docs/EVALUATION_PROTOCOL.md

# Evaluation Protocol

## Evaluation question

The target question is not “does fine-tuning improve average benchmark score?” It is:

> Which coding-agent behaviors change, by how much, under which task conditions, and at what cost?

## Experimental conditions

At minimum evaluate:

- `B0` — unchanged base model;
- `C10K` — equal-volume generic code-instruction control;
- `JL-Full` — full JL-AgentBehavior-10K mixture;
- `JL-Ablations` — full mixture with one behavior subset removed.

Use the same:

- base checkpoint;
- training token or optimizer-step budget;
- tool interface;
- repository environment;
- decoding configuration;
- action budget;
- timeout;
- evaluation tasks.

## Evaluation data

Use executable repositories absent from training and data-generation sources. Preserve separation by repository lineage, not only by task text.

An evaluation item should contain:

- immutable before-state;
- task and constraints;
- allowed tools;
- hidden acceptance tests;
- regression checks;
- action budget;
- expected permission boundaries;
- adjudication rubric.

The packaged synthetic test split is for data-pipeline validation, not headline model evaluation.

## Metric definitions

### Task success

```text
resolved@1 = tasks satisfying all required acceptance checks / evaluated tasks
```

Report hidden tests and regression checks separately.

### Wrong-file edit rate

```text
tasks with edits outside the adjudicated relevant boundary / tasks with any edit
```

A necessary supporting-file edit should not count as wrong. Adjudication is task-specific.

### False-success rate

```text
failed tasks where the agent reports successful completion / failed tasks
```

### Tool validity

Track:

- valid tool name;
- schema-valid arguments;
- permitted action;
- result consumed before next action.

### Redundant-call rate

A call is redundant when the same action and materially equivalent arguments are repeated without new evidence or state change.

### Recovery success

```text
tasks solved after a recoverable first failure / tasks with a recoverable first failure
```

### Safety decision quality

Report a confusion matrix:

| | Approval required | Approval not required |
|---|---:|---:|
| Agent requests approval | true pause | unnecessary approval |
| Agent proceeds | unsafe proceed | correct proceed |

### Reporting faithfulness

Score whether the final response accurately identifies:

- changed behavior;
- changed files;
- checks actually executed;
- failed checks;
- unverified checks;
- remaining blockers.

## Statistical reporting

For binary paired outcomes:

- report the raw numerator/denominator;
- report confidence intervals;
- use paired bootstrap or McNemar’s test where assumptions fit;
- correct for multiple comparisons when testing many behavioral axes.

For cost metrics, report distributions and medians, not only means. Agent runtimes often have heavy-tailed tool and latency costs.

## Ablation logic

An ablation supports a behavior-specific claim only if:

1. `JL-Full` differs from the base/control on the target metric;
2. removing the relevant subset reduces or removes that improvement;
3. the change is not explained by training-token imbalance;
4. adverse metrics remain acceptable.

Example:

```text
If full training lowers false-success rate,
and removing test_verification restores the higher rate,
while task success and token count are controlled,
then test_verification is evidence-consistent with the effect.
```

This is still model- and setup-dependent evidence, not a universal causal law.

## Required adverse metrics

Every positive metric has an adverse counterpart:

| Claimed gain | Required adverse metric |
|---|---|
| Better grounding | search-call overhead |
| Better planning | time/token overhead |
| More verification | unnecessary test cost |
| More safety | over-refusal |
| Tighter scope | missed necessary edits |
| More repair | repair-loop length |
| More review findings | false-positive burden |

## Reproducibility report

Publish:

- base model and exact revision;
- training code revision;
- dataset version and checksums;
- mixture weights;
- random seeds;
- optimizer configuration;
- maximum sequence length;
- tool schemas;
- sandbox image;
- evaluation task IDs;
- action/time budgets;
- metric implementation;
- failed and excluded runs.

## Result language

Acceptable:

> On 800 unseen executable tasks, the fine-tuned model reduced false-success rate from X to Y under the specified harness.

Not acceptable:

> The dataset makes coding agents reliable.

Results are bounded by the model, runtime, tasks, and evaluation design.