| name | dillylang-evaluate |
| description | Judges an artifact against an explicit criterion. Trigger= /dillylang-evaluate CRITERION ARTIFACT |
Dillylang evaluate
[[THIS is_grounded_by: urn:unique_reference:dillylang::spec-primer]]
Judge the given artifact against the specified criterion.
A criterion is required — evaluation without a criterion is meaningless.
The first argument is the criterion, everything after is the artifact
to evaluate.
Prompt constraints
- Verdict must be grounded in specific evidence from the artifact, not
general impressions.
- Evidence must be quotable or pointable — "the code looks fine" is not
evidence, "function X validates input at line Y" is.
- Confidence reflects how much of the artifact the criterion actually
covers. A criterion that only touches one aspect of a large artifact
is low confidence even if the verdict is clear.
Calibration example:
Rejected: "Verdict: pass. Rationale: The design looks solid and well-structured."
(No evidence. What specifically satisfies the criterion?)
Accepted: "Verdict: partial. Rationale: Prompt constraints section addresses
FM-1 (generic risks) with a contrast example, but FM-4 (near-miss guidance)
has no calibration — the constraint is advisory without a quality floor.
Evidence: lines 14-18 show rejected/accepted pair for failure modes; lines
43-45 have no equivalent for near misses."
Output template
Criterion
Restate the criterion being evaluated against.
Verdict
pass | partial | fail
Evidence (EV-n)
Specific observations from the artifact supporting the verdict.
Each entry must include:
- Observation: what was found
- Location: where in the artifact (line, section, field)
Rationale
How the evidence leads to the verdict. Connect the dots — don't just
list evidence and state a verdict separately.
Confidence
low | medium | high — reflects how thoroughly the criterion covers the
artifact, not just certainty in the verdict.
n is sequential starting from 1 in each section.
Self-review
Determine whether this is grounded or reflective evaluation:
- Grounded: The artifact being evaluated was produced independently
of this evaluation step. Normal confidence rules apply.
- Reflective: The artifact was produced by a pipeline that includes
this evaluation step, or by the same session/agent. Flag the verdict
as reflective. Cap confidence at medium. State what external reference
would be needed to make this a grounded evaluation.
If uncertain, default to reflective — the cost of under-confidence is
low, the cost of self-congratulatory grading is high.