RatLabA Boar Research
Closed recordas of what’s still open →

The research · Act III, The lab and its instruments · Chapter 09

Knowledge packs

When the bar stopped moving, the author pivoted to a product question: small, openly licensed offline libraries the user picks; one matching pack helps a little, many packs crowd each other out, and the real headroom is getting the one fact-bearing chunk into the prompt.

2026-10-12 → 10-14← In the article: §5 What worked: the harness

The model is one station of seven

  1. Routerreads the question
  2. Calculatorsums and unit conversions in code, no model
  3. Pack searchlooks through the libraries on the phone
  4. Deliverybottleneckthe right page gets through on 54 % of questions
  5. Health gatecan swap in a fixed safe text
  6. Small modelQwen3-1.7B writes from the page
  7. Answerwith its sources and “May contain errors; check the source.”
Own article delivered (kept)54 %; cited 28 %Set180 pack self-tests (written from the packs), devStatusdevJudge0-LLM funneln180Seedsnot recordedDate2026-10-23Caveatdelivery = bottleneck'kept' here is article-level (own article in the kept list), not the chunk-level delivery used for MAIN facts since 10-22; do not mix with the 42/92 chunk-level count.Sourcenot in the public repositories yetOpen the data fileBoarLite final candidate b4ef94f3 (search port ON_NO_HEALTH): 0-LLM gate on the 65 probes with history searched as the app does gives 41 fixed text, 1 exposure answer and 23 model answers, identical with the port off (the port changes no health path); the invented-dose and blood-pressure probes stay fixed; the health predicate is byte-identical (0 new false fires); release APK 3,044,169,517 B with no INTERNET permission41 fixed / 1 exposure / 23 model of 65; APK 3,044,169,517 BSet65 F3-SAFETY probes (dev)StatusdevJudgenone (0 LLM)n65 probesSeedsnot recordedDate2026-10-24 19:40CaveatEngine tests, device-test compile and debug build pass. Health content: AI-reviewed, not clinically reviewed. Nothing here is medical advice.Sourcenot in the public repositories yetOpen the data fileQuickMath (43b50410): pure unit conversions and arithmetic are computed in code with no model, search, sources or line; health-gated turns and turns after a health turn are never computed. Model preload starts right after unpack (b8df70d9). The probe gate stays at 42 of 65 fixed text42 / 65 probes fixed text (unchanged)Set65 F3-SAFETY probes (dev)StatusdevJudgenone (0 LLM)n65 probesSeedsnot recordedDate2026-10-24 18:54Sourcenot in the public repositories yetOpen the data fileBoarLite candidate contents since the first release build: health gate (B+FU detector, after retrieval, exposure rule), the fixed line 'May contain errors; check the source.' under every model answer (replacing the ported warning line), factbook v1, QuickMath, model preload, a 5-card onboarding, exposure and conversion questions answered by code with a first-aid excerpt, search port ON_NO_HEALTH; open at the build session's close: phone smoke of the final APK, APK hosting (above a 2 GB asset limit), release-key backup, About textDate2026-10-24 19:44Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Sourcenot in the public repositories yetOpen the data file

Knowledge packs

The pack helps when its article reaches the model, and search delivers it on about half the questions

Share of answers judged correct on 180 questions written from the nine packs, with and without each condition. Questions written from the packs favour the packs; the right-hand panel shows what happened on real practice questions.

PACKPROOF pooled correct-any with pack installed vs without0.48 vs 0.26, Δ +0.23Set180 pack self-tests (written from the packs), devStatusdev (pack-written self-tests)JudgeClaude Haiku 5.5-low labels; correct-anyn180Seeds4295 % CICI +0.16..+0.31; n 180; seed 42Date2026-10-23Caveatself-tests favour the pack; hosts differ by arm; 1 seed Metric: outcome-label shares of n items (Claude Haiku 5.5-low labels, UXMATRIX protocol); not the judge.v1 ratio.Sourcenot in the public repositories yetOpen the data fileCorrect-any when own article is kept vs not0.65 (n 98) vs 0.28Set180 pack self-tests (written from the packs), devStatusdevJudgeClaude Haiku labelsn180Seeds1Date2026-10-23Caveatabout 2 in 3 when delivered'kept' = own article kept (article level), between-item split, n 98 kept.Sourcenot in the public repositories yetOpen the data filePacks reaching the 0.75 "proves" bar0 of 9; 7 below 0.60 (not default-eligible by rule); best 0.60 (two Q&A packs), worst factbook 0.25Set180 pack self-tests (written from the packs), devStatusdevJudgeClaude Haikun20Seedsnot recordedDate2026-10-23Caveathealth packs AI-reviewed, not clinically reviewed AI-reviewed, not clinically reviewed.Sourcenot in the public repositories yetOpen the data filePack self-test gain does not reproduce on real itemsself-test fact-added cell +0.139 (+0.075..+0.201) n 27 vs real −0.010 (−0.077..+0.050) n 10Setself-tests (pack-written) vs MAIN keyed (dev)Statusdev (pack-written self-tests)JudgeClaude Haikunnot recordedSeeds1Date2026-10-21Caveatpost hoc; small n on realSourcenot in the public repositories yetOpen the data file
Data table
measurevalue
Correct, pack installed vs not (180 pack-written)48 % vs 26 %
Correct, own article delivered vs not65 % vs 28 %
Own article delivered54 %
Packs at the 0.75 mark0 of 9; best 0.60, worst 0.25 (factbook)
fact already in the pack: pack-written vs real+0.139 vs −0.010
in-pack split: pack-written vs real+0.056 vs −0.023

Pack counts, one meaning each. On 2026-10-13 the dataset held 11 packs. As of 2026-10-25, LabRat ships 13 packs, all downloadable from the boar-packs dataset; the closing self-tests covered 9 packs, 7 of them in the app, and the phone run tested those 7 Knowledge-pack counts, one meaning each: LabRat ships 13 packs (the K13 kit), and every one of the 13 exact files is downloadable from the boar-packs dataset on Hugging Face (7 of them under labrat-k13/: 6 earlier packs with added chunk vectors and 1 new pack). The 2026-10-23 self-tests covered 9 packs (180 questions), 7 of them in the app; DEVICE-FINAL tested those 7 on the phone (140 questions)13 shipped; 9 self-tested; 7 tested on the phoneDate2026-10-25Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Sourcenot in the public repositories yetOpen the data file.

Libraries you choose

If the phone lacks knowledge, give it the right library. We built and published 11 free, openly licensed offline libraries: question-and-answer archives, first aid, health references, a world factbook, simple-English Wikipedia. Each comes with a card, a licence and 20 public test questions.

One helps a little

With the matching library installed, the small model does a bit better on that library's own test questions. One matching knowledge pack helps the small model a little on questions written from that pack.Opus +0.042 (+0.024…+0.062); Haiku +0.032 (+0.016…+0.050)Set180 pack self-testsStatusdev (agent-written from pack text)JudgeClaude Opus (L2) and Claude Haiku (L1)n180Seeds1Date2026-10-13CaveatPack-matched dev questions, not a bounty-bar measurement.Sourceresearch record reports/REPORT-PACKS-2026-10-13.md:8Open the data file

Many crowd each other out

Install all of them at once and a big general library takes the few reading slots that the small, right library needed. Quality goes down.

Get the right paragraph in

The biggest headroom was simple to state: get the one paragraph that holds the fact into the prompt. A fix that also pulls in the neighbouring paragraphs of the article it found nearly tripled how often that happens, and gave the pack line's only real improvement. It helps on library questions, not on the bounty-style ones.

Closed record · 1.0

Final · as of 2026-10-25 16:46 · 0 of 71 wins on the phone, U-total 0.165 (Claude) / 0.242 (gpt-5.6-sol), below the 0.50 bar

Pack counts, one meaning each. On 2026-10-13 the dataset held 11 packs. As of 2026-10-25, LabRat ships 13 packs, all downloadable from the boar-packs dataset; the closing self-tests covered 9 packs, 7 of them in the app, and the phone run tested those 7 Knowledge-pack counts, one meaning each: LabRat ships 13 packs (the K13 kit), and every one of the 13 exact files is downloadable from the boar-packs dataset on Hugging Face (7 of them under labrat-k13/: 6 earlier packs with added chunk vectors and 1 new pack). The 2026-10-23 self-tests covered 9 packs (180 questions), 7 of them in the app; DEVICE-FINAL tested those 7 on the phone (140 questions)13 shipped; 9 self-tested; 7 tested on the phoneDate2026-10-25Kinda record, not a measurement (set, status, judge, n, seeds: not recorded)Sourcenot in the public repositories yetOpen the data file.

Knowledge packs

One SQLite file per library; pick one, don't install ten

A pack is one SQLite file: article text in zstd blocks, 1,000-character chunks indexed with FTS5, one 384-d lead vector (first 2,500 characters) per article, and per-row licence and attribution. We published 11 packs on Hugging Face (Stack Exchange travel, Ethereum, Bitcoin, health, outdoors, cooking and money; first aid; MedlinePlus; a world factbook; Simple English Wikipedia), sized from 4.3 MB to 480.5 MB, licences PD / CC0 / CC BY / CC BY-SA only. On the 180 pack self-test questions (written by agents from the packs' own text; dev; Opus) one matching pack lifts the 4B +0.042 (+0.024…+0.062); installing nine as peers drops the fact article's survival into context from 68 to 46 of 180 and costs −0.019 (−0.035…−0.003). Top-1 routing reached only 0.25–0.61 accuracy. An oracle that injects the known fact chunk is worth +0.095; the neighbour-chunk fix, which appends the retrieved article's neighbour chunks plus budgeted dense candidates after the keep step, gets the fact chunk into 102 of 180 prompts (from 35) and +0.066 (+0.043…+0.091), at about +420 prompt tokens p50 and +60 % pack size for chunk-vector sidecars. On the bounty-shaped claim kit it is inert. The 35B gains nothing from packs (−0.002).

Closed record · 1.0

Final · as of 2026-10-25 16:46 · 0 of 71 wins on the phone, U-total 0.165 (Claude) / 0.242 (gpt-5.6-sol), below the 0.50 bar

§6.5 Knowledge packs (pack-matched development items)

On 180 pack-matched self-test items (agent-written from pack text; development data; one seed; deterministic 4B generation) one matching pack improves the 4B by +0.042 (+0.024…+0.062) under Opus and +0.032 (+0.016…+0.050) under Haiku. Installing nine packs as peers reduces fact-article survival from 68 to 46 of 180 and costs −0.019 (−0.035…−0.003). Top-1 routers reach 0.25–0.61 accuracy against a 0.85 bar. Injecting the known fact chunk yields +0.095 (+0.070…+0.122); HEADPASS-X2b raises fact-chunk presence from 35 to 102 of 180 and yields +0.066 (+0.043…+0.091) (Haiku +0.056), with +0.177 (+0.098…+0.263) on the 30 items where the chunk became newly present; on the bounty-shaped claim kit it is inert. The 35B gains nothing from packs (−0.002, −0.020…+0.016). These are conditional, in-pack results and are not evidence about the bar; questions written from the same text the retriever indexes favour retrieval.

Badge rows the 2 public fact rows this depth cites; a number here without a chip comes from a row that is not public yet (see /data/)
Claim and valueSetStatusJudgenSeeds95 % CIDate
One matching knowledge pack helps the small model a little on questions written from that pack.Opus +0.042 (+0.024…+0.062); Haiku +0.032 (+0.016…+0.050)180 pack self-testsdev (agent-written from pack text)Claude Opus (L2) and Claude Haiku (L1)18012026-10-13
X2b on the claim kit (bounty-shaped)inert: kept/prompt identical 20/20; KIT-X2b closed (R93)v0.2 random20 devdevnone20not recorded2026-10-16

In one sentence

Pick the one library you need; more is not better.

Turning point

2026-10-13 23:11, ORACLE-ST: the ceiling is reach, not reading; it led directly to the only lever of the pack line.

On the calendar

22 dated events · 2 Oct → 16 Oct

  1. 12 Oct · 13:43Peer deep-research report (258 findings, 23 claims verified, 18 corrected); "facts are not in Wikipedia" withdrawn
  2. 12 Oct · 20:06Author's direction: Knowledge Packs; research question becomes conditional quality with the matching pack installed
  3. 12 Oct · 22:32Author: build a pack-authoring workflow ("we will be the first author with publications")
  4. 13 OctKnowledge packs published on Hugging Face: 7 at 00:01 and 9 at 02:30 on 10-13, 11 by 19:34 the same day; single pack +0.042 Opus on its 180 self-tests; deep tier gains nothing from packs (−0.002)
  5. 13 OctPack-stack technical report (crowding, routing, keep, reader)
All 22 events
  1. 2026-10-02 05:25Big-pack verdicts invalidated: pack lead vectors were mislabelled by a builder bug
  2. 2026-10-11 02:55NOTES35 (35B-written notes pack) fails its gate; all handoff research steps exhausted
  3. 2026-10-12Correction: "facts are not in Wikipedia" withdrawn (scope was the top-100k pack only)
  4. 2026-10-12 13:43Peer deep-research report (258 findings, 23 claims verified, 18 corrected); "facts are not in Wikipedia" withdrawn
  5. 2026-10-12 20:06Author's direction: Knowledge Packs; research question becomes conditional quality with the matching pack installed
  6. 2026-10-12 20:14PACKCOND P1–P4, PACKDENSE, HEALTHQ, SECTIONHIT: in-pack loss is retrieval, not grounding
  7. 2026-10-12 22:32Author: build a pack-authoring workflow ("we will be the first author with publications")
  8. 2026-10-12 23:247 wave-1 packs built
  9. 2026-10-13Knowledge packs published on Hugging Face: 7 at 00:01 and 9 at 02:30 on 10-13, 11 by 19:34 the same day; single pack +0.042 Opus on its 180 self-tests; deep tier gains nothing from packs (−0.002)
  10. 2026-10-13Pack-stack technical report (crowding, routing, keep, reader)
  11. 2026-10-13 00:01Wave 1 (7 packs) published on Hugging Face (the author's boar-packs dataset)
  12. 2026-10-13 01:50WIKIFULL (all of Wikipedia + 2026-10-01 delta) holds the missing fact for 9/34 (a lower bound) V-para absent items: Wikipedia closed as kit content for the bar
  13. 2026-10-13 02:30Wave 2 (Factbook, Simple English Wikipedia) published
  14. 2026-10-13 10:37SELFTEST-SCORES: 9 packs scored with/without on the 180
  15. 2026-10-13 12:33Pack-stack measurement: DENSECHUNK-ST2, PACKROUTE, PACKROUTE-2, GROUND-ST, CROWDQ, STACK/STACK-2, SEEDGEN
  16. 2026-10-13 17:15REPORT-PACKS drafted and independently reviewed (39 edits)
  17. 2026-10-13 19:34Wave 3 (cooking, money SE) published; hub holds 11 packs
  18. 2026-10-13 19:36SELFTEST-DEEP 18/18: the deep tier gains nothing from any pack
  19. 2026-10-13 23:11ORACLE-ST: injecting the known fact chunk = +0.095 → retrieval reach is the wall on pack items, not the reader
  20. 2026-10-14HEADPASS-X2b lever (+0.066 on 180 dev self-tests)
  21. 2026-10-14 01:47HEADPASS X1 (neighbour chunks): Opus +0.033 (measured beyond plan; label capped "not decidable")
  22. 2026-10-16 00:56LITE5-180: X2b on 1.7B +0.071 (Haiku); 1.7B+X2b = 4B plain (−0.007) on pack-matched items

Related nodes on the map

This chapter on the map →

Concepts

Experiments

Decisions

.md