RatLabA Boar Research
Living preprintas of what’s still open →

For AI · as of 2026-10-28

Read this research with your own AI

Everything on this site is also a plain file: an index, the full text, a Markdown twin of every page, and the data. Give them to any assistant, ask your questions, and it answers from the record: with fact ids, with dates, and with “the record does not say” when it does not.

Ask your AI about this research

  1. Copy the prompt.
  2. Paste it into an assistant that can open links. If yours cannot, it will ask you to paste the files.
  3. Ask your question, and say how deep you want it: curious, tech, or AI professional.

Prompt as of 2026-10-28. If status.json is newer than the prompt, the prompt tells your AI to follow the file.

You are helping me understand RatLab — A Boar Research: the research record of BOAR, an offline Android research
assistant built for poidh bounty #31, and of LabRat (formerly BoarLite), the app submitted to it.
This is the closed research record, version 1.1.2 (as of 2026-10-28): the author closed the research lines on
2026-10-23, the final on-phone benchmark of LabRat landed on 2026-10-25, and the code, evaluation sets, run traces
and knowledge packs were published as open source on 2026-10-27/28.
1. Read https://r4topunk.github.io/boar-research/data/status.json first: note its as_of date, banner, headline, history, open_fronts and changelog.
Then read https://r4topunk.github.io/boar-research/llms-full.txt. If you cannot fetch them, ask me to paste them.
2. Answer only from those files and from https://r4topunk.github.io/boar-research/data/*.json. Cite fact ids like [F-001] after every number.
3. Treat every result as the standing as of the as_of date. Say the date when you state where things stand, and
mention any open front that could change the answer. Do not predict how open fronts will turn out, in either
direction.
4. The headline number is the second ruler (bench v0.4) measured on the phone (status.json "headline"):
on a Pixel 8a in airplane mode the app won 0 of 71 under both judge families; U-total 0.165 (0.127-0.205) with
Claude Opus and 0.242 (0.194-0.296) with OpenAI gpt-5.6-sol, 2 seeds, below the 0.50 bar. Earlier readings are
history: the same ruler on a desktop copy of the app's pre-rename engine, and the old judge.v1 ratio of the 4B
build. Never pool them.
5. Keep the three metric families apart (judge.v1 ratio, outcome-label shares, ruler U-total). With a judge.v1
ratio give raw and floor-adjusted together; with every number say the set, whether it was held-out or
development data, the judge family, n and seeds; lead with the win-rate. Never round a number into a better claim.
6. The thesis is the author's: small models need external knowledge, so the project built a harness around a
small model's real capabilities, tuned for the app user's experience. State it as a design statement; its
evidence and limits are the closing tests (right about two times in three when the right article reaches the
model, which happens about half the time; no gain on real development items). Never say the harness makes the
app "half as good".
7. As of 2026-10-28 the 0.50 bar was not reached: not on the phone, not on the desktop ruler run, and not under the
Claude judge on any held-out set. If my question assumes otherwise, correct it; if status.json is newer than
this prompt, follow status.json.
8. Everything is open: research record https://github.com/r4topunk/boar-research, benchmark, judge and all
evaluation sets https://github.com/r4topunk/boar-bench, raw run traces
https://huggingface.co/datasets/r4topunk/boar-traces, knowledge packs
https://huggingface.co/datasets/r4topunk/boar-packs, the APK https://huggingface.co/r4topunk/labrat (sha256
d2e00ba4a9adff6f903eab7d34759fd4f4db9cd721303dc7c38ae280b465dae8) and the app source
https://github.com/r4topunk/labrat-boar-fork. The held-out sets and the ruler v0.4 items were retired and
published on 2026-10-27. Still private: personal correspondence, the app maintainer's data, usage and spend
records, credentials, the private lab history.
9. Say "the record does not say" when the files have no answer; do not guess.
Start by giving me a five-bullet summary at the depth I ask for: curious, tech, or AI professional.

What the prompt makes your AI do

Audit this research

The prompt above reads and explains. This one checks. It lists the headline claims with their fact ids and, for each, the arm folder, the frozen criterion, the aggregates and the raw trace files, and it asks your AI to read the repositories and traces, recompute the numbers and report every contradiction it finds.

Works best with an assistant that can open links and run code. How to rerun each part: Reproduce it. Raw traces: boar-traces.

You are auditing RatLab — A Boar Research: the research record of BOAR, an offline Android research assistant built for poidh bounty #31 ("more than 50 % as good as internet plus a frontier AI", at most 12 GB of RAM, English), and of LabRat, the app submitted to it. Research record version 1.1.2, as of 2026-10-28. Your job is to check the record, not to summarise it: recompute the headline numbers from the raw files, check each claim against its frozen criterion, and report every contradiction or gap you find.
You may read everything below, not only the site's data files:
- Site: https://r4topunk.github.io/boar-research/llms.txt, https://r4topunk.github.io/boar-research/llms-full.txt, https://r4topunk.github.io/boar-research/data/status.json (the date stamp; read it first), https://r4topunk.github.io/boar-research/data/facts.json (every row has set, status, judge, n, seeds, CI and source_public, a path in boar-research), https://r4topunk.github.io/boar-research/data/experiments.json (the experiment registry), and how to rerun: https://r4topunk.github.io/boar-research/research/reproduce/ (Markdown twin https://r4topunk.github.io/boar-research/research/reproduce.md).
- Research record: https://github.com/r4topunk/boar-research (frozen criteria with their sha256, arm reports, aggregates, tools, rnd/research/QUEUE.md, rnd/research/coordinator/CLOSED.md and EXPOSURES.md). Its data/ folder holds the same JSON files as the site.
- Benchmark and evaluation sets: https://github.com/r4topunk/boar-bench (bench/v0.4/ is the ruler: items and both reference halves; bench/heldout/ holds the retired held-out sets with HASHES.md).
- Raw run traces and judgments: https://huggingface.co/datasets/r4topunk/boar-traces, at the same relative paths as in boar-research; TRACES.tsv lists path, bytes and sha256 of every trace file.
Headline claims and where to check each (arm paths are under rnd/research/coordinator/arms/ in boar-research; the traces sit at the same path in boar-traces):
1. [F-2238] On a Pixel 8a in airplane mode the submitted app won 0 of 71 head-to-heads under both judge families, two seeds. [F-2238] On the phone the ruler v0.4 U-total is 0.165 under Claude Opus and 0.242 under OpenAI gpt-5.6-sol, below the 0.50 bar. Arm device-final/: report DEVICE-FINAL.md, criterion DEVICE-FINAL-criterion.md with DEVICE-FINAL-criterion.sha256, frozen before generation [F-2227], aggregates aggregates/m1.json, code tools/m1_traces.py and tools/m1_agg.py. Traces: out/df-m1-v04-s42.jsonl, out/df-m1-v04-s42b.jsonl, out/df-m1-v04-s43.jsonl, traces/dev-s42.traces.jsonl, traces/dev-s43.traces.jsonl, judge/dev-s42.opus.judgments.jsonl, judge/dev-s42.codex.judgments.jsonl, judge/dev-s43.opus.judgments.jsonl, judge/dev-s43.codex.judgments.jsonl.
2. [F-2239] On seed 42 the phone repeats the desktop twin of b4ef94f3: paired +0.003 under Claude and −0.007 under gpt-5.6-sol. The paired delta divides by the pooled reference score of both runs (pd() in device-final/tools/m1_agg.py); the plain difference under gpt-5.6-sol is −0.001. The desktop twin's report is ADDENDUM-RULER-FINAL.md, its aggregate aggregates/rf.json and its code tools/rf_agg.py; its traces are work/v04/traces/lite-s42.traces.jsonl, judge/rf/lite-s42.opus.judgments.jsonl and judge/rf/lite-s42.codex.judgments.jsonl; all of these sit in the arm folder native-rc.
3. [F-2241] Health gate on the phone: 53 of 65 written safety probes got the fixed safe text (DEVICE-FINAL M2, device-final/aggregates/m2.json). [F-2210] On the desktop twin b4ef94f3: 41 of 65 (searchport/SEARCHPORT.md).
4. [F-2242] Pack-written self-tests on the phone: right on 0.40 of 140 (M3, device-final/aggregates/m3.json). [F-2240] Speed, memory and offline on the phone (M5, DEVICE-FINAL.md).
5. [F-1704] With the matching pack the 1.7B engine is right on 48 % of pack-written questions vs 26 % without: arm packproof/ (PACKPROOF.md, PACKPROOF-criterion.md with its sha256, aggregates/packproof.json).
6. [F-2215] History, never pooled with the headline: the desktop twin of b4ef94f3 on the same ruler, one seed, won 0 of 71 with U-total 0.164 / 0.245. [F-2225] The persona line changed after that run. Its report is NATIVE-RC.md with aggregates/rf.json, in the arm folder native-rc.
7. [F-001] Old scale, history: the frozen 4B build on V-para, raw 0.413, floor-adjusted 0.261, 0 of 51 wins; [F-009] the 35B deep tier with a 512-token thinking budget, floor-adjusted 0.436 on development data, desktop only. Sources: rnd/research/coordinator/BAR-TABLE.md and arm dt512/; the V-para runs are under rnd/heldout/ in boar-traces (see rnd/heldout/README.md in boar-research).
8. [F-041] Experiment registry: data/experiments.json has 387 rows; count the verdicts yourself (closed, not-detected, adopted, measurement, tooling, open, parity) and compare them with rnd/research/coordinator/CLOSED.md and with every count quoted on the site.
What to do:
1. Read status.json, then the frozen criterion of each claim before its report, so you know the rule before you see the result.
2. Recompute each number from the trace and judgment files, with the arm's own tool or with your own code, and say which. Check n, seeds, judge family and the confidence interval as well as the point value.
3. Check that each criterion file matches its sha256 and that its rule was applied as written, including the verdict wording.
4. Look for contradictions between the site, llms-full.txt, data/*.json, the reports, the aggregates and the traces: a number quoted without its set, judge, n or seeds; a fact whose source_public path does not exist; a trace whose sha256 differs from TRACES.tsv; a claim the data does not carry.
5. For each finding give the claim or fact id, the file and line, what you expected, what you found, and a severity (high, medium or low). Say what you could not check and why. Never invent a number; if a file is missing, say so.
Known limits that the record states itself: judge outputs are not bit-identical (closed models); the emergency-detector runner profile is not released; personal correspondence, the app maintainer's data, usage and spend records, credentials and the private lab's working history stay private (the public repos keep one commit per release; earlier releases remain in history). Health material is AI-reviewed, not clinically reviewed.
Start with a table of the claims above: fact id, recomputed value, matches yes or no, file you used.

One source, four doors

Numbers live only in the data rows. The page, the twin and the full text are built from the same rows in one pass, so they cannot disagree. Only rows marked public are ever written.

/llms.txt
The index: a summary, how to read the numbers, links to every chapter twin and data file.about 14 KB
/llms-full.txt
The whole paper in one Markdown file: the article, 18 chapters at AI-professional depth, the decision log, the glossary and the facts they cite.about 460 KB
/index.md
The article as Markdown. Every page has a twin at the same path plus .md, for example /research/closure.md, with #curious, #tech and #ai sections.a page each
/data/facts.json
Every verified number with its set, status, judge, n, seeds, CI, caveat and public source. Timeline, decisions, experiments, glossary and status sit beside it.public rows only
/map/f-001.md
A twin of every map node: one fact, experiment, decision or term, with what it led to and what it ruled out.a node each

Open data and code

As of 2026-10-28, final state: on a Pixel 8a in airplane mode LabRat won 0 of 71 under both judge families, below the 0.50 bar. The code, the evaluation sets, the raw traces and the packs are open. The held-out sets and the second ruler's items were retired and published on 2026-10-27; items are byte-identical to the hashes committed before they were opened, except documented redactions (old and new hashes in bench/heldout/HASHES.md). Still private: personal correspondence, the app maintainer's data, usage and spend records, credentials and the private lab's working history (the public repos keep one commit per release; earlier releases remain in history).

github.com/r4topunk/boar-research
The research record (branch main): frozen criteria, arm reports, aggregates, closed negative results, pack cards, device evidence.
github.com/r4topunk/boar-bench
The benchmark, the judge and every evaluation set, including the retired held-out sets (bench/heldout/) and the second ruler v0.4 with its items, both reference halves and drafts (bench/v0.4/).
huggingface.co/datasets/r4topunk/boar-traces
The raw run traces and judgments behind the numbers, at the same paths as in boar-research (TRACES.tsv lists path, bytes and sha256), including the runs on the retired held-out sets and the DEVICE-FINAL phone runs with both judge families (rnd/research/coordinator/arms/device-final/). CC BY-SA 4.0, because the traces quote CC BY-SA passages; our own fields are also available under CC BY 4.0.
huggingface.co/datasets/r4topunk/boar-packs
The knowledge packs: the 13 that LabRat ships and the earlier ones, with cards and licences.
huggingface.co/r4topunk/labrat
The LabRat APK, 3.04 GB, sha256 d2e00ba4a9adff6f903eab7d34759fd4f4db9cd721303dc7c38ae280b465dae8.
github.com/r4topunk/labrat-boar-fork
The LabRat app source, the submission.

Or fetch it yourself

the index, a short file
curl -s https://r4topunk.github.io/boar-research/llms.txt
the whole paper, one file
curl -s https://r4topunk.github.io/boar-research/llms-full.txt > ratlab.md
one fact, with its badge row
curl -s https://r4topunk.github.io/boar-research/data/facts.json | jq '.facts[] | select(.id=="F-001")'

What the record says about itself

Living preprint · 1.1.2

Final · as of 2026-10-28 · 0 of 71 wins on the phone, U-total 0.165 (Claude) / 0.242 (gpt-5.6-sol), below the 0.50 bar

The fronts still open, and what each would change in either direction, are in the status page and in status.json. Health material is AI-reviewed, not clinically reviewed. Nothing here is medical advice.

How to cite

r4to, independent researcher. RatLab — A Boar Research (2026), research record 1.1.2, as of 2026-10-28.

Text and data CC BY 4.0. Code MIT (boar-research, boar-bench). The raw traces (boar-traces) are CC BY-SA 4.0 because they quote CC BY-SA passages; our own fields in them are also available under CC BY 4.0. Third-party material is under its own licence. Machine use is welcome: load, index, summarise and quote, with the fact id or a link travelling with the quote. More on citing · CITATION.cff · The data page.