| name | naga-validate |
| description | Scores fidelity of a generated artifact against a source pattern via N1 Zhang-Shasha tree-edit + N4 Salton-Wong-Yang cosine. Returns (score, ci_low, ci_high, N) per the honest-numbers contract. Use when the user runs /naga:validate or asks "how close is this to the source pattern?". Do not use for fingerprinting (see /naga:observe) or for generation (see /naga:match).
|
| model | haiku |
| tools | ["Read","Grep","Glob"] |
naga-validate
Preconditions
- Both
<new> and <source> paths exist and are readable.
- If the source has not been fingerprinted yet, the skill runs
/naga:observe <source> first.
Inputs
- Slash command:
/naga:validate <new> <source>
- Arguments:
new — path to the artifact under validation.source — path to the reference pattern.
Steps
- Resolve or compute the source fingerprint at
plugins/naga-observe/state/patterns/<hash>.json.
- Spawn the naga-fingerprinter Haiku agent with
<new> to produce its feature vector.
- Compute N4 cosine of the new-artifact vector against the source vector.
- Compute per-feature similarities and pass through
bootstrap_ci (1000 iterations) to derive (score, ci_low, ci_high, N).
- Reject the result if N is missing (F11 guard).
- Publish
naga.fidelity.measured with {generated_path, source_pattern, score, ci_low, ci_high, N}.
Outputs
- One
naga.fidelity.measured event on the bus.
- Console table:
path | score | ci_low | ci_high | N | verdict. Verdict is pass if ci_low >= p10_threshold, else hold.
Handoff
If verdict is hold, the developer can re-run /naga:match <source> <new> for a fresh attempt.
Failure modes
- F11 reward hacking — never widen the CI to make
ci_low cross the threshold.
- F02 fabrication — never validate against a fingerprint that does not exist on disk; auto-observe instead.
