Purpose In. Proof Out.

TELOS Reference Architecture · the record of trust, end to end

This is the architecture we are building toward. Signing and public-key checks apply where the signed-receipt profile is switched on, and what each deployment can prove varies. Today the public record is unsigned: you can check the payload hashes and their order. The private review capsule goes further, checking its signed chain against a demo public key included with it.

1 · Purpose In A human declares what the agent is for. Nothing is governed that a human didn't declare. Human authority
1

Write the purpose

The owner writes the agent's purpose as plain documents: what it's for, where it works, where it stops, which tools it uses.

Purpose Anchor (PA) · the corpus IS the spec
2

Sign it

The human authority reviews and signs the purpose. This signature is the root of everything that follows. Without a signed purpose there is no record.

Ed25519 signature · Class-4 human sign-off
3

Compile it

The signed purpose becomes a measurable model the engine can score against: purpose, scope, boundary, and tool, each measured on its own yardstick.

compiled purpose model · embedding centroids
2 · The Run The agent works. TELOS watches and measures, and never blocks or rewrites the work. Observation-only
4

The agent works

Your agent runs exactly as it normally does, on its own runtime and tools. TELOS sits beside the work, not in its way.

any agent runtime · SDK / adapter attach
5

Every action is witnessed

Inside a declared observation boundary, every change the agent makes is either recorded or explicitly marked as a gap, so nothing is silently missed.

recorded-or-gapped · declared boundary · declared residuals
6

Scored against the purpose

Each witnessed action is measured against the compiled purpose across six fidelity dimensions. The output is a score, not a permission.

fidelity cascade · 6 dimensions per action
3 · Proof Out The work becomes a record of trust that anyone can check and no one can rewrite unseen. The record of trust
7

Signed receipt per action

Where the signed-receipt profile is on, each witnessed, scored action becomes a signed receipt, tamper-evident from the moment it is written. Not every surface signs yet, so each one states plainly where its signing stops.

signed receipt · Ed25519 · tamper-EVIDENT
8

Receipts chain together

Receipts hash-link into an append-only chain. Change one byte anywhere and the chain shows exactly where the story broke.

hash-linked chain · append-only JSONL lanes
9

Anyone can verify

An offline verifier checks whatever evidence that deployment publishes. The public record covers payload hashes and their ordering. Checking a signed chain needs the signed payloads themselves, plus verification material you pinned independently.

offline verifier bundle · pinned head anchors it
10

The human reviews

Scores and receipts roll up into plain reports for the human authority. What the review finds becomes corrections to the purpose itself.

reports + replay · human stays the authority
The loop closes:  corrections from step 10 flow back into the purpose corpus at step 1: respec, revise, re-sign. The purpose evolves under the same signature discipline it started with.
What this produces: the record of trust

A durable artifact: a hash-linked chain of receipts documenting what an agent did, scored against the purpose a human declared. Tamper-evident, and checkable offline by anyone holding it, against a head hash they pinned independently. Where the signed-receipt profile is on, each receipt carries a signature too.