Every piece of it: what each one does, a link you can click, and where it lives right now. Nothing here is a proposal. Nothing has been changed.
So what. The manual you called slow was one page of nine megabytes, and 44% of it was an animation engine that no part of the page ever switched on. The rebuild cut it into 44 pages that open fast, and replaced the pretend-interactive figures with six pages that take numbers you set and work out an answer. The catch, and the reason this page exists: the 44 pages are live, the six computed pages are not published anywhere. They sit in the shared code repository and on the M4 laptop. That is the thing you have never seen.
Each takes the facts you set and works out the answer the manual's own rules give, then names the rule that decided it. Every one has a test file that proves it refuses a wrong answer, not just that it runs.
The frame builderAssemble one frame of intent — what it is for, who may act, which resource, constraints, horizon, proof — and the page runs the eight rules in order and refuses it at the first that fails. The image open on your M4 is this page.§8 Frames · 18 checks · 114 assertions · 60 KB · the one you photographed The promotion benchA shortcut learned from logs asks to decide live. The page puts the 95% interval on the improvement, then runs the six conditions in order and stops at the first that fails.§20 Policies and Learning · 30 assertions · 30 KB The capability checkOne request, one verdict: allowed, needs an approval, or refused — plus the missing bit and who repairs it. Checks run in the manual's order and the skipped ones say why.§5 Identity and Capability · 139 assertions · 50 KB The funnel maskA search request narrowed column by column until one record class remains, naming the column that emptied the result and the one repair that brings it back.§15 Record Search · 179 real capability rows · 82 KB The resolution loopTwo observations disagree: which one wins, on which rule — or the single piece of evidence that would settle it. Six stages, each showing its arithmetic.§100 The Resolution Loop · 5,362 assertions · 31 KB The Decision LadderSeven facts about a real decision in; out come the rung that can settle it, the authority that must sign, every rule that fired with its numbers, and the training record it writes.§13 Decision Systems · the reference standard · 22 assertions · 17 KBOnly reachable from your own devices, which is where it has been since 25 October. A chapter reads well on a phone; the tables become cards below 600 px wide.
| Piece | Where it is now |
|---|---|
| 44 manual pages | live manual.berkle.space/latest/ · pages built from the shared code repository |
| 6 computed pages | repository only then this copy — never published until now; the links above are a copy I put up so you can open them |
| The one-page original (rev22) | repository research/fabric-timeline-sketches/rev22/index.html — 9.1 MB, kept frozen on purpose so every split chapter can be checked against it |
| The measurements, the plan, the task sheet, the name register, the claims register | repository rev23/AUDIT.md, PLAN.md, TASKS.md, NAMES.md, CLAIMS.md |
| The four checkers that decide what may ship | repository weight, fidelity, interaction, and house-style — each also proves it fails a deliberately wrong page |
| 8 screenshots of it running | repository rev23/captures/ and rev23/*.png — phone, wide, catalogue at 3440 px, glossary open, the merge rundown at 390 and 1366 |
| A revision 24 | does not exist — no folder, no branch, no pull request, as of today |
Filled from the transcripts on the M4 and this Mac; each row names the session it came from.
| When | Session and thread | What it delivered |
|---|---|---|
| 23 Oct, 13:14 | 01a168e5-f400 "GOOSE: Evaluate Goose as Tricon substrate" — the root of the chain, in the worktree s-20261023T131422Z on the M4 | the worktree everything below was built in. Seven of the eight sessions in it carry this same title with "(curated)" added each time the thread was forked: the whole documentation rebuild is filed under a thread named after a different question — whether to use the "Goose" agent tool as a base. That is why you never found it by name. |
| 24 Oct, 17:08 → 25 Oct, 04:29 | 01a16ee2-b772 → 01a1703a-f1ce → 01a17113-6dad (plus forks 01a170aa, 01a1714f, 01a17151) four generations of the same forked thread | 01a17113 is the builder. The audit, the plan, the split into 44 pages, and all six computed pages: the six page builds run in its own rundown file, and its browser log shows the frame builder open at a local address. Its first instruction was a "gk-prompts battery" run — five writing formulas answered, then executed with real tools. |
| 25 Oct, 15:34 → 27 Oct, 10:01 | 01a173b2-9cbb "DOC: Clean up GK doc ui" | the pass over the manual's look — the one you now find part incredible, part hard to read — and a later re-touch of the frame page. The likely author of the last pull request, #3259. |
| 25 Oct, 21:40 | 01a17501-b44b "throughput rebase worker", this Mac | the merge that got the 44-page manual through: 62 conflicting files, all under the rebuild folder, all resolved by regenerating the built pages, then the project's own quality check came back clean. Pull request #3232 merged at 21:54 that evening. |
| 25–26 Oct | three more rebase workers on this Mac (01a172a5, 01a17318, 01a177d8) | carried the sibling branches through the merge queue; one of them wrote the merge record for #3232 at 21:54:49 |
Nine pull requests landed this work. In order: #3150 the audit, the plan and the first computed page · #3175 whole-manual search · #3194 and #3195 three plain sentences at the top of every page · #3196 then #3197 — the first time the manual was reachable by URL at all, at its older address under artifact.berkle.space · #3200 the publish checker made ten times faster · #3232 the 44 pages and the move to manual.berkle.space, which is the address you are looking at · #3259 the contents menu and glossary. One earlier attempt, #3196, was superseded the same hour by #3197.
The rebuild sits inside a folder that holds much more. All of it is in the shared code repository; none of it is published anywhere. Counted from the repository today.
| What | What it is | Size |
|---|---|---|
Six earlier versions of the manualrev17–rev22 | Each one a complete earlier build of the same book as a single page, oldest smallest. They are kept on purpose: every page of the rebuild is checked against the last one, so nothing can drift by being split. | 885 KB → 9,126 KB each 54 MB over six folders |
The manuscriptOperator-Manual-rev15…22.md | The book as text, seven versions. Version 22 is the one the rebuild is cut from and may not be edited. | 226 KB → 685 KB each |
A working prototypeapp/ + proof/ | Eight small files that draw a scene, cards and an insight panel around this material, plus 38 screenshots that prove each state it can reach — card agreement, focus, key handling, overlap. Nobody has shown this to you. | 0.1 MB + 10 MB of proof |
A conversation minerminer/ | Reads past agent sessions and builds an index of them, with a 205 KB report on what it found. This is where the material for the book's claims was gathered. | 1.2 MB |
Two data bundlesbundles/ | Session material frozen as JSON, one of them 3.6 MB, the other 289 KB. | 3.9 MB |
Three pictures of the timelinesummagator3d.html, log-summagator.html, timeline-volume.svg | A 3-D view of the project's timeline, a summary of one long log, and a printed volume diagram. Self-contained pages, never published. | 0.1 MB |
The sheet has 59 rows: 44 finished, 15 not. Eleven of the fifteen have a decision of yours inside them; the other four are already closed — one is a deliberate decision not to act (C4), one is measured and ready (D1), two are findings written down with nothing asked for (F7, F27). None of the eleven is about the manual's words being wrong; seven are the publishing checker refusing the manual for reasons that are not the manual's fault.
| Row | What it says |
|---|---|
| D1 | Every code path, command and decision record the manual names should resolve. Measured: 45 of 47 paths resolve; nothing else outstanding. Marked ready, not done. |
| F8 | The checker treats the words "helpers" and "simulated only once its post was observed" as jargon the glossary must explain. The fix is a rule change to the glossary's shorthand word list — your file, not the manual's. |
| F13 | The publishing checker reads the manual's own prose and refuses it, because a specification that quotes a frozen text always trips its "two opinions" rule. Either the manual's text may be edited — which breaks the thing this work exists to protect — or the checker gains a tier for quoted specification text. |
| F14 | The checker treats each numbered clause (GRC-10) as an unexplained word, so the glossary grows one entry per clause the manual prints. |
| F16 | The manual follows your Mac's light/dark setting and draws a light theme; the house style is the dark night field. Needs a call. |
| F17 | The checker lists vocabulary a reader may not know, and the glossary answers it one word at a time. The class fix is yours to decide. |
| F18 | A phrase the checker learned from one page blocks every other page that uses it, including a specification quoting it. |
| F19 | One sentence the checker marks wrong refuses an entire release. |
| F20 | The version note the publisher writes is itself judged by the checker, on a page where no glossary can explain it. |
| F21 | Third round of the checker: three must-fix rows, none about text the manual owns. Same request as F13 and F19 — a page that reproduces a frozen document cannot be judged on that document's prose. |
| F22 | Fourth round: one must-fix row against a figure's own caption inside the frozen original. |
| F25 | A figure says "step 1 of 9" while the loop is drawn as five stations, so a reader cannot tell if the loop has five parts or nine. Glossary done; the caption waits on you. |
44 of the 59 rows are done, each with the proof it names. Worth knowing in plain words:
The full sheet, one row per task with its proof: TASKS.md. The audit: AUDIT.md. The plan: PLAN.md.
I added no words to any page of the manual, changed no style, published nothing of the manual, and deleted nothing. The only thing I made is this page and the copy of the six computed pages beside it, so you could open them. Say what you want kept, deduplicated or thrown out, and it gets done to your call, not mine.