Evidence and claims

An agent tells you it fixed the bug and the tests pass. Do you believe it?

You cannot answer that from the sentence. You answer it from the diff, the command output, the test receipt — and the cost of getting to those is what determines whether you actually check or just nod.

Tenon's position: compression is only safe when it is reversible. Every condensed claim keeps a direct path back to the thing it condensed.

The rule

A context capsule and an attention signal are navigation aids. They do not replace the terminal, transcript, diff, command result or test receipt. Those remain the source evidence, always.

So every material claim carries:

• source — exactly which artifact it came from;
• freshness — when that was captured;
• authority — how strongly it is known;
• a return path — a direct way back to the raw thing.

The measured version of this, for the supervision experiment that is not yet built: every material capsule claim must resolve to an exact source identity, an immutable location and hash, capture time, freshness, and authority level. Traceability errors and claims unsupported by their own cited evidence are counted separately, with zero accepted in the reviewed sample.

Reported is not observed

An agent saying it ran the tests, and Tenon watching a test command finish, are two different facts. Flattening them into one row would be a lie of formatting.

Agent Lens keeps them distinct. The timeline shows which claims are reported by the agent and which were directly observed, because those warrant different amounts of trust and the difference is exactly what a supervisor is for.

Folding, not deleting

Finished work collapses into one quiet row so the timeline stays scannable. Expanding it returns every original fact with its stable ID.

Nothing is deleted to make a view tidy. A supervision surface that discards detail to look calm has optimized for the wrong thing — it looks like less work because it is showing you less, not because there is less.

Instructions get the same treatment from the other direction: system, developer, project and skill instructions stay collapsed as context, because they explain the agent rather than record what it did.

Degrade explicitly, never guess

When Agent Lens cannot bind a pane to a provider session with authority, it says so and hands you the terminal.

It would be easy to pick the newest transcript in the sessions directory by modification time. It would usually be right. Tenon does not do it, because "usually right" in an evidence tool is the same as untrustworthy: once you know it guesses, you have to verify everything it says, and then it has saved you nothing.

A stale process, a child-agent fact, a mismatched process group, or a rotated surface token are all rejected for the same reason.

Evidence is required to ask

The clearest place this principle is enforced rather than merely stated is agent.ask.v1.

Its input schema requires evidence, with a minimum of one entry. An agent literally cannot ask you for a decision without attaching what the decision should be made from.

{
  "question": "Two callers depend on the old signature. Break them or adapt?",
  "choices": [ /* … */ ],
  "evidence": [
    {"label": "caller A", "url": "file:///Users/me/app/src/auth.ts"},
    {"label": "caller B", "url": "file:///Users/me/app/src/session.ts"}
  ],
  "recipient": {"kind": "human"},
  "timeoutMs": 55000
}

A question with nothing behind it is an interruption. A question with its evidence attached is a request for judgment, and judgment is the thing you are there to supply.

The same standard, applied to this site

The intent reference here is generated from Tenon itself, not typed out. Where the generator cannot see a contract's schema, the page says so and sends you to the runtime, rather than printing a hand-copy that will quietly go stale. Where the schemas came from a build older than the catalog beside them, the index says which build and which contracts are affected.

Documentation is a compression of a codebase. The same rule applies to it.