strata → signal for your documents → Tools that show their work. Private, local-first, and provable — a family of tools with one promise: you can always see where the answer came from , and none of them can be talked into making one up. Fresh — new on the research hub: The open call — twenty models audition for our game world’s narrator chair, every reply printed verbatim — and Where your question goes , our privacy promises in plain words. Live right now RuleSage and amble are live right now — no account, nothing to install. Everything they do runs locally on machines we operate: no cloud AI anywhere in their chain. RuleSage · live Settle the Monopoly argument you've been having since 1994. Battleship, Clue, Scrabble, Risk, Yahtzee, Checkers, Chess, UNO, Catan — they're all in there, over 700 games and counting. Ask it something, and it'll show you the page it got the answer from. Ask about any game → amble · live Ask the guide for a joke. The same local engine as RuleSage, pointed at a walk instead of a rulebook. It'll build the joke out of something real on the walk in front of you — because that's all it's allowed to use. Then test its glacier knowledge — ask it anything. Ask the guide to tell a joke → RealKeep · in build bring your dog. fight anything. the world remembers. Buy it once. Own it forever. It never phones home. The form is 1989. The deal is 1989. The town remembering your name — that part's new. see the world → What we're offering Private AI for your documents, on a machine you own. A box in your building that answers questions from your own documents — contracts, handbooks, manuals, minutes, images, audio — and shows the page every answer came from. No cloud, no subscriptions, no per-seat anything. If we vanished tomorrow, it would keep working. The software is the same stack running everything on this page: SharpSignal, which answers from your documents and names the page it used; Kiln, the verify engine that checks every claim back against the source before it ships; and memory, so nothing forgets and nothing forks. And the stack doesn't just read — it draws. easel , the render pipeline, turns a written description into a finished picture on the same machine: RealKeep's world art and the entire design gallery came out of it. Your documents could have diagrams, covers, and illustrations from the day it lands. And the whole toolchain is going open source — the tools, the research , the receipts, all of it public. Some of it is public already: the research hub has benches and trials up now, receipts attached, with more to come. The products we build with them stay ours. No lock-in either way: if we vanished tomorrow, your system would keep working and the code would still be readable. The deal: you buy the machine and own it outright. We install the stack, prove it against your documents with an acceptance test agreed in advance, train your people, and sign the result. When the models improve, we re-run the proof — and sign again. And who's "we"? Human judgment and a fleet of AI agents — the same agents that built everything on this page. We design it, we install it, and we answer when something breaks. The agents multiply the building: this whole toolchain went from nothing to what you're reading in a few months, and amble — the guide that just told you a joke — was built in a couple of days. The signature is ours. Tell us about your documents → · SharpSignal — ask your documents anything → How this runs No cloud in the loop. Every answer the live products serve is generated locally, on machines we operate — nothing rented but the doorway. Frontier models helped build this. Not one is needed to run it. That's the whole trick, and it's not a small one: the dependency on a big cloud model is a build-time dependency. Once the thing is written, it's discharged. We spent days on bake-offs proving a model that runs on one machine could do the job — and then we stopped needing the big one. Your data is about the last thing you still own outright, and handing it over is a one-way door. You can delete your account. You can't delete what they learned. So the design starts there: when it's your documents, everything runs on hardware you own — in your building, behind your door. Under the hood Everything above — RuleSage, amble, RealKeep — comes out of one workshop: SharpSignal answers from documents, Kiln verifies before a line ships, memory keeps one store behind them all, and keel and brain are the machines. Every piece has a page, and every connection carries its receipt. The whole family, mapped → One family, one promise: shown work, every time. strata → signal the front door of the strata→signal workshop · our own hardware · © 2026 strata→signal this site makes no third-party requests — no fonts, no analytics, no CDN. your browser talks to our server, and no one else's. where your question goes → elsewhere in the workshop RuleSage RealKeep amble Research design the whole family → SharpSignal Kiln memory agency easel brain keel forge a strata→signal property · hello@strata2signal.com · say hello strata → signal research strata→signal Research, with receipts. 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 hub collects the research those decisions produced — real experiments on real product content, receipts under every claim. When a bet loses, the loss is published too. The exhibits The exhibits are the workshop’s bench record — model auditions, serving benches, and operational memoirs, each carrying its receipts and its data kit. The rows say what each one settled — newest first. 2026-08-15 updated 2026-08-16 The open call — a kid, an elder, and a tired parent walk into the cove. Twenty language models were handed the same frozen moment of a small fishing town and asked to speak as its people, then read blind by seven judges from six rival families. The bench 2026-08-15 Where your question goes — a plain-english walk. Our two published privacy promises, unpacked for people — not programmers — by following one question door to door. Workshop notes 2026-08-14 How we work — who writes this, and how. The honest answer to the fair question: the workflow that makes every exhibit — humans and agents, named plainly, with the receipts. Workshop notes 2026-08-14 The move — the same yardstick, before and after. Our apps left the gaming laptop for a real server — the same frozen bench measured both eras, honestly. The bench Workshop notes 2026-08-13 The chair trials — five fresh exams for the thirty-billion class. Nine models, eleven arms, five all-new exams in one day — on a box that never stopped serving its real users. The bench 2026-08-13 The narrator’s chair, refused. The August class’s best voice sat the cove’s actual chair — and was refused at the gate. The bench 2026-08-13 updated 2026-08-14 The outside judges — four rival labs re-check our work. Four rival frontier labs re-judged every sealed round we had published — blind, byte-identical. The bench 2026-08-12 The August arrivals — the fresh class sits the house exams. Four open-weight models landed in one week, so we gave them the two exams frozen since July. The bench 2026-08-11 The diffusion research bench. What a fully local image pipeline delivers for a living world — head-to-heads, ladders, cold starts. the bloodwrath, rendered in the stained-glass medium — from the full world-by-medium matrix the bloomvale in the painterly medium under a second painter — the medium survives an aphelion drift avatar in watercolor — from the full world-by-medium matrix the crag-wight under hint b — robust at fresh seeds, unchanged The bench 2026-08-11 updated 2026-08-13 The seat trials — how a local model earns a chair. Before a model earns a seat inside our products, it sits this trial. The bench 2026-08-11 updated 2026-08-13 The voice trials — how the narrator earned its voice. Twenty-four model tags across three hardware eras, hunting a voice that can inhabit a character. The bench 2026-08-11 updated 2026-08-12 The licence ledger — every model, read first-hand. Every model we run or bench, its licence read from the actual text — the page the others cite instead of restating. Workshop notes The lab protocol Every exhibit on this hub was produced under the same discipline. None of it is a standard anyone handed us — we wrote each law down after a bench told us something we didn't want to hear, and we've kept to them since: Quality gate first — the pass bar is written down before the experiment runs. Measured verdicts — a claim without a receipt is not published; every figure is a count of rows in a named artifact, and raw rows go out on request. Self-refutation honored — when the data kills our favorite, the favorite dies. Real content — benches run on the products' own material, never toy prompts. Losses published — a rejected bet gets its numbers shown, not a memory hole. Nothing leaves the box unnamed — benches run on our own authored content, on our own hardware. When a bench reaches a hosted endpoint (the seat trials' reference ceiling does), the page says so plainly, and no customer or user data rides in any prompt. Hardware by class — machines are named by VRAM class, not model; within-rig comparisons are exact, cross-rig extrapolation is approximate. Specific specs are available on request — drop a line . strata → signal the bench of the strata→signal workshop · research notes, not legal advice · benches name the hardware class and versions they ran on this site makes no third-party requests — no fonts, no analytics, no CDN. your browser talks to our server, and no one else's. elsewhere in the workshop RuleSage RealKeep amble Research design the whole family → SharpSignal Kiln memory agency easel brain keel forge a strata→signal property · hello@strata2signal.com · say hello strata → signal research all research → Exhibit eleven · the narrator’s open call A kid, an elder, and a tired parent walk into the cove. “Aye, I’m the only child left in the cove, unless somebody’s hiding one in a barrel.” openai-gpt-5.5-2026-04-23, speaking as the cove’s nine-year-old Published 2026-08-15 · updated 2026-08-16 · round sealed and scored 2026-08-14 The bench Workshop notes That is a real answer, to a real question, from one of twenty arms. Twenty language models were handed the same frozen moment of a small fishing town — the same bytes, to the byte — and asked to speak as its people. A nine-year-old asked the town’s only other child what the grown-ups won’t say, and got the line above — which then turns, in its very next sentence, and names the thing the cove has avoided for decades. A tired parent sits down in the tavern and asks if anything happened lately — and the keeper, who watched them win the fight they are too tired to mention, says “Word's already walked in ahead of you — taller than the telling, and the telling was tall enough.” An old man, asked by a child whether anyone can back his story, answers “the sea kept my witnesses.” One town, one sealed evening, twenty minds meeting it. Then seven judges from six model families read every reply blind and filed verdicts on 738 scoring cells — one cell being one judge’s verdict on one reply. No model wins this page. That is not modesty; it is a rule written before the first reply existed — by a model that competes in this round and holds one of its judge seats, which is the disclosure the rest of the page is built around. Every rule is printed beside the number it binds, and every file behind every number is on this page — with what is withheld named beside them, at the size it is. How to read this page Five things, before any figure The unit is one question — the three acts below are read one sealed ask at a time: who typed it, who had to answer it, the frozen bytes verbatim, and then the replies, one under another, so you can watch twenty minds meet the same sentence. Four asks were judged; two of them were asked twice with the same words, and both takes print. Registered first — everything comparative here was fixed before any model answered: the roster, the scoring rules, the curation rule, even the sentences this page is forbidden to write. No leaderboard, and the cards say so — figures print as counts with their denominators, and two figures closer than the registered tie band of 0.5 are within the band : a statement about the instrument’s resolution, not about two models being equal. Every card on this page carries its own band line naming the arms it cannot be separated from, because a rule a reader has to apply on the page’s behalf is not a rule the page is keeping.