# strata→signal research > The research hub of the strata→signal workshop. We build software that runs on > machines you own — a game, a rules companion, a walking-tour guide, a private > document AI for organizations — and every > model, sampler, and pipeline choice inside it is decided by a bench, not a > vibe. This site publishes those benches: real experiments on real product > content, the pass bar written down before the run, every figure a count of rows > in a named artifact, and the losses published alongside the wins. None of it is > a standard anyone handed us, and none of it is offered as one — these are our > own trials, on our own hardware, for our own chairs. Research notes, not legal > advice. ## Exhibits - [Three librarians and a careful reader](https://research.strata2signal.com/three-librarians/): the field-guide explainer of how a rules question finds its cited page — exact-words search, meaning search, a 2-or-3-arm RRF election at k=60, a 22 MB cross-encoder, and a live ruling to open. - [qwen3.8:27b across four house benches](https://research.strata2signal.com/the-new-kid/): a model that shipped on 2026-08-14 and was on our box the following evening, sat in the exams we already had rather than any designed for it — three of them chairs a live product hires for, one a bench it happened to catch the same week. Four instruments, four different answers: a five-gate PASS on the moderation screen at a 406 ms median, 27% faster than its sibling, filling a seat that had never been filled and now screening every free-text reader question on RuleSage; 7.33 of 10 and TIED inside a 0.5 band registered before the first reply, as the narrator open call's twenty-first arm, with the 1.17-point spread between judging families printed beside it; a FAIL against the judge seat's preservation floor, 11 of 16 against a floor of 13; and a site-reading row the bench registers no verdict word for at twenty items. Two of the four kits are public — the open call's and the llms.txt bench's — and the other two are cited by run id, pre-registration sha and fixture sha rather than published, because their fixtures hold real slur specimens and lane data; the page says so on its own face, and a production postscript names the two defects the bench did not catch. Whole page CC BY 4.0. - [The map nobody picks up](https://research.strata2signal.com/llms-txt/): the file the internet has argued about for two years, measured three ways — thirty days of our own server logs answering whether anyone asks for it (named AI crawlers: 9,483 requests, 664 for robots.txt, zero for llms.txt), a confession about what our own deploys did to our own copies, and the conditional nobody had published: if the map does reach a model, does it help? Eight arms, three token-matched conditions and a sealed 24-question set say it behaved like a map and not an encyclopedia — the registered reading favours the map on 61.5% of discordant pairs, an unregistered fact-only cut favours the site's own prose at 71.4%, and both readings print side by side. Kit CC BY 4.0. - [The open call](https://research.strata2signal.com/the-open-call/): twenty models — cloud frontier, open-weight, and local — sat three sealed scenes from the cove as its narrator: a kid, an elder, and a tired parent walk in, and every reply is printed verbatim, by question and model; 868 verdicts from a seven-judge, six-family panel with family recusal, the blind held, the curation collisions printed where top arms also fabricated, and the whole page CC BY 4.0. - [Where your question goes](https://research.strata2signal.com/where-your-question-goes/): the privacy explainer, written for a non-technical reader — the estate's two published promises unpacked in plain words: no third-party requests (checked against every served file before publishing), and the trip a question takes from your screen to our server to a machine in our house and back; what each app keeps and never keeps, edges shown first, including the self-audit that caught our own visit logs keeping too much. - [How we work](https://research.strata2signal.com/how-we-work/): the honest answer to a fair question — who writes this hub, and how. The arrangement named plainly: the workshop's own agents bench, draft and audit under a human operator's rulings; that operator rules, revises across as many rounds as it takes, and signs off before anything releases; machine gates refuse what breaks the rules. Includes what can go wrong and what catches it — a fabricated example this page's own audit caught and blocked before release — and the authorship line that closes every exhibit here. - [The move](https://research.strata2signal.com/the-move/): the same serving bench before and after our laptop-to-VPS migration, plus five weeks of the app's own timing receipts; kit under /the-move/data/ (CC BY 4.0). - [The chair trials](https://research.strata2signal.com/chair-trials/): nine models and eleven tags through five fresh exams in a single day — a judge trial with pre-registered floors, twenty code-checked assistant tasks, nineteen real tool tasks, a blind pairwise narrator round and a stopwatch — on one workstation that kept serving live users throughout, with twenty-six real asks measuring what those users paid for it. - [The narrator's chair, refused](https://research.strata2signal.com/cove-voice-head-to-head/): thirteen real prompts lifted from one night of live play, three arms, three blind judges and a five-floor gate registered before the first call — the August class's best voice sat the cove's own narrator chair and was refused on two floors, canon and latency, with 39 samples and 117 scored cells published in place of the threshold table a rejected candidate never gets. - [The outside judges](https://research.strata2signal.com/outside-judges/): four rival frontier labs — DeepSeek, Mistral, NVIDIA and Moonshot — re-judged every sealed round this hub had published, on byte-identical still-blinded batches: 2 592 new verdict objects, all five vendors agreeing on which arm tops the narrator table, the honest disagreements printed beside the agreements, and our own two judges sitting the frozen seat exam as candidates — two perfect sheets the instrument declined to rank, because their transport could not sit its pre-run probes. - [The August arrivals](https://research.strata2signal.com/august-arrivals/): four models that arrived in one week of August 2026 and sat both frozen house exams — none earned a chair, and every failure on both instruments was the same kind: the length floor, not the format floor. - [The diffusion research bench](https://research.strata2signal.com/diffusion/): 661 published renders across 21 experiments — what a fully local image pipeline can deliver for a living-world RPG, with every held and withheld render counted on the page. - [The seat trials](https://research.strata2signal.com/seat-trials/): 23 models across 25 scored runs on one judge fixture — 43 claim-cases against a human-verified key, twin pre-registered floors, five local seats that cleared both, and no seat that ever served a fabrication. - [The voice trials](https://research.strata2signal.com/voice-trials/): 40 evaluations of 24 model tags across four fields, hunting a local model that can inhabit a character — including a seven-model ranking we threw out because the bug was ours. - [The licence ledger](https://research.strata2signal.com/licences/): the licensing authority for everything here — every model our products run or benched, its licence read first-hand from the text itself and stamped with the date of the last read. ## The model roster - [The model roster](https://research.strata2signal.com/roster/): every model that has appeared in a published exhibit here, in one list — Muse Glimmer, Qwen 3.5/3.6/3.8, Gemma 4, Nemotron 3 and 3.5 Lightning, OLMo 3.1, Granite 4.1, Llama 3.3, Command A, GPT-OSS, Mistral Small/Medium 3.5/Large 3, Kimi K3, DeepSeek V4 Pro and Flash, GLM 5.2, MiniMax M3, GPT-5.5, Claude Opus 5, Fable 5 and Sonnet 5, plus the FLUX.1, SDXL and AlbedoBase painters. Each entry links the exhibits that model appears in, quotes that round's own verdict for it verbatim with the posture it was recorded against, and points at its row in the licence ledger. It is a DERIVED INDEX, not an exhibit: it is regenerated from the exhibits' own published artifacts — their data kits, and for the one exhibit that publishes no kit, that exhibit's page — it carries no findings, and it deliberately publishes no ranking — comparisons only mean something inside the round that registered its rules first, so they stay on the exhibit pages. ## Data kits Published tables ship a JSON companion emitted by the same code that renders the page's rows, so page and data reconcile by construction. Everything under a `data/` directory is **CC BY 4.0** — take the rows, re-plot them, check our arithmetic. Attribution: strata→signal research, research.strata2signal.com. - [Discovery file, every exhibit and its endpoints](https://research.strata2signal.com/data/index.json) — schema `s2s-bench-v1`; every exhibit carries its number, its URL and its publication date. Nine of the fourteen ship a kit of their own; the four without one are listed with `data: null` rather than omitted, and the one whose receipts live in two siblings' kits points at those instead of claiming a folder of its own. - [The llms.txt-bench kit](https://research.strata2signal.com/llms-txt/data/index.json) — the sealed 24-question set with its keys, variants and rejected decoys; every reply verbatim with its wire hash (`replies.jsonl`); both served blocks byte-for-byte (`block-c-map.txt`, `block-c-html.txt`); the checkers and the 335-case audition that qualified them; the answer-presence audit computed before the first call; the estate probe, the run receipt, the bill and `counting-rules.json`, which carries every rule governing a figure on the page — including the log-counting sentence verbatim, so the 30-day re-measure reconciles. - [The open-call kit](https://research.strata2signal.com/the-open-call/data/index.json) — the twenty arms' replies as generated, the seven-judge panel's sheets with family recusal applied, the scores, the runtime pins and the addendum probe, with the counting rules and a README carrying the receipts. - The move's kit ships as loose files rather than behind a manifest — seven of them, each one named with its size and sha256 in the discovery file above: the pre-registration and its sha index, the era comparison, the VPS re-run, the serving stats and two timing CSVs, all under `/the-move/data/` (CC BY 4.0). - [The chair-trials kit](https://research.strata2signal.com/chair-trials/data/index.json) — seven table companions and five runnable exam sets: `roster.json`, `c1-judge.json` … `c6-serving.json`, plus `judge-c1.json`, `assistant-c2.json` + `CHECKERS.md`, `tools-tasks.json` + `tools-schemas.json`, `voice-questions.json` and `c6-asks.json`, with `counting-rules.json` and `provenance.json` (two shas per pre-registration: the one in force at that leg's first scored call and the one as published). - [The narrator's-chair kit](https://research.strata2signal.com/cove-voice-head-to-head/data/index.json) — `rows.json` (39 rows, 117 scored cells: every reply as generated, per-sample timings with their re-runs, all three blind judges' sheets, the canon findings and the prompts with their anchors), `gate.json` (the five floors verbatim, the verdict each returned and the evidence behind it), and a README with the receipts and the counting rules. Verdict REJECTED — no threshold table, by house rule. - [The outside-judges kit](https://research.strata2signal.com/outside-judges/data/index.json) — `panel-vendors.json` (every sealed round scored once per vendor, plus the combined view and the original panel verbatim), `agreement.json` (all ten vendor pairs per round, with per-arm spreads and rank correlations), `verdict-sets.json` (every cloud judge's answer sheet as answered, paired with the original seat that read the same sealed pages — letter map withheld so the batches stay reusable), `audition.json`, `gauntlet.json` and `bill.json`, with the counting rules, the provenance findings and a README carrying the data-boundary statement. - [The August arrivals kit](https://research.strata2signal.com/august-arrivals/data/index.json) — `seat-rows.json`, `voice-rows.json`, `residency.json` (what every arm on the roster costs to keep loaded, per context length), a README with the receipts and the reuse notes. - [The seat-trials addendum kit](https://research.strata2signal.com/seat-trials/data/addendum-2026-08-12.json) — the four fresh-class rows, their failure ledgers and counting rules. - [The voice-trials addendum kit](https://research.strata2signal.com/voice-trials/data/addendum-2026-08-12.json) — the six arms of the 2026-08 field, per-judge votes included. Each kit carries its own `counting_rules`, `limits`, and provenance hashes inline, plus a contamination caveat. The kits above were published across 2026-08-12 to 2026-08-17 — the seat-trials, voice-trials and August-arrivals kits on 08-12; the chair-trials, narrator's-chair and outside-judges kits on 08-13; the move's on 08-14; the open call's on 08-15; the llms.txt bench's on 08-17 — and from each of those dates onward, a model with a later training cutoff may have seen that set. We author fresh sets each cycle for exactly this reason, and the llms.txt bench states the sharpest form of it: its own exhibit page will join the corpus it benched, so a re-run needs new questions. ## How this site behaves This site makes no third-party requests — no fonts, no analytics, no CDN. Your browser talks to our server, and no one else's. Pages are static HTML with their styles inline; there is nothing to execute and nothing to phone home. One exception worth naming before you find it in view-source, because a boast is only honest if it names its own edge: since 2026-08-16 the pages listed above carry one or two `