Purpose in. Proof out.
Know what your agents actually did.
TELOS makes the actions recorded on your instrumented paths visible and reviewable. If anyone changes the record, it shows. Check it yourself, offline, without trusting us.
For the people who have to answer for what their agents did.
TELOS is forming the Record of Trust for agentic work, at runtime.
The Record of Trust is a lasting, tamper-evident chain of hash-linked receipts. It records agent actions scored against a human-declared purpose. Anyone can check it offline without trusting the agent, vendor, or us.
How do you know what an agent actually did?
Identity tells you who acted. Logs tell you what was called. Neither tells you whether the work stayed inside the job a human set. The record answers that, action by action.
Five questions, answered for every action.
PURPOSE
What was this for?
What it means: On instrumented paths, actions are scored against the human-declared purpose to show how closely the work matched the job.
SCOPE
Did it stay inside?
What it means: The declared scope of the work. The record shows whether the work stayed within it.
BOUNDARY
Did it reach a line?
What it means: The lines a human said to watch. When an action reaches one, the record notes it and the decision goes to a person.
TOOL
What did it touch?
What it means: The tools and systems the agent used. The record captures which ones the agent touched.
CHAIN CONTINUITY
Does this follow the prior work?
What it means: Coherence with prior submitted work, measured by the similarity of consecutive action embeddings. Coherence can accompany a wrong direction.
Check it yourself.
The record format and verifier are open. Anyone can read the code and recompute the hashes offline.
telos-receipt · PyPI
Mint a receipt and verify it offline. No account, no keys.
pip install telos-receipt
v0.1.0. Runs on synthetic data, observation only. It is the receipt contract, not the calibrated engine.
Try the MCP governance server interface for scoring, receipts, audit-chain queries and counterfactual replay (CCRS).
pip install telos-mcp
v0.1.1, Apache-2.0. Observation only: scoring, audit, and replay use synthetic stubs; no live governance engine is wired. Receipts use the same unsigned payload-hash contract as telos-receipt; a hash match does not prove issuer identity. Server source on GitHub.
Verify it yourself
The public Apache 2.0 verifier uses only Python’s standard library. Run it offline on the included valid and altered receipts to see it pass and fail.
Work with us.
Open to read, open to check.
Licensing
The receipt code is Apache-2.0. The review content and framework mappings are CC BY 4.0. Share and adapt them, with credit.
Published datasets
Validation datasets are archived on Zenodo with forensic audit trails, including the adversarial set (DOI 10.5281/zenodo.21629088) and the XSTest over-refusal set (DOI 10.5281/zenodo.21630399). The full table is in the whitepaper.
Author transparency
TELOS AI Labs Inc. builds and maintains this work. The framework mappings are self-assessed, not independent certifications.
What this does and doesn't claim
TELOS observes and records; it does not block the action. Coverage is scoped to the paths you instrument. Receipts are integrity-hashed and unsigned in this prototype. The record is evidence, not a certification.