Four working modules in the Fabric folder (920 lines) hold the missing pieces Semantic Tricon now needs; the other 38 manual pages and 72.8 MB of copies are dead weight.
Who benefits: You stop treating a 44-page book as a build plan and wire only the four missing pieces into Semantic Tricon's live Rust turn path.
Second-order effect: The Frame Builder on your M4 becomes Tricon's real confirmation card, and its 75-byte parcel becomes multiplayer sync across M3, M4, and M6.
Risk if left alone: Agents keep polishing 11 prose-linter complaints on a frozen book while Tricon stays stuck on a 4-bit world mask.
Recommendation & decision: Freeze the 44-page manual, drop the 72.8 MB of old copies, and build the 4-step stack rank below — proved by one Dock demo.
Every row below was checked in the repository and in the manual's own source-conformance chapter (implementation-map.html). None of these produces runtime value for Semantic Tricon.
| Item | Size / count | Why it is not valuable — proof |
|---|---|---|
Six older single-page builds of the manualrev17–rev22 + Operator-Manual-rev15…21.md | 57.3 MB | Each is a full copy of the same book (885 KB → 9,126 KB) kept only so rev23/gates/fidelity.ts could diff chapter splits against rev22. Keep Operator-Manual-rev22.md (685 KB) if you want the text; the other 56.6 MB are historical build outputs. |
Inline Manim animation engine inside rev22 | 4,010 KB per page | rev23/AUDIT.md measured that 44.1% of the 9.1 MB manual was a single inline script that mounted 0 KB of live figures at runtime (zero live mounts in its own audit table). |
Duplicate "anatomy" interactive pilot pagessystem/gk-plane/docs/anatomy/ | 5.8 MB | These are the earlier "37 chrome buttons, 1 slider" pages you called horrific. The six computed pages in rev23/instrument/ replace them and pass gates/interaction.ts, which explicitly fails the anatomy pages. |
Toy cards-and-scene prototype + 38 proof PNGsresearch/fabric-timeline-sketches/app/ + proof/ | 10.1 MB | A static mock of cards and an insight drawer built before semantic-tricon shipped its real 3-screen KokonutUI workbench (PR #85–#88, live at tricon-workbench.berkle.space). Superseded by real software. |
Frozen session JSON bundles & 3-D timeline sketchesbundles/, miner/, summagator3d.html | 5.2 MB | Point-in-time dumps used to write the book's prose. Meanwhile semantic-tricon's real nightly miner (mine_tricons, PR #74) mines live sessions every night into ~/.local/share/tricon-workbench/store/. |
| 38 of the 44 manual chapters as a "build spec" | 38 pages | The manual's own Implementation map (GRC-101 crosswalk) admits that chapters on Path Search (C6), Context View Compiler (C7), Circuit Promotion (C8), Control-Plane Sequencer (C9), and Console Sync (C10) have "None — no inspected source", while chapters on P1–P5 projections, trigram search, and the 25 workflows are already surpassed by running Rust code in semantic-tricon. |
10 of the 11 open task rows on rev23/TASKS.mdF8, F13, F14, F17–F22, F25 | 10 rows | All ten are the prose-publishing linter arguing with the frozen text of the manual (treating GRC-10 or quoted sentences as violations). None of them touches Semantic Tricon or changes a single line of working code. |
Why the fragmentation felt overwhelming: two tracks ran in parallel without a shared map. Track 1 (semantic-tricon in Rust, live on gk-m6-01) built the fast recognition and execution engine. Track 2 (fabric-timeline-sketches in TypeScript) built the world model, the frame envelope, capability chains, and multiplayer wire parcels as standalone sketches.
| Capability | In Semantic Tricon today (Rust, live) | In the Fabric / Manual repo (TypeScript) | Verdict |
|---|---|---|---|
| P1–P5 projections & typo repair | live in Rust src/turn/stream.rs + tricons.rs — runs in 1.7 ms, emits 24-bit words, 3-clause windows, 8×8 Kronecker matrix, 75 B Merkle-XOR parcel, 8 keyboard/voice repair laws (PR #86–#88) | gk-grammar.ts (298 lines) — toy TypeScript prototype over a 196-cell array | Keep Rust. Already 100× further along. |
| Workflow routing & execution | live in Rust src/turn/fill.rs + jobs.rs — 25 real workflows (FileRead, CodeSearch, GitCommit, PullRequestLand, ScreenCapture, voice tab), 0-token turn in 12 ms | funnel-mask.html + resolution.html — static demonstrations over 179 sample rows | Keep Rust. The engine already runs real processes safely without a shell. |
| World state (what exists & who owns it) | 4-bit stub src/router/validate.rs:34 — WorldSymbol is literally a 4-bit integer mask (CurrentTab=1, WorkingDirectory=2, ActiveMachine=4, ConversationContext=8) | valuable gk-registry.ts (124 lines) — bit-indexed resource graph (gk://kind/id), owned-by cycle breaker, O(1) column AND queries | Port to Rust (Rank 1). Tricon cannot reason about real objects until its 4-bit mask becomes this Registry. |
| Frame of intent (constraints, horizon, law, proof) | missing fill.rs extracts flat slot strings (repo, pr, branch) — no actor, no authority, no expiry, no constraint or proof slot | valuable rev23/instrument/frame.ts — the 8 Frame rules: (Action, Object, Surface) + Actor + Authority + Bindings + Constraints + Timestamp/Horizon + Law + Proof | Port to Rust (Rank 2). Turns a bare CLI command offer into a governed commitment that refuses stale or unpermitted acts. |
| Permissions, delegation & spend-once approvals | binary only every changing workflow is mode: "confirm" — always stops to ask the human, with no delegation chain or spendable permit | valuable gk-capability.ts (110 lines) + §17 Proof Ledger — explicit > derived > owner-chain > no, plus SPENDABLE → SPENT single-use approval tokens | Port to Rust (Rank 3). Lets you grant an agent or teammate a bounded permit (e.g. "land PR #89 once checks pass") that spends once and cannot be replayed. |
| Multiplayer across machines & people | single-node tricon-workbench runs on m6-01 alone; M3 and M4 only view it in a browser; no signed peer log | valuable gk-wire.ts (66 lines) + gk-fountain.ts (288 lines) — Ed25519-signed Parcel of changed cells between world revisions, RaptorQ fountain spray/catch | Port to Rust (Rank 4). Connects your Mac, M4, m6-01, and a second person's world over one signed log. |
Each step adds one working capability to semantic-tricon and unlocks the step after it. No step writes prose or touches the 44-page manual.
WorldSymbol stub with the Fabric Registry in Rust
What it adds: src/router/world.rs and validate.rs gain a live in-memory + memory-mapped Registry of real resources (gk://repo/…, gk://pr/…, gk://host/…, gk://tab/…, gk://user/berkeley) and their owned-by / opened-by / grouped-in edges, ported directly from gk-registry.ts (124 lines).
Why it comes first: Right now fill.rs has to guess which repository or machine you mean from text in the current message. With the Registry, Tricon knows which repo, PR, tab, and host are open and who owns them in one bitmap AND — and Ranks 2, 3, and 4 all index into Registry row IDs.
confirm offers in the 8-Rule Frame Builder
What it adds: Every changing turn in src/turn/ produces a full Frame (the exact structure in the frame builder) instead of a bare command preview: Actor, Authority, Bindings, Constraints, Horizon (expiry timestamp), Law, and Proof. If any of the 8 rules fails (expired horizon, missing constraint, untyped binding, authority mismatch), the engine refuses before execution and names the failing rule and fix.
Why it comes second: This makes the Frame Builder you liked in Preview the actual confirmation card inside the live Semantic Tricon chat and workbench — wired to the 25 real workflows, not dummy dropdowns.
gk-capability.ts + spend-once approvals into jobs.rs
What it adds: When a Frame asks to run a changing workflow (GitPush, PullRequestLand, ServiceRestart, ScreenCapture), Tricon checks may(principal, act, resource) across the Registry's owner chain, and admits a signed approval token that flips from SPENDABLE to SPENT in the job ledger (state_dir/jobs/) the instant the command runs.
Why it comes third: Today Tricon only has two modes: read_now (no check) and confirm (always block on a human click). This step gives you safe delegation: you can approve a Frame once (or delegate write on a repo to an agent) and know the permit cannot be spent twice or used on a different resource.
m6-01 via signed Parcels
What it adds: Port gk-wire.ts + gk-fountain.ts's Parcel (schema · world · from_rev · to_rev · xor_cells · ed25519_sig) into Tricon's existing 75-byte P5 parcel path. Every world change on M3, M4, or m6-01 seals a signed parcel; peers verify the Ed25519 signature against the owner chain, advance their Registry revision, and refuse any out-of-order or unauthorized flip.
Why it comes fourth: Once Ranks 1–3 exist, multiplayer is no longer a 44-page theory — it is just shipping the 75-byte P5 parcel Tricon already computes on every turn between two live Registries.
One end-to-end scenario, zero mocked data, exercising Ranks 1 → 4 on your actual machines:
gk://user/berkeley@m3) and m6-01.gk://repo/lincicomb/semantic-tricon, gk://pr/88, and gk://host/gk-m6-01 from live world state.read_now runs PullRequestRead + CheckInspect immediately; the changing act ServiceRestart is held inside a Frame with constraint checks == green and horizon 18:00.from rev 41 → 42). You tap Approve once on the M4. That signs a single-use approval parcel back to m6-01, flips the permit SPENDABLE → SPENT, restarts the service, and updates both Macs' world revision to 43.permit out_42 is SPENT). Change the horizon to yesterday — refused by Frame Rule 6 (EXPIRED). Have an unpermitted guest principal sign the parcel — refused by gk-capability (may-not: not in owner chain).Only two of these six (The frame builder and The capability check) are direct inputs to Ranks 2 and 3 above; the other four are reference calculators for routing, resolution, and learning thresholds.
1. The frame builder — direct input to Rank 2Assemble 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.§8 Frames · 18 checks · 114 assertions · 60 KB · the image open in Preview on your M4 2. The capability check — direct input to Rank 3One 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 3. The Decision Ladder — reference for rung selectionSeven 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 · 22 assertions · 17 KB 4. The promotion bench — reference for nightly store promotionA 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 5. The funnel mask — already superseded by Rustsrc/router/search.rsA 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 capability rows · 82 KB
6. The resolution loop — already superseded by Rust src/router/rank.rsTwo observations disagree: which one wins, on which rule — or the single piece of evidence that would settle it.§100 The Resolution Loop · 5,362 assertions · 31 KB
gk-m6-01:8787, commit 2d8d6ba, 28 Oct 2026)01a17113-6dad in worktree s-20261023T131422Z (filed under thread title "GOOSE: Evaluate Goose as Tricon substrate (curated)").