insightSynthesisData/thinking-design-v3-harness

thinking_design_v3 — does any multi-turn harness beat single-pass ideation?

Arm outputs, harness turn traces, and judge verdicts for a matched-idea-budget comparison of single-pass ideation against two multi-turn harnesses, on Qwen3-4B-SI and Qwen3-8B-SI over a fixed 100-instance slice.

Layout

path	what
arms/arms_<model>_<arm>.jsonl	3-idea sets per instance. _goldlc = same candidates, gold compressed to the candidate length
turns/twoturn_h{2,4}_{4b,8b}.jsonl	harness turns: turn1/turn2/turn2b prompts + raw completions, vote, drift
judge/<tag>_with_evals.jsonl	raw judge output incl. both orderings' reasoning
judge/h2h_*, h2hfix_*, sj30_*, opus46_*	arm-vs-arm verdicts
impacts.json	compact {condition: {model: {arm: {instance: [impact]}}}} — what the viewer reads
gold_compressed.jsonl	gold compression cache, before/after token counts

Arms

A0 single pass (draw 1) · A0' random 3 of 36 · A0" gte-diverse 3 of 36 · H1 pool→cluster→prune · H2 two-turn propose→select→write · H4 extract→recombine→select

A0" is the bar that matters: beating A0' only shows selection beats random selection.

Read this before using the arm-vs-arm files

h2h_* was produced by a renderer that showed the judge only the inspiration field — one of four, and the one the repo excludes from judging by construction. Those files are invalid for ranking arms and are kept only as evidence of the bug. Use h2hfix_*, the same comparisons re-judged after the fix.

The vs-gold files (gold_*, goldlc_*, sjgold_*) were never affected — run_all_evals renders through _candidate_idea_text(si=True), i.e. SIIdea.card_text().

Impact scale

1.0 both A/B orderings agreed the idea beat gold · 0.5 the orderings disagreed (position instability, not a genuine tie) · 0.0 lost.

Viewer

python thinking_design_v3/view_arms.py --source hf --hf-repo insightSynthesisData/thinking-design-v3-harness

Code: thinking_design_v3/ in the insight_synthesis repo.