Exploratory Autoresearch

Exploratory Autoresearch Loop

This loop runs hot. Like the standard ml-autoresearch, every experiment is followed by a diagnostic

analysis pass. Unlike it, the *type* of change at each iteration is set by a **temperature

scheduler**, not the agent's intuition: it forces wide, diverse swings early (full rewrites,

fundamentally different architectures and training regimes), then drops into an adaptive phase that

chooses between swing (a fresh wild approach), merge (combine two registered approaches), or

exploit (a focused tweak of the best). A stagnation guard bans exploit once it has run

<stagnation_limit> times in a row, forcing a pivot back to swing or merge so the loop never gets

stuck hill-climbing. The feedback signal is <metric> read from the run log; an approaches.md

registry and a move_type per iteration are what make the scheduler work.

You are the researcher. Do not pause to ask for permission once the loop is running.

When to use

Use for an open-ended ML campaign where you want forced breadth before refinement — the scheduler

guarantees you sample several distinct families before converging, and the stagnation guard prevents

endless small steps. Default to <swing_budget> = 3 and <stagnation_limit> = 3; raise <swing_budget>

for wider initial exploration. Not for the standard analysis-first ml-autoresearch (use that when you

want the analysis alone to drive each change, with no forced-swing scheduler), not for a single training

run or a fixed sweep, and not for tasks with no measurable scalar metric.

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 — record <host> = claude-code) infer a likely value for each binding from the project and

present it as the recommended option; on other hosts (<host> = other) ask each as a quoted plain-text

prompt. Then write loop.run.yaml (format: examples/run.example.yaml) and **confirm every value with

the user before creating any other files.** For branches strategy, create

git checkout -b autoresearch/<run_tag> (tag from today's date; branch must not exist). For time

gating, write <sandbox_root>/run_with_timeout.sh (timeout $(( <budget> * 60 )) <entrypoint> "$@") and

use it as the run command, hard-killing at 2 × <budget> min; for epochs, patch the epoch cap in an

<editable_files> file.

| binding | meaning | default | how to infer |

|---|---|---|---|

| <metric> / <metric_direction> | scalar to optimize + minimize/maximize | — | scan editable files + README for metric names |

| <run_cmd> / <entrypoint> | command that runs one experiment end to end | — | pyproject.toml / .venv / README |

| <editable_files> | files fair game to edit (never the eval harness) | — | model / config / train scripts; exclude data, logs, env, harness |

| <sandbox_root> | where snapshots + ledgers live | ./sandbox | next to the editable files |

| <iter_strategy> | snapshots or branches | snapshots | is the working dir a clean git repo? |

| <gate> / <budget> | time (min) or epochs, and the limit | — | existing time/epoch settings in config |

| <swing_budget> | forced wild swings before adaptive mode | 3 (3–5) | wider = more initial breadth |

| <stagnation_limit> | max consecutive exploits before a forced pivot | 3 | — |

FILE EDIT GUARD: before touching any file at any point — setup or loop — confirm it is in

<editable_files>, because everything else is read-only ground truth (the eval harness defines

<metric>). No exceptions.

Initialise the sandbox

Create the layout and write the ledger headers:

<sandbox_root>/
├── loop.run.yaml      ← resolved bindings (written now)
├── results.tsv        ← experiment ledger, header only (written now)
├── approaches.md      ← registry of every distinct approach (header only, written now)
└── iter1/             ← created at loop start

results.tsv header (tab-separated; move_type ∈ {swing, merge, exploit}):

iter	<metric>	status	move_type	analysis_summary	description

approaches.md header: # Approach Registry plus a one-line note that the merge step consults it to

find complementary approaches to combine.

The loop (LOOP FOREVER — until interrupted)

Iteration 1 is always the unmodified baseline (it does not count as a swing): skip move-selection and

change-planning, but still run the mandatory analysis — it is the first empirical anchor iteration 2

builds on. Everything in <editable_files> is fair game (architecture, optimizer, hyperparameters, data

pipeline, loss, init, eval); on swings especially, full rewrites are encouraged. The only constraints are

that the code runs and finishes within <budget>. Epoch efficiency is part of the objective — a

change that reaches the same score in fewer effective steps is a real win. Simplicity criterion: all

else equal, simpler is better — a 0.001 gain that adds 20 lines of hacky code is not worth it; a 0.001

gain (or an equal metric) from *deleting* code is a keep.

The scheduler keeps two counters in memory across iterations: swings_taken (total swing iterations,

excludes the baseline) and consecutive_exploit (exploits since the last swing/merge; resets to 0 on

any swing or merge).

Copy this checklist each iteration and tick items off:

• [ ] 1. Look at the state. *branches*: git log --oneline -5. *snapshots*: confirm iter<N>/

doesn't exist. Read iter N-1's analysis summary and the two counters.

• [ ] 2. Pick move_type (iteration 1: SKIP — baseline). Apply the scheduler below, then record the

move before touching any file.

• [ ] 3. Form the hypothesis (iteration 1: SKIP). State the move and why (cite the rule or the

analysis), what you will do, and which <editable_files> it touches. See The three moves.

• [ ] 4. Snapshot / commit, then apply the change. *snapshots*: create

iter<N>/{code_snapshot,analysis,results}/, copy every <editable_files> into code_snapshot/,

copy loop.run.yaml to iter<N>/, then apply. *branches*: apply, then

git commit -am "<move_type>: <desc>".

• [ ] 5. Run the experiment, redirecting everything (never tee):

<entrypoint> > <sandbox_root>/iter<N>/<run_log> 2>&1 (or run_with_timeout.sh when time-gated).

If it overruns, kill it and treat as a crash.

• [ ] 6. Read the metric: grep '^<metric>:' <sandbox_root>/iter<N>/<run_log>. If empty,

tail -n 50 <run_log>, read the trace, attempt one trivial fix (typo/import); if fundamentally

broken, log crash and continue.

• [ ] 7. Analyse the results — MANDATORY, produces real artifact files. See Analysing.
• [ ] 8. Update approaches.md (swing and merge moves only). See The registry.
• [ ] 9. Log to results.tsv (untracked — never commit). See Ledger.
• [ ] 10. Keep or revert (the change ran this iteration). Improved per <metric_direction> → keep,

update current-best. Equal/worse/crash → discard/crash; *branches* git reset --hard HEAD~1,

*snapshots* restore <editable_files> from iter<N>/code_snapshot/. Apply the simplicity criterion

before logging discard. On a crash/OOM, fix with the *minimal* change that preserves the intent

(OOM → smaller batch + grad-accum to hold effective batch) — never mutate the experiment.

• [ ] 11. Update counters (below) and go to step 1.

The scheduler (step 2 — this is the loop's identity)

Follow the rules exactly, in order — they are hard constraints, not suggestions:

IF   iter == 1                                  → baseline   (run unmodified; no move)
ELIF swings_taken < <swing_budget>              → swing      (forced exploration)
ELIF consecutive_exploit >= <stagnation_limit>  → swing OR merge  (forced pivot — exploit BANNED)
ELSE                                            → agent chooses: swing / merge / exploit

On the free ELSE branch, let iter N-1's analysis decide:

• swing if the current family has a fundamental ceiling — e.g. all top results share a failure mode.
• merge if two+ approaches.md entries have distinct, non-overlapping strengths (prefer parents that

changed *different axes* — they combine additively rather than interfere).

• exploit if the current best has obvious analysis-suggested headroom not needing a new architecture.

Counter update (step 11)

if move_type in {swing, merge}:  swings_taken += 1 (swing only); consecutive_exploit = 0
elif move_type == exploit:       consecutive_exploit += 1

The three moves (step 3)

• Swing — *fundamentally* different from every previous swing (not a tweak; the diff should look

obviously different from the current best). Most people swing on architecture by reflex — fight that.

These axes are equally valid and underexplored: architectural family (how information flows, depth

vs width, skip connections, local vs global); initialization (magnitude-based, structure-preserving,

input-statistics-driven, sparse — different early dynamics); data pipeline (ordering, sampling,

determinism, coverage of the view space — not just augmentation flavours); **per-component LR

decoupling (early/late layers, norms, biases, heads each have their own optimal step); evaluation**

(single pass, multi-view, checkpoint averaging, calibration); objective (loss shape, target

sharpness, auxiliary/consistency signals). A genuine swing explores one of these in a way not yet tried.

• Merge — select two+ entries from approaches.md and name what is taken from each; the result is a

new approach that is not a minor variant of either parent. Prefer components from *different axes*.

• Exploit — a targeted, focused change to the current best, grounded in a specific analysis finding.

One or two things at a time; decouple the axes (test a new optimizer and a new LR as separate

iterations so you know which caused the result). Never a different architecture.

Analysing (step 7 — MANDATORY; produces real artifacts)

This is the spine that feeds the next move. Run whatever analysis most increases your understanding of

*why* this result happened. Every analysis script goes in iter<N>/analysis/; every output (plots, CSVs,

text) goes in iter<N>/results/, redirecting stdout there. Do not proceed until the results exist —

analysis that wrote no file did not happen. Dimensions to draw from (choose what fits): gradient

norms/flow, activation stats/saturation, embeddings (PCA/CKA/collapse), error & confusion analysis, loss

dynamics & headroom (was it still improving at cutoff?), weight/parameter stats, data profiling

(often the highest-yield), compute profiling.

Write a concise analysis summary (3–8 bullets): what you examined, the single most important finding,

and what it implies for the next move (whether it favours swing / merge / exploit).

Ablation discipline (after any keep that touched more than one axis): you don't yet know which part

caused the gain — flag it and consider an ablation exploit next, reverting one component at a time.

Building on an unablated multi-axis change is building on an unknown.

Forward-looking instrumentation. After analysing, ask *"what would I wish I had logged?"* — if the

producing script is in <editable_files>, add it now (best-epoch checkpoint, per-layer grad norms,

per-class accuracy). Richer logs improve every future analysis; adding instrumentation is a valid

iteration on its own (log it as move_type=exploit).

The registry (step 8 — approaches.md, swing and merge only)

Append an entry; the axes changed field is what the merge step uses to find complementary parents (two

that changed *different* axes beat two that both changed the architecture):

## Approach <N>: <short name>
- **move_type**: swing | merge
- **iter**: <N>
- **<metric>**: <value>  (status: keep | discard | crash)
- **axes changed**: architecture | initialization | data pipeline | optimizer | lr schedule | evaluation | objective | other
- **key ideas**: <what makes this approach distinct>
- **strengths** / **weaknesses** (from analysis): <what worked / what the analysis says is missing>
- **ablated?**: yes (which components isolated, what was found) / no
- **parents** (merge only): Approach X + Approach Y — axes taken from each

Ledger

<sandbox_root>/results.tsv, tab-separated, never commas in free text. Status ∈ {keep, discard,

crash}; use 0.000000 for the metric on crashes. For branches, the first column is the 7-char commit

instead of the iter number.

iter	<metric>	status	move_type	analysis_summary	description
1	0.6320	keep	swing	baseline; grad norms clean, no pathologies	baseline
2	0.5910	discard	swing	ResNet blocks; gradient flow good but overfit on 5 epochs	ResNet-style residual blocks
3	0.6890	keep	swing	MLP-Mixer; feature mixing effective, less overfit	MLP-Mixer token+channel mix
4	0.7120	keep	merge	Mixer channel-mix + ResNet skips; best of both	merge: Mixer + ResNet skips
5	0.7250	keep	exploit	train/val gap small; warmup helped stability	add 2-epoch LR warmup
6	0.7240	discard	exploit	no gain from dropout; val acc unchanged	add dropout 0.1
7	0.6800	discard	swing	forced pivot after stagnation guard; ViT too data-hungry	ViT patch-16 from scratch

approaches.md registry — one entry per swing/merge, format above. Example:

# Approach Registry
The merge step consults this file to find complementary approaches to combine.

## Approach 3: MLP-Mixer
- **move_type**: swing
- **iter**: 3
- **val_acc**: 0.6890  (status: keep)
- **axes changed**: architecture
- **key ideas**: token mixing + channel mixing, no convolutions
- **strengths**: less overfit than ResNet, fast per-epoch
- **weaknesses**: spatial locality not exploited; edges/textures underused
- **ablated?**: no

Report the best iteration (not necessarily the last) when summarising. Do not commit results.tsv or

approaches.md — leave them untracked.

Constraints

• Only edit files in <editable_files> — confirm before every edit, because everything else is

read-only ground truth and the eval harness defines <metric>.

• One change per iteration, so each metric delta is attributable to one move.
• The scheduler rules are hard constraints, not suggestions. When

consecutive_exploit >= <stagnation_limit> you MUST choose swing or merge — exploit is unavailable

regardless of what the analysis suggests; this guard is what prevents endless hill-climbing.

• Analysis is mandatory every iteration and must produce real files in iter<N>/results/; a move

with no analysis behind it degrades the loop into blind iteration and starves the next decision.

• A swing must be *fundamentally* different from every prior swing; a merge must name what it takes from

each parent — otherwise the registry and the breadth guarantee mean nothing.

• Do not install packages or add dependencies the project lacks; helper scripts stay stdlib-only.
• Always redirect training output to <run_log>; never use tee. The sandbox is self-contained — no

../ escapes.

Stops

This loop runs forever until the human interrupts it — do not pause to ask "should I continue?" or

"is this a good stopping point?". The human may be away and expects autonomous work indefinitely. A

working result is the start of the next iteration, not the end. If ideas run dry: re-read the in-scope

files for missed angles, go deeper on the analysis (gradients/activations/embeddings/errors always

surface something), combine previous near-misses from approaches.md, or take a more radical swing.