Tournament Autoresearch

Tournament Autoresearch Loop

An ML autoresearch loop whose single "form a hypothesis" step is replaced by an idea tournament. The artifact is an experiment ledger; the feedback signal is the **realized ` delta** of the change that won the tournament. Each iteration ResearchAgents propose competing architecture changes, a **Judge** critiques and ranks them, the proposers refine, and the Judge selects **one** change to run. The Judge is the orchestrator and **self-calibrates**: it scores its predictions against realized results, so it learns which kinds of ideas actually pay off. The experiment mechanics (snapshot → run → mandatory analysis → keep/revert) match the sibling ml-autoresearch` loop.

When to use

Use this for open-ended ML experimentation where competing ideas should be vetted before compute is spent and the picker should improve over time. You are the Judge: adopt roles/Judge.md and spawn the proposers with roles/ResearchAgent.md. Default to ` competing proposers with one refine round; widen or add rounds when ideas are converging too fast. Not for running a single pre-decided experiment, and not for analysis-only exploration over a dataset — for one uncompeted hypothesis per iteration use the sibling ml-autoresearch` loop.

The cast and files (all in this folder):

• roles/Judge.md — your behavior: critique, rank, decide, self-calibrate.
• roles/ResearchAgent.md — the proposer role, spawned `` times each round.
• rubrics/rubric.md — the scoring criteria (shipped defaults; copied to a working copy at setup).
• schemas/idea.schema.json — what a proposer returns (one proposed change).
• schemas/verdict.schema.json — what the Judge records per idea (scores, rank, decision).

Setup

Resolve bindings interactively. If loop.run.yaml exists in the working dir, load it, confirm the values in one line, and skip to the loop. Otherwise: on Claude Code (the AskUserQuestion tool is available) infer a likely value for each binding and present it as the recommended option; on other hosts ask each as a quoted plain-text prompt. Then write loop.run.yaml (format: examples/run.example.yaml) and confirm the values before creating any other files.

Record ` (claude-code or other) once — it also decides spawn-or-degrade: on Claude Code spawn real Agent` subagents for the proposers, all in one turn; otherwise adopt the ResearchAgent role inline, one proposal at a time.

| binding | meaning | default | how to infer | |---|---|---|---| | ` | scalar metric to optimize | — | infer from code/README; ask direction | | | minimize or maximize | — | infer from the metric's meaning | | / | command that runs one experiment end to end | — | pyproject.toml/.venv/README | | | model/config/training files the loop may edit; never the eval harness | — | scan for the model + training script | | | where snapshots + ledgers live | ./sandbox | — | | | branches or snapshots | snapshots | git present → offer branches | | / | time (minutes) or epochs, plus the cap | epochs / — | infer epoch arg from the script | | | proposers competing each round | 3 | recommend 3 — competition without a crowd | | ` | propose→critique→refine rounds before the Judge decides | 1 | — |

If ` is branches, create git checkout -b autoresearch/ (must not exist). If is time, write /runwithtimeout.sh (timeout $(( * 60 )) "$@") and hard-kill at 2 × min; if epochs`, cap the epoch count in an editable file.

Initialize the sandbox (after confirmation):

/
├── loop.run.yaml       ← resolved bindings (written now)
├── results.tsv         ← experiment ledger, header only
├── calibration.tsv     ← Judge predicted-vs-realized ledger, header only
├── judge_lessons.md    ← append-only Judge lessons (header only)
├── rubric.active.md    ← copy of rubrics/rubric.md; the Judge self-refines THIS, never the shipped one
└── iter1/              ← created at loop start

Copy rubrics/rubric.md → /rubric.active.md. Write the headers (see [Ledger](#ledger)).

The loop

` = the file capturing training output for an iteration (default /iter/run.log). Everything in is fair game; code must run and finish within . **Simplicity criterion**: equal metric but simpler code is a keep`. The tournament yields exactly one change per iteration. Iteration 1 is the unmodified baseline — skip the tournament; just run + analyse to seed the first analysis summary.

Copy this checklist and tick items off, looping until interrupted:

• [ ] State. branches: git log --oneline -5. snapshots: confirm iter/ is new.
• [ ] Tournament (iter 2+) — run it as the Judge (roles/Judge.md): propose (spawn ` ResearchAgents) → critique & score against rubric.active.md → refine `× → select the single top-ranked change.
• [ ] Snapshot/commit, then apply the winning change.
• [ ] Analysis plan → iter/analysis/plan.md: deliverables table covering the winner's prediction plus useful steps from losing ideas; ≥1 row on a not-yet-measured dimension.
• [ ] Run (redirect, never tee) → read metric.
• [ ] Analyse — execute every plan.md row → iter/results/; check the winner's prediction; write a 3–8 bullet summary ending in the empirical anchor for next round.
• [ ] Log the results.tsv row and the realized delta + hit to calibration.tsv.
• [ ] Keep or revert (simplicity criterion before logging discard).
• [ ] Self-calibrate (roles/Judge.md): update judge_lessons.md, refine rubric.active.md, update the hit-rate.

In detail, each iteration:

1. State. branches: git log --oneline -5. snapshots: confirm iter/ is new.
2. Tournament (iter 2+). As the Judge (roles/Judge.md): spawn `` ResearchAgents

(spawn-or-degrade by `) with roles/ResearchAgent.md, the latest analysis summary, and schemas/idea.schema.json. Critique & score each against rubric.active.md — gate (reject ideas with no testable prediction, unscored) → pointwise 0–5 per axis as the learning signal → **de-biased pairwise to rank** (compare each pair in *both* orders, keep only consistent verdicts). Refine ×, re-score, then **select the single rank == 1 change** (no merging). Write ideas to iter/ideas/, one verdict per idea to iter/verdicts/, and log the Judge's predicted outcome for the winner to calibration.tsv`.

1. Snapshot/commit, then apply the winning change. snapshots: copy `` →

iter/code_snapshot/, copy loop.run.yaml → iter/, apply the change. branches: apply, git commit -am ": ".

1. Analysis plan → iter/analysis/plan.md: a deliverables table including the winner's

prediction and any useful analysis steps from the losing ideas. ≥1 row must cover a not-yet-measured dimension.

1. Run (redirect to `, never tee) → read the metric (grep '^:' `;

on empty, tail -n 50, one trivial fix, else log crash).

1. Analyse — mandatory, real artifacts. Execute every plan.md row → files in iter/results/;

verify none missing; interpret; check the winner's prediction; write a 3–8 bullet analysis summary ending in the empirical anchor for the next round.

1. Log. Append the results.tsv row (0.000000 on crash). Append the realized delta + hit to

calibration.tsv against the prediction.

1. Keep or revert. Improved → keep, update best. Equal/worse/crash → discard/crash

(branches git reset --hard HEAD~1; snapshots restore from code_snapshot/). Apply the simplicity criterion before logging discard.

1. Self-calibrate (roles/Judge.md): update judge_lessons.md, refine rubric.active.md

(weights + anchors, bounded, from realized outcomes), update the selection hit-rate.

1. Go to step 1.

Never stop. Once running, do not pause to ask "should I continue?" — the loop runs until manually interrupted. If ideas run dry: push proposal diversity, mine results.tsv/calibration.tsv for under-explored directions, go deeper on analysis.

Ledger

Three append-only files under ``, all tab-separated, never commas in free text.

results.tsv — the experiment ledger (same format as ml-autoresearch). Header:

iter		status	analysis_summary	description

status ∈ {keep, discard, crash}. Example:

iter	val_acc	status	analysis_summary	description
1	0.6320	keep	baseline; grad norms even, no pathologies	baseline
2	0.6890	keep	layer-2 activations near-saturated; BN helped	iter2-a1: add BatchNorm after conv2

calibration.tsv — the Judge's track record (predicted vs realized). Header:

iter	idea_id	grounding	impact	feasibility	pred_direction	pred_magnitude	confidence	realized_delta	hit

Example:

iter	idea_id	grounding	impact	feasibility	pred_direction	pred_magnitude	confidence	realized_delta	hit
2	iter2-a1	5	4	4	improve	+2%	high	+0.057	1

judge_lessons.md — append-only prose, 1–3 bullets per iteration: which axis tracked gains, what kind of idea was over/under-rated, and the reason for each rubric.active.md refinement. Example:

## iter 2
- grounding tracked the gain (BN tied to the dead-unit finding hit +0.057, as predicted).
- bumped grounding weight 0.40 → 0.45; tightened the impact anchor (impact=5 picks over-promised).

Report the best iteration (highest keep metric), not necessarily the last, plus the running selection hit-rate. Leave results.tsv, calibration.tsv, judge_lessons.md, rubric.active.md, and iter*/ untracked.

Constraints

• Only edit files in `` — confirm before every edit, because everything else

(especially the eval harness) is read-only ground truth defining ``.

• Exactly one change per iteration (the tournament winner) — no merging ideas — so each metric

delta is attributable to one change.

• An idea with no testable prediction is rejected before scoring; the rank is decided by

de-biased pairwise comparison, never by the pointwise scores (which only feed calibration).

• The Judge edits only rubric.active.md (the working copy), never the shipped rubrics/rubric.md.
• Always redirect training output to `; never tee` (it floods your context).
• Do not install packages or add dependencies the project lacks; helper code stays stdlib-only.
• Do not modify the evaluation harness, and do not pause the loop to ask for direction.
• The sandbox must be self-contained — no ../ escapes.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: gaasher
• Source: gaasher/Agent-Loop-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.