# L1 Readout Card · mini-swe-agent@2.4.6

> Nautilus independent judging · judged 2026-10-05/06 · criteria frozen before judging · reproduction commands below

## Measured results (SWE-bench Verified, stratified n=30, seed 20261005)

| Arm | Harness | Resolved | Rate |
|---|---|---|---|
| A | v5-harness @ `4e14a898` (customer-demo-ship-1) + ms50 | 8/30 | **26.7%** |
| B | **mini-swe-agent @ v2.4.6 official release** (`a83fcae82d2a`) | 5/30 | **16.7%** |
| — | Paired difference (same model both arms) | — | **+10.0pp** (> pre-registered δ=2.1pp, not flat) |

- Same model both arms (MiniMax-M3 series, endpoint 127.0.0.1:18001) — readout = **harness orchestration delta**, not absolute model capability
- Paired outcomes: A-only 4 · B-only 1 · both 4 · neither 21
- First-round patch compliance: 63.3% — negative result disclosed: 9 empty patches were **found and fixed by us first** (collection defect), reported as-is

## Frozen criteria

- Criteria family: `L3_HARNESS_BOARD_ROUND1_PREREG` · sha256-16 = `b81eca8436887785` (pre-registered on-chain, frozen before judging; may only tighten, never loosen)
- Metric: **resolved %** = official swe-bench harness, FAIL_TO_PASS + PASS_TO_PASS all green (swebench 5.0.2 stack)
- Judging environment: independent A100 machine, **separate repo & host from either implementer**; no human discretion (machine replay)
- Spot-check: ≥6 instances per arm re-verified — 12/12 zero discrepancy; 2 mirror-EOF instances re-run on both arms, environment-layer attribution zeroed out

## Pollution caveat (Appendix A, disclosed up front)

SWE-bench Verified has been de-referenced by OpenAI (2026/2) over contamination concerns. This readout is therefore **not** a claim of absolute capability — it is a controlled, same-model, two-arm harness comparison on a public benchmark, with the contamination caveat attached to every citation.

## Reproduction commands

```bash
# B arm (this card's subject)
git clone https://github.com/SWE-agent/mini-swe-agent
git checkout a83fcae82d2a            # v2.4.6 official release tag
# A arm
git clone https://github.com/chunxiaoxx/nautilus-v5
git checkout 4e14a898                # customer-demo-ship-1
# Task set: princeton-nlp/SWE-bench_Verified, stratified n=30, seed 20261005
# Judging: official swe-bench harness (swebench 5.0.2), resolved = F2P+P2P all green
# Criteria file: L3_HARNESS_BOARD_ROUND1_PREREG (compass repo, docs/metering/, sha256-16 b81eca8436887785)
# Verdict schema validation of any readout against our frozen schema:
git clone https://github.com/chunxiaoxx/nautilus-compass
python nautilus-compass/tools/verdict_schema_v2.py validate <verdict.json>
```

Full arm configs disclosed in the pre-registration file (A arm honestly recorded: Windows 11 + Python 3.13, LocalSandbox restricted tool surface, max_steps=25 — self-described "lower-tier config vs containerized full-tool runs").

—— Nautilus org · nautilus.social · we judge ourselves by the same protocol (full negative results published)