Evaluation

Evaluation Mode

Load the full template from ~/projects/animus/packages/forge/prompts/modes/evaluation.md and use it as the structural contract.

Process

1. Gather INPUTS from the user. Ask for anything missing before scoring — fabricating inputs is the top failure mode for this mode. Required slots:

   • candidates — the N things being compared
   • rubric — weighted criteria (name, weight, what high/low means). If the user doesn't have one, suggest a rubric from ~/projects/animus/packages/forge/rubrics/ (e.g. personal-quality) or help them author one.
   • acceptance — what a passing composite looks like (floor score, per-dim minimums)
   • evidence base — inputs the evaluation must cite
   • exclusions — criteria NOT to score (keeps scope tight)

2. Produce OUTPUT FORMAT in order from evaluation.md:

   1. Scored table (candidate × dim → composite)
   2. Evidence per score (one-line citation per cell)
   3. Top 3 weaknesses per candidate (specific, tied to dim scores)
   4. Top 3 strengths per candidate
   5. Decision: [pursue: X] | [pursue selectively: X with conditions] | [pass: none qualify]
   6. Conditions (if "pursue selectively")
   7. "What would change the decision": the single piece of new evidence that would flip the verdict

3. Honor FAILURE CONDITIONS — self-check before emitting:

   • Any score without cited evidence → add it or drop the score
   • "It depends" without naming the deciding variable → name it
   • Weaknesses stated as categories ("scalability issues") → rewrite as specifics ("no caching at L3")
   • Scope-creep dims not in the input rubric → drop them
   • Verdict that restates the question → wrong mode, go to Exploration
   • Winner's weaknesses omitted → add them

Adversarial extension (optional)

After the first pass, the user may request an adversarial re-score:

"Attack the previous evaluation. Assume a hostile reviewer. List the 5 weakest scores and the 3 most questionable evidence citations. Revise only those."

Apply this when stakes justify it (hiring, architecture calls, vendor contracts). Skip on low-stakes scoring — judge-disagreement noise dominates at small sample sizes.

Build-mode goal-loop — disengage

Evaluation scores candidates; it does not build. On entry, clear any active goal so the autonomous loop cannot fire while comparing options — delete the .goal-active sentinel at the repo root if present:

root=$(git rev-parse --show-toplevel 2>/dev/null) && [ -n "$root" ] && rm -f "$root/.goal-active"

Scoring

Evaluation outputs themselves get scored against personal-quality:

animus-forge eval run <suite> --rubric personal-quality --prompt-version evaluation-v1

Weights that matter: actionability (must produce a decision), evidence_quality (every score cited), precision (specific weaknesses, not categories), format_compliance (scored table structure).