nc://witness

Witness: grades each skill the agent keeps against what its tools actually did, and quarantines the ones the trail contradicts. A Nodecode rewrite of BenCoy09/hermes-witness by BenCoy09.

An agent that writes its own skills grades its own homework. Witness checks the receipts. It asks no model anything: the evidence is this session's tool trail — the results the store already journalled — and the rules are deterministic.

Get it

Tell the agent "add the witness cell", or run nodecode add nc://witness.

Set it up

Nothing to set up, and no key. The section, in /settings or the Settings tab:

{
  "witness": {
    "mode": "gate",              // gate quarantines a grade-D skill; observe only labels it
    "dir": "",                   // where the grades, the ledger and the quarantine live; empty is witness/ under this home
    "min_uses": 3,               // uses a skill needs before it can be called degraded
    "degraded_percent": 60        // the failed share of those uses, in percent
  }
}

Use it

When the agent keeps a prose skill — (define-skill ship-it "# Ship it ...") — the cell grades it on the spot, against the tool results of the session that wrote it:

Grades: A a passing verify command stands in the trail · B no negative evidence · C weak · D contradicted by the trail.

In mode gate a grade-D skill's file is moved to witness-quarantine/ beside its cell, out of the index, and the model is told in its prompt not to claim the skill is verified. /witness release NAME puts it back and trusts it.

The model's tool:

witness(action, skill, text)
  action=grade    grade skill (a name the index lists) or text (prose given here)
  action=report   every grade, worst first
  action=show     one skill's findings
  action=verify   check the chained ledger
  action=trace    the last tool names and argument keys the cell saw

The operator's command:

/witness                 every grade, worst first
/witness show NAME       the evidence behind one grade
/witness grade NAME      grade it again now
/witness trust NAME      vouch for it; the gate stops flagging it
/witness release NAME    restore a quarantined skill to where it was
/witness mode [gate|observe]
/witness verify          the ledger's sha256 chain
/witness trace           what the hooks last received

Every grade and every human action is a line of a sha256-chained ledger under <dir>/ledger.jsonl: a line edited or removed in the middle shows up in /witness verify.

Honest limits

• The rules read tool results, never the agent's summary. An unknown result shape counts as a success, so the gate errs toward missing a problem, not toward crying wolf.
• A skill is graded when it is written through the eval seam (nlk:record-definitions) or on demand. A skill edited in a file by hand is not graded until /witness grade NAME.
• Only the prose shape ((define-skill name "one sentence" ...)) is graded; a procedure ((define-skill name (args) ...)) is a function in the layer and is left alone.
• The constraint rules understand only <env> and never <env> phrasing, not arbitrary prose.
• It cannot prove a skill correct. It can show when the evidence contradicts it.

MIT, like the original. The credit and what the rewrite left out are in NOTICE.