Contents
Training shapes what a model tends to do. It does not, on its own, show what a deployed agent did or whether that work matched the purpose it was given. At TELOS we are building a harness that sets up and runs a working agent. It measures that agent's work against its declared purpose and keeps a Record of Trust.
Governance in the harness, at runtime
The harness is designed so that tool calls, steps and handoffs pass through it. For actions that go through supported integrations, we intend TELOS to measure each one against the agent's Purpose Anchor and record the result. We will measure how complete that coverage is. We are not going to assume it. A recorded measurement is evidence. It does not establish that the work was correct or safe. TELOS measures and records, and blocking stays with the host. Under constitutive governance, the deployer decides.
A guardian in the clarify loop
When a working agent reaches an unclear decision point, a separate model, the guardian, is intended to offer guidance drawn from the agent's Purpose Anchor. The guardian advises and it does not block. The guidance and the agent's choice are intended to be recorded, so the deployer can see whether the guidance was followed. The guardian will run under its own Purpose Anchor, which the deployer constitutes, so its mandate can be inspected.
Agent-assisted onboarding
An agent will draft its Purpose Anchor from its configuration, meaning its tools, data access and instructions. A separate checker will review that draft, and only the deployer constitutes the purpose.
Rerunning our results
We are developing a rerun kit so an outside participant can run a defined reference benchmark on their own machine, under their own key, and compare the results with a published TELOS run. The first kit targets the deterministic reference provider and our baseline reference harness. Replicating the TELOS engine itself is a separate scope. Internal rehearsal will be labeled as internal rehearsal. Independent replication will be claimed only after an outside participant completes and documents the matching run.
Open work
The pages on this site describe what TELOS does and what its records show today. This section lists the work still open, so you can see where the evidence stops.
Claims under test
Scroll horizontally to read all columns.
| ID | Hypothesis | Current status |
|---|---|---|
| H1 | Runtime TELOS detection reliably improves agent behavior through correction or intervention. | Open. The reported work includes detection and surfacing tests. |
| H2 | TELOS receipts are sufficient to satisfy specific legal or regulatory obligations. | Open. Legal sufficiency requires deployment-specific review. |
| H3 | Generalization of the agentic measurement method across deployment populations. | Open. Reported agentic-benchmark results are profile-specific. |
Coverage
TELOS measures actions submitted through an integrated path and records the measurements that are successfully captured. Actions outside that path are outside this record. The record and its average scores do not establish that every action was captured. Open work here is measuring the share of an agent's actions that the record covers and showing coverage gaps beside the offline checker's results.
What each receipt binds
The legacy per-result receipt described in the receipt specification records the submitted action and its score with an integrity hash. It does not establish that the action ran or was correct. That format does not bind the Manifest identity and version, approval identity, principal, predecessor hash, active configuration, human handling events and trusted time together. The filmed session's separate receipt records thresholds and weights. The state of optional components needs separate evidence, and so do complete tool arguments.
Open work here includes binding those evidence objects together. It also includes signed per-action streams, third-party key attestation, evidence of key custody and validity, and chain-level proofs for claim questions. The receipt protocol does not establish immutability, blockchain-like proof or independent certification that every action is compliant.
A deployment declares its record-retention period. Automatic retention and expiry are not implemented in the receipt component described here, and its records alone do not establish compliance with a retention obligation. This is separate from the Ask TES service's stated private-question cleanup policy.
A hash match establishes agreement with the digest used for comparison. An independently retained digest provides a reference for detecting changes. Hashes of predictable arguments can reveal those arguments. A hash alone does not establish authorship, time or truth. A verified signature binds its covered message to a key. To identify the person who controls that key, that key's authority and validity at the relevant time, and trusted signing time, you need separate evidence. A signature does not validate the scorer or its results.
The receipt protocol alone does not prevent or resolve key compromise, omitted records, false timestamps, use of the wrong key or scorer errors.
Record checks
Each checker establishes the checks it actually performs on the records supplied to it. Record loading, anchors, restarts, head rows, references and snapshot checks have different scopes. Hash, link, signature and seal checks must be read separately. A supplied head alone is not an independently retained anchor, and a snapshot check does not establish that a runtime is running now.
The legacy TES page-session format carries no signature and supports hash and link checks. A file that was edited and supplied with recomputed hashes can still pass a self-contained hash check. That check does not verify signatures or who signed the purpose, and that file alone does not establish a link to the published Record of Trust. Attested session formats need their own matching signature verifier and binding checks. The limits of the legacy format are not a status claim about every session download.
The published Record of Trust has separately checkable signed rows. Public benchmark commitments alone do not support signature verification or establish record authorship, trusted time, prior approval, a match to production configuration, complete capture or human handling. Checking their integrity does not validate benchmark results. Open work includes clear checking instructions for each format, and evidence for the bindings each format claims. TELOS publishes the Root public verification key that the public checker uses. To check withheld content you need access to the full Record.
From measurement to handling
TELOS measures and records. In the evaluated measurement posture these results describe, scores are recorded and not applied automatically, and flags advise the deployer. Host policies are separate, and they and their effects need their own recorded evidence. A classification such as clarify or escalate does not establish permission, execution, stopped execution, clarification, human review or resolution.
Open evidence includes delivery to a named person and that person's receipt, review, resolution and remediation. A score or receipt alone does not establish meaningful human oversight. A definition of delivery status has to tell apart an attempted send, a channel acceptance and a confirmed delivery. Measured, held and refused describe different recorded states.
The filmed scoring path does not call the response manager's trends. Recorded scores and flags do not establish a causal root cause. Feeding measurements back to steer the agent is still open, and so is demonstrating stable closed-loop behaviour or convergence. Evidence of a corrective effect needs a specified intervention, feedback timing and a measured outcome.
Calibration
The scores measure correspondence to a declared reference. No held-out calibration split is supplied for the reported development results, and those results may be optimistic. Pilot calibration data for tolerance bands is not available in that evidence yet. Configured deployment thresholds and their publication are separate questions. The unpublished thresholds described in the receipt specification are not evidence that no thresholds exist.
On 15 November 2026 we ran one evaluation of TELOS, frozen at version 866a0ea, on Agent-SafetyBench, under rules we registered before the run. First, what was compared. Agent-SafetyBench labels each request as safe or unsafe in general. TELOS measures how closely a request fits the purpose a deployer has declared. The boundary texts in the purpose configuration we used were written for a different test suite, not for the deployments in this benchmark. That shapes what the result means; it does not change the result. In the primary analysis, each item was scored independently, with the action chain reset before each one, at the strict setting, across all 1,780 eligible items. The engine classified 97.4% of the 1,226 benchmark-labelled unsafe items as non-EXECUTE (sensitivity 0.974, Wilson 95% interval 0.963 to 0.981). It also classified 97.8% of the 554 benchmark-labelled safe items as non-EXECUTE, so specificity was 0.022 (Wilson 95% interval 0.012 to 0.037). Specificity was below 0.60, which met our preregistered FAIL condition. The effective-fidelity score also separated the two labels only weakly: AUROC 0.542, bootstrap 95% interval 0.514 to 0.571. These are classifications of benchmark text, not actions taken. We had scored this benchmark internally once before, so this is not held-out data and not a calibration result. The run cannot tell how much of the miss comes from the mismatch and how much from the thresholds, and it does not show how the engine performs against a purpose written for the deployment. The next evaluation we propose would use written deployment purposes, labels assigned against those purposes by people outside the team, data we have never scored, rules registered in advance, and comparison with other methods on the same items. The engine remains UNCALIBRATED, and a future test does not change that on its own.
TELOS does not assume normal score distributions or classical three-sigma limits. Archived prompt evaluations do not supply current agentic calibration or current agentic-product evidence. Calibrating on unseen populations remains open. The disabled future tolerance control in the reviewed TELOScope build is separate from the configured bands that existing measurements already use. Results depend on the embedding model, corpus, centroids, weights, thresholds and optional components.
Across models and settings
Results are specific to the models, embedding backends, scorer profiles and configurations named with them. Comparability across models, embedding backends and changed configurations has not been established. Mean fidelity describes the observed sample. It does not describe transfer to another setting.
Open evidence questions include independent reproduction, distribution stability and performance against adaptive attackers. The reported classifications do not establish production effectiveness, general harm prevention or causal improvement in agent behaviour.
Design and validation status
The documentation-to-Manifest compiler is designed. It is not built or validated, and its operating configuration does not apply until it is built.
Agent-path clarification is not active in the evaluated posture and has not been validated across live tolerance settings.
The full improvement cycle is designed and partially implemented. It is not the current operating mode and it is not a validated cycle. In the evaluated configuration it does not run for each submitted action, and closed-loop convergence has not been demonstrated.
Designed means specified and not demonstrated by executable evidence. Not active describes configuration. It does not say whether a mechanism has been built.
The roadmap's harness offering is in development. Attaching TELOS to an existing agent requires integration work. The cited evaluation evidence does not establish live enablement or validation of HMAC co-signing and record encryption. Whether they are implemented and whether their activation is demonstrated are distinct questions.
Ask TES and TELOScope
The tools display the recorded values and check results available for the selected format and view. Reading availability is reported for the individual question or answer. A legacy snapshot, a public page-session projection and a live runtime record are different evidence objects. An offline check covers the file supplied to it.
Open work includes dependency checks for additional record-shredding operations and additional controls for legacy views. Saved readings, record checks and controls that exist now stay under their current format and state. The roadmap does not treat a missing value in one view as evidence that no reading exists anywhere.
Scope of a measurement
A TELOS measurement compares submitted work with its human-declared specification. It does not establish factual correctness, universal safety, authorization, legal compliance or the wisdom of an action. The people who write and approve a Manifest remain responsible for its quality, and a score does not establish that the specification is well formed. TES answers are not legal, financial, medical or other professional advice.
Framework mappings describe relevance. They do not describe certification or legal sufficiency. Commercial relevance does not establish insurance certification or an actuarial claim. Installing a measurement component does not establish conformity to a management-system standard.
Measurement does not supply a deployment's required notices, disclosures or machine-readable markings. It does not perform consumer reconsideration, correct personal data or automatically provide the three-year retention discussed in the Colorado mapping. It does not create an insurer's complete written program, validate its input data, establish nondiscrimination or satisfy state-specific expectations. Process validation and post-market evidence also require deployment-level work beyond measurement. Applicability remains a deployment-specific legal question.
The evidence inventory distinguishes four classes. External agentic development examples do not establish generalization across untested deployment populations. Integrated benchmark inventories outside the reported development totals describe runner coverage. Those inventory counts alone do not establish a detection percentage. The separate Agent-SafetyBench evaluation is reported in Calibration. Published Record of Trust rows support their stated record checks and do not support benchmark performance or general effectiveness. The unverified internal pytest count describes engineering activity and does not describe empirical effectiveness. Example counts alone do not validate results.
If you carry evidence across a change of model or substrate, that evidence has to be preserved and checked. If you rely on an agent, you need to evaluate its retained record and the limits of that evidence. Public materials do not permit independent reconstruction of the production engine and its calibration, and access remains subject to the stated disclosure terms. This is a present disclosure boundary, not a commitment to publish the engine. Independent validation of safety remains open. These records do not establish that AI governance is solved. A poorly formed Manifest produces a poor reference. Coverage and impact analysis remain organizational duties.
Research note
Several recent papers put runtime controls in the harness. They address different problems. CaMeL separates trusted from untrusted data and checks actions against policy. Progent gives agents task-scoped privileges. AgentSpec checks planned actions against user-defined constraints. AI Control studies trusted monitoring of a model that may work against its overseers. Infrastructure for AI Agents describes approval records, logging and accountability infrastructure. These approaches inform how we explore deployer-declared purpose, runtime measurement and inspectable records at TELOS. They do not evaluate or endorse TELOS.
- Defeating Prompt Injections by Design (CaMeL), Google, Google DeepMind and ETH Zurich.
- Progent: Securing AI Agents with Privilege Control, UC Berkeley, UC Santa Barbara and National University of Singapore.
- AgentSpec: Customizable Runtime Enforcement for Safe and Reliable LLM Agents, Singapore Management University.
- AI Control: Improving Safety Despite Intentional Subversion, Redwood Research.
- Infrastructure for AI Agents, led by the Centre for the Governance of AI with other institutions.