A minimal, auditable recursive harness-improvement loop.
A frozen model. An editable harness around it. A recursion that proposes, probes, contrasts and composes — and only judges after a freeze.

Quick start · The loop · Real executor · Results · Research notes · 简体中文

---

nanoRSI v0.5 is a from-scratch, standard-library implementation of the recursion in Harness² (Google Cloud AI Research, arXiv 2610.14224): keep the model frozen, make the harness around it editable, and improve the harness from execution evidence without a verifier in the loop. The whole core is sixteen modules and about 1,400 lines; every rollout, proposal, contrast, composition and judgement is a hash-chained ledger event.

What is editable	How edits are admitted	What stays honest
system prompt · guardrails · memory · tool guidance · skills · scripts	matched-configuration probe → mitigates / neutral / regresses contrast, no ground truth	scores exist only after freeze; budgets charged per call; ledger verify fails on any rewrite

Two tracks live under this name. The engineering track is this loop. The research track is a daily living document of RSI papers, classified on the two axes of the 1,250-paper RSI survey: docs/research/RSI_NOTES.md.

Quick start

git clone https://github.com/xxwtiancai/nanoRSI.git && cd nanoRSI
python3 -m venv .venv && source .venv/bin/activate && pip install -e .

nanorsi new ./lab --goal "Improve the harness for document work"
nanorsi configure --workspace ./lab --model YOUR_MODEL --base-url https://api.example.com/v1 \
  --api-key-file ~/.nanorsi/keys/provider.key        # the key stays outside the workspace
nanorsi doctor --workspace ./lab --check-model

nanorsi evolve --workspace ./lab      # base → INV/DOM proposals → probes → contrasts → compose → final, per task
nanorsi freeze --workspace ./lab      # pin the chosen harness per task; evolve is refused afterwards
nanorsi judge  --workspace ./lab      # score base and final deliverables with the workspace judge
nanorsi report --workspace ./lab && nanorsi verify --workspace ./lab

The starter ships a four-file base harness, two offline document tasks the built-in agent can execute, and a keyword judge — so the full lifecycle runs on any OpenAI-compatible endpoint with no benchmark installed.

The loop

per task                                       mode agg (parallel, width)      mode seq (sequential, depth)
  base1  ── reference rollout                  k pairs probed vs the base,     k rounds; each composition
  INV    ── one cross-domain edit  ─┐          then k compositions             becomes the next base
  DOM    ── one domain edit        ─┤ probe each under matched config
  contrast ─ mitigates/neutral/regresses per probe (no scores)
  compose  ─ adopt / merge / keep, every decision tied to an observation
  final    ─ rollout of the composed harness (a keep reuses the reference)
  select   ─ blind choice over anonymised deliverables when finals > 1
  lesson   ─ one evidence-backed entry into the workspace library

Everything above is a ledger event with its token usage; reports/report.md is a derived view. The paper's two remaining components, subagents and plugins, need a host runtime nano does not ship — the specification lists every boundary.

Attach a real executor

The core never learns what runs a task. Set runner.kind = "command" and point runner.command at a script that reads NANORSI_REQUEST (task, harness directory, rendered always-on text) and writes NANORSI_RESULT (trajectory, deliverables, usage). The judge is the same shape (NANORSI_JUDGE_REQUEST → n_passed / n_criteria). The LAB + OpenCode adapter runs the paper's own legal benchmark through the paper's own executor substrate this way.

Results

Study	Setup	Base → final partial
nanoRSI v0.5 on LAB	this loop (final-confirmation round), OpenCode executor, GLM-5.3 as executor/improver/judge	0.862 → 0.850 (−1.2 pp); averaged −1.5 pp — confirmation killed the collapse losses, contrast verdicts misdirect (2/5)
Original Harness² code on LAB	google-research/harness2, same models	0.872 → 0.892 (+2.0 pp, 5W/1T/2L)
Paper (GLM-5 / OpenCode)	authors' report	0.596 → 0.714 (+11.8 pp)

Earlier studies on the v0.4 core (BBH, ALFWorld, HumanEval, OpenEvolve and ShinkaEvolve tasks, parameter learning) remain published unchanged under examples/results/ as historical evidence; their task packs and runners live in examples/v0.4/ and no longer run on the current core.

Honesty rules

Author results are not local reproductions. Negative and null results are kept. No claim that general recursive self-improvement is solved. Scores never influence selection. Budgets, failed calls and unknown costs stay visible.

License

Apache-2.0. Harness² (Apache-2.0) informed the mechanism shape; no upstream code is copied. LAB tasks (harveyai/harvey-labs, MIT) are used as an installed dependency.