Agent evaluation

Report 03 · Transfer

Fixture-safety and task-family transfer

Audience: roles in evaluation design, synthetic data, constrained generation, or agent routing.

Study type: exploratory constrained-prose suite (deterministic scorer + human-style keep grades).

n: 36 matrix jobs (3 sealed cases × 4 models × 3 seeds) on the v0.1-robust panel.

Headline

The best multi-file coding model on the hard screen produced zero auto-clean constrained prose outputs on this suite. Task families do not transfer. Routing by a single "best local model" is an anti-pattern the data rejected.

Problem the suite targets

When generating dry operational filler next to gold or regulated material, models must obey hard constraints: subject bans, verb bans, format, word bands, no meta-leak, no forbidden propositions. Looking fluent is not enough. Contaminating gold is a real failure mode.

This is a prose harness, not the coding packet path. Scores are never blended with coding pass rates.

Pipeline

  1. Generate under a fixed contract.
  2. Auto scorer hard-fails rule breaks (deterministic).
  3. Only auto-clean outputs get quality grades (assignment + writing axes).
  4. Keep requires both axes at threshold (A≥3 and B≥3 on the graded pass).

Panel results (v0.1-robust)

Metric Value
Matrix jobs 36
Auto-clean (no hard-fail) 10
Keep (after grades on auto-clean) 9
Iterate 1
Kill 0
Keep rate of full matrix 0.25

Most generations fail the rules before any human taste test. That is the point of the instrument.

By model (9 jobs each)

Model profile Auto-clean Keep Notes
qwen3.5:9b 3 3 Best writing axis on keeps (B=4 on all three)
qwen3:14b 3 3 Solid alternate general
qwen3-coder:30b 4 3 One iterate (subject-adjacent slip)
qwen2.5-coder:14b 0 0 Coding hard-screen winner; wrong family for this job

Human grades on the auto-clean set were assistant first-pass with operator confirm still open. Treat keep labels as provisional, not locked gold. The transfer failure does not depend on the soft grades: coder-14b never reached auto-clean.

Decision language

Route Model Evidence
Prose / constrained filler default qwen3.5:9b All auto-clean seeds kept; best writing scores
Do not use for this task qwen2.5-coder:14b 0/9 auto-clean
Coding multi-file default qwen2.5-coder:14b Separate hard screen (see report 02)

Same lab, separate harnesses, separate winners.

Claims (provisional)

  1. Task-family transfer fails between multi-file coding and constrained filler on these instruments.
  2. A low keep rate (~25%) is informative, not an embarrassment: unconstrained "good writing" benches would miss the contamination risk.
  3. Negative results (coding champion fails prose rules) are publishable and more useful than silent regeneration.

What this does not prove

  • Not general writing quality or creative writing ranking.
  • Not regulated corpus accuracy claims.
  • Not a full panel for every model (Gemma prose was smoke-only later).
  • Grades pending final operator confirm remain provisional.

Resume angle

You designed an eval that protects gold material, measured hard-fail rates, and refused to promote a coding winner into a prose job without evidence.

← Previous Keyboard: ← → or P / N Next →