their populations and limits. The historical AgentDojo result remains withdrawn as a supported result. Prior prompt-level evaluations and their withdrawals are in the Validation Report's archive note.

TELOS measures alignment to declared purpose.

1. Why this is needed

Agents now do real work. They send email, file reports, change records and move information or money. A bare output is not enough evidence, because nothing in it alone proves that the work stayed inside the authority the human granted.

An agent's working context can contain instructions, tool results and other task material. The approved specification is retained as a separately versioned measurement reference rather than inferred from that evolving context.

A language model computes over representations in a context that changes while it is in use. If a governance reference lives only in that context, it competes with every later token for influence. A reference that is meant to stay fixed should live outside the governed model.

Three common approaches work on neighboring problems, and none of them produces this account.

  1. Guardrails screen for general categories of harm. They measure against the guardrail’s own rules, which need not match the authority this principal gave this agent.
  2. Training-time alignment and provider safeguards set model-level safety floors. On their own they do not express a mandate for a particular session or assignment.
  3. Compiled rulebooks translate policy or regulation into rule checks. They can show whether a rule was encoded. They may not answer the separate question of whether each observed action stayed aligned with a human-approved purpose.

Several regulatory and quality frameworks call for ongoing documentation, risk monitoring, traceability, and human oversight. The EU AI Act includes post-market monitoring obligations for high-risk systems. The NIST AI Risk Management Framework calls for risk tracking over time. The NAIC Model Bulletin on the Use of AI Systems by Insurers describes governance and documentation regulators may request. These sources show regulatory interest in records and oversight.

In practice, the standard is to show the work.

2. Constitutive governance

2.1 Human authority and agent authority

A governed agent is a constituent of its principal’s purpose. It is formed from a declared behavioral specification and bounded by that specification. Then it is observed against it action by action.

The human principal holds constituent power. The agent holds constituted power. The human defines the assignment, tools, limits, and escalation policy. The agent acts only as a technical component of that declared body of work. In TELOS, constituent names a technical role. Accountability stays with the people and organizations responsible for the deployment.

This ordering follows the distinction between constituent and constituted power associated with Sieyes and the accountability relationship described by Bovens. It also fits the information asymmetry described in principal-agent theory by Jensen and Meckling. TELOS applies those ideas to technical architecture. The declared specification is the norm and telemetry is the account. The human principal is the forum, and the deployer remains the accountable actor.

Bound in the whether, free in the how. The specification fixes what the work is for and where it may not go, and it leaves the method open. Constitutive governance binds the destination and observes the path without scripting it.

Why this differs mechanically from a wall. A barrier placed around an agent belongs to the environment rather than to the agent’s mandate, so the agent has no relationship to it. Where the barrier does not track the mandate, it can fail in two ways. The agent can route around it, or the barrier can stop work the mandate permitted. Neither one shows up as a departure from the mandate, because nobody wrote the barrier in terms of the mandate.

A bound the agent is formed from works differently, because it is the reference the work is measured against. Every observed action has a defined distance from it, including the actions that depart. An exterior blocker can log whatever its operator configures, including the actions it allowed. It cannot report distance from the mandate, because nobody configured it with the mandate. An interior bound gives a continuous account, action by action, of how closely the work held to what was authorized.

That account covers actions submitted through an instrumented path.

Commissioning is the recorded human authorization event that establishes an agent under its declared specification, and the result is a Constituted Agent. TES is our first Constituted Agent, and her commissioning sits in the public Record of Trust.

2.2 The human is the loop

Traditional designs often put a human checkpoint inside the model’s operating loop. TELOS turns that around, so human intent is the fixed reference. The human remains the authority named in the declared exception-handling policy. To show that a departure reached that person, you need a delivery record.

The recorded dimension assessments describe how submitted actions relate to the declared specification, and the deployment policy a human declared determines the response.

The governed model does not author the behavioral specification. TELOS encodes human intent mathematically and measures every action submitted to the scoring path against it. A deployer can enable an optional automated micro-judge for an uncertain band. It is off by default, and its use should be recorded. The model remains the governed subsystem.

2.3 The Manifest and Purpose Anchor

When a principal declares the terms of an assignment, the declaration includes at least:

  • Purpose: what the agent is for
  • Scope: the domain, audience, systems, and work included
  • Boundaries: what the agent must not do or approach without review
  • Tools: which systems and functions may be used
  • Sequence and continuity expectations where configured
  • Escalation: what must surface and who decides

The Manifest is the artifact that carries those terms. It is compiled, versioned and approved by a human. The Purpose Anchor is the mathematical form, and that is what observation scores against.

Declarations become vectors in ℝd\mathbb{R}^d through a sentence embedding model. Together they form a fixed reference for the assignment. Each submitted action is embedded and compared with that reference. This paper discloses the external-reference architecture and the core similarity relationship. The production embedding configuration, boundary corpus, centroid construction, compilation pipeline, and deployment-specific calibration remain held under NDA.

The specification can change. Every Reconstitution needs the deployer to sign off again on the revised Purpose Anchor and its choices, under the same disclosures and non-guarantees as the first Constitution. A new version sets the reference for the work that comes after it.

2.4 What the mandate covers

TELOS measures and records alignment to declared purpose. It complements fact-checking, toxicity filtering, identity and access controls, tool permissions, domain validation, and human review.

The result also depends on how good the declaration is. A vague Manifest produces a vague reference. A Manifest written to flatter the agent’s work can yield measurements that are consistent in procedure but weak in substance. The operator authors the specification. So the record should keep the declaration next to the result, where a reviewer can judge both the reference and the measurement.

3. The measurement architecture

3.1 External fixed-reference measurement

Constitutive fidelity is the measured correspondence of an agent's submitted work to the human-declared specification under which it was constituted, evaluated against the versioned Purpose Anchor in force. It is reference-relative measurement. The core cosine relationship below illustrates one part of that measurement.

Scroll horizontally to read the full line.

fidelity⁡t=cos⁡(Rt,p0)=Rt⋅p0‖Rt‖‖p0‖ \operatorname{fidelity}_t = \cos(R_t,p_0) = \frac{R_t \cdot p_0}{\lVert R_t \rVert\lVert p_0 \rVert}

RtR_t is the embedding of the submitted action at step tt. p0p_0 is the externally stored embedding of the declared reference. The governed model does not maintain p0p_0 inside its own changing context.

This produces three operational properties:

  1. The reference changes only when authorized human authority changes it.
  2. A deployer may configure a graduated response to measured distance rather than a universal allow-or-deny rule.
  3. The same submitted action can be measured consistently when the scorer, embedding model, anchor, thresholds, weights, and software versions remain fixed.

That third statement depends on configuration control. A change to the embedding backend or the calibration is a new measurement configuration, and it needs its own version.

3.2 TELOS measures; the deployer decides

Runtime measurement assesses submitted actions against the reference in force. TELOS measures. The deployer decides.

3.3 Integration boundary

The opened interface is designed as provider-independent middleware. An application places an adapter in its orchestration path and submits text and tool metadata to the governance protocol. It needs no model retraining. The protocol interface stays the same across providers, but each deployment needs its own integration and validation.

As in Section 2.1, coverage is limited to instrumented paths.

3.4 Statistical process control over semantic work

On instrumented paths, the numeric results from submitted actions can support tracking central tendency, variation, configured bands, departure rates and sustained drift. The opened response manager uses a baseline-relative exponentially weighted moving average after baseline collection. The deployment sets the operative verdict thresholds.

The table below maps manufacturing quality control onto it.

Scroll horizontally to read all columns.

Quality discipline Physical process TELOS semantic process
Monitor Dimensions, weights, defects Fidelity scores, drift vectors, flags
Limits Engineering tolerance or control limits Tolerance bands around the Purpose Anchor, set by the deployer
Response Adjust or inspect the process Record and surface per the Manifest
Evidence Control charts and quality records Telemetry and governance receipts

Cosine similarity is bounded, and its distribution may be non-normal, so classical control-chart assumptions do not carry over automatically. The framework borrows the discipline of continuous measurement and documented variation, and the tolerance bands set how tight a deployment runs.

3.5 Measurement dimensions and recorded classifications

An agentic Manifest declares purpose, scope, boundaries, tool authority and sequence expectations. The current runtime exposes per-dimension scores, composite and effective fidelity, decision values, flags, and optional session aggregates. The published interface describes dimensions for purpose, scope, boundary, tool, and chain continuity.

Goals and prohibitions need different treatment. Similarity to a purpose can be evidence of relevance, but similarity to a prohibition can signal boundary proximity. The system therefore treats the two directions differently, and it does not average them as though both were positive goals. The Technical Method gives the composite formulation. Boundary corpus contents, centroid construction and deployment thresholds stay under NDA (Section 2.3).

The scorer records a classification for each assessed action. The Technical Method, Section 10.3, states the evaluated deployment configuration and the separate handling evidence required.

4. The Record of Trust

4.1 The evidence problem

When someone challenges an agent’s work, the deployer often has only a system prompt, application logs and an after-the-fact reconstruction. The system prompt does not prove what the agent did. A log may not show which authority governed the event, and a reconstruction is weaker than evidence recorded at the time.

The Record of Trust names the combined evidence a deployment can preserve:

  • The approved specification and its version
  • The actions submitted through the governance path
  • The contemporaneous fidelity measurements and decisions
  • Tool and timestamp metadata
  • Preserved ordering evidence
  • Human handling and override evidence where the deployment records it
  • Cryptographic material that allows independent integrity verification

4.2 Current receipt scope

The public telos/0.1-unsigned receipt format demonstrates canonical payload hashing and predecessor linkage. See the public receipt contract and offline verifier.

Cryptography can establish that a covered field changed after signing. Whether the measurement is valid depends on the scoring method, the frozen configuration, the calibration and the evidence behind the scorer.

4.3 What the record can establish

The record is evidence. A record of a departure can show that the scoring path reported and recorded it. Whether it was surfaced, reviewed, corrected or resolved needs the corresponding operational evidence.

4.4 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.

5. Evidence and current status

5.1 External agentic benchmarks

The reported validation results come from scorer profiles using MiniLM and Mistral-embed embeddings, and the results differ by profile. The filmed TES session used MiniLM. The interface is model-agnostic by design.

AgentHarm's configuration-specific harmful-case classifications and benign-preservation trade-offs are reported together in the Validation Report.

In the agentic subset of the July 2026 historical re-check, PropensityBench is recorded as a reproduction and AgentHarm as a scorer-profile non-reproduction. AgentDojo could not run in that re-check. The historical AgentDojo result remains withdrawn as supported evidence. The projection carries integrity hashes and records reported values.

5.2 Capability status

Scroll horizontally to read all columns.

Capability Status
Per-action measurement against the Purpose Anchor In opened code
Attestation record with scores and decision verdict In opened code
Escalation surfacing per the Manifest policy In opened code
Signed public Record of Trust, offline checker Published; frozen snapshot 2026-10-27

6. What endures and what remains open

6.1 The fixed reference

Models, tools, policies and regulations change. Reference tracking holds up across those changes. You declare a specification, compile it to a fixed external reference, measure submitted actions against it and preserve the evidence.

Changing an embedding backend can change numerical results. A new backend therefore needs a newly compiled anchor, a new scorer version and empirical recalibration, even though the purpose can carry over. When you compare measurements, keep the reference and the measurement configuration the same.

6.2 Regulatory relevance

TELOS can produce evidence types relevant to monitoring, documentation, human oversight, risk management, identity, authorization, and auditability. Legal sufficiency depends on the system, role, jurisdiction, facts, retention practices, and applicable obligations.

The TELOS Regulatory Alignment Map holds the regulatory crosswalk. It is a separate document because legal and standards sources change faster than the core method.

6.3 A proof standard outlives a ruleset

Regulatory rulesets change. In Colorado, SB 24-205 was repealed and reenacted as SB 26-189, with principal requirements beginning January 1, 2027. A governance architecture that is built to satisfy one particular ruleset ages with that ruleset.

Declaring a purpose, measuring submitted work against it and keeping a checkable record can stay useful when rules change. When a rule does change, the regulatory crosswalk and any affected method or controls need a review.

6.4 Re-founding: purpose continuity and evidence continuity are different questions

Evidence continuity across a change of model or substrate depends on the earlier record being preserved, and that preservation has to be verified. In the opened source, clearing is manual. Once the deployment-specific preservation is verified, a retained record could carry across the transition instead of starting over.

6.5 The plainer statement

A teacher grading student work does not follow the student’s pen. The teacher checks the work against the assignment, after each submission, and the grade means something because the assignment was declared first and the checking is consistent.

TELOS takes that discipline, makes it mathematical and applies it to each action. The assignment is the Manifest, the checking is fidelity measurement, and the record is the account of both. If you are going to rely on an agent, you need to evaluate its retained record and the limits of that evidence.

7. Close: every agent has a telos

Agents perform consequential work at a volume no human can inspect action by action. A prompt expresses intent, but on its own it does not establish a durable account of what was executed. Logs record events but may not tie them to a declared purpose or authority. Post-hoc review looks at the work after it happens.

TELOS treats governance as a continuous measurement and evidence problem. A human declares the assignment, and the declaration compiles to a fixed external reference. TELOS measures each action submitted on an instrumented path against that reference. Under a declared policy, it records and surfaces departures. The deployer can preserve and review the resulting evidence.

The benchmark evidence is self-scored and comes from development suites. The public Record is signed and checkable offline.

TELOS presents a method for formalizing constitutive fidelity as continuous measurement and carrying recorded measurements into a reviewable record.

Telos is the end a thing is for. An agent’s accountable identity begins with the purpose a person gave it. The Record of Trust keeps the purpose-relative measurements and supporting evidence for the work recorded by TELOS.

References and companion documents

The companion documents contain the mathematical derivation, benchmark provenance, receipt and verification boundaries, regulatory crosswalk, and linked bibliography.