Written 2026-09-24, after the seven books of Series One were drafted. It surveys the whole tree: the root documents, 40-odd sub-repositories, the review bundles, the claim ledger, the Super clone at `~/Documents/Codex/2026-09-12/le/work/super-live`, `~/Projects/Travelmonstr`, and the operator's memory notes. The notes were used only as leads.
How to read this file. Every story in §2 was checked on 2026-09-24 by opening the file it cites and finding the quoted words or numbers. Where a story could only be found in a memory note under ~/.claude/projects/-home-travis-ProjectAmp2/memory/, it is marked memory-only, unverified. Each story also carries the rung of its underlying record, using the doctrine's ladder (spec · in_tree · live_local · live_deployed · external). That rung describes the record behind the episode, not the book the episode would appear in. Nothing in this file is external.
Private-repo rule. ProjectAmp2 is private. Every path below is for the author. A published chapter names and hashes a private file; it never links it and never quotes it (ruled 2026-09-15; see books/README.md).
Depth scale: deep means a book or most of one. chapter means one chapter. paragraph means a passage inside a chapter. mention means a name or a table row. none means it appears nowhere in books/book-*/ or in the Patterns catalog (opensentience.org/_patterns/data/patterns.json, 32 patterns and 12 anti-patterns). Coverage was measured by grepping every chapter on 2026-09-24.
| Subsystem / idea | Covered by | Depth | What is missing |
|---|---|---|---|
| Interaction calculus, confluence, TRVM | Book 1 ch1–2 | deep | — |
WRL sealing, sem-…, meaning-is-a-hash | Book 1 ch3; Patterns meaning-is-a-hash, identity-is-a-seal | deep | The legacy_edge join trap: a zip by index produces a complete, plausible and wrong mapping (WRL/test/conformance.mjs:4725) |
| Films, exact replay | Book 1 ch5; Patterns exact-replay | deep | — |
| TRAAVIIS / trvs, three judges | Book 1 ch6 | chapter | Why the chain stops by name at bundle- and not episode- (TRAAVIIS/examples/end-to-end/README.md:68-76) |
Compiled backend (CompilePlanV1, laws_gate.py) | Book 1 ch7 | chapter | The flags policy (-march=native is znver5); the Python printer as the real bottleneck (4,424× for the step alone, 297× with render); refuse_narrowing (BENCHMARK_LANE.md §1, §4; TRVM/compiled/FLAGS_POLICY.md) |
| WRLM proposer | Book 1 (4 mentions) | mention | The 448-of-768 "cells that could never be inhabited" retraction (TRVM/WRLM_RESEARCH_BRIEF.md:426-442) |
| TRVM governance rounds (E-42/E-43, the grid) | — | none | "Six consecutive rounds found the defect in the instrument rather than the engine" (TRVM/DERIVE_R16_18_REVIEW.md:3405) |
| TRVM vs HVM4 / Bend bench | Book 1 ch8, Book 6 ch6 | paragraph | Bend 2.0.4 as a row; laws 5 and 23 in Bend (FOUNDATION_LANE.md §3ao–§3aq) |
ComputeDriven floor (cd-kernellet, census, 8 KB gate) | Atlas row only | none | The whole arc (§3, book S2-2) |
| FCB, the floor competitive benchmark | — | none | Everything |
| Floor attribution (T4b) and its outside review | — | none | Five withdrawn claims |
| sflr / locus multiplexing (R7.2.1x) | Patterns carrier-multiplexing, progress-aware-placement | pattern only | The 2C saga; F163 hidden by the rendezvous |
| FPLA (Floor Programmable Locus Array) | — | none | A design proposal: "no implementation, no measurement" |
| Settlement boundary (R89–R91) | Book 5 ch7 | chapter | — |
| Subsystem / idea | Covered by | Depth | What is missing |
|---|---|---|---|
| box-and-box kernel, eight rungs, bridge | Book 3 | deep | — |
| Compose / CC2 admission rounds (15 outside reviews; attestation; WIRE) | Book 3 ch8, Book 4 ch2 | paragraph | The review saga itself: COMPOSE_ADMISSION_ROUND.md, WIRE_ROUND.md, WIRE_CLOSURE.md, register R25–R67 |
| [&] Protocol, PULSE, PRISM, κ | Book 4 | deep | — |
| Graphonomous v0.4 | Book 4 ch5 | chapter | — |
| Graphonomous v2 (G0/G1, D-001…D-077) | Book 4 ch5 | paragraph | The decision log; the tri projection; the 6.4 GB assertion |
| Invariant frontier / Periodic Table / Mosaic / claim ledger / invariant-r10 | — | none | Everything (book S2-5) |
| WEK R0–R3 (model, laws, mutants, verifier) | Book 2 ch6 | chapter | The four falsified premise families; the checker that kept its GREEN receipt; the eight-document R1 round |
| AGENCY.md: agent as an active locus | Patterns active-locus and 4 more; Book 2 ch1 | pattern plus mention | The argument at book length (book S2-6) |
| DOCTRINE, the evidence ladder, the register | Book 6 | deep | — |
| the-residency (multi-agent board) | Book 3 ch9 (the algebra finding only) | mention | Gate 0 closed negative (the-residency/DIRECTION.md:53-72) |
| weave (static cost certificate) | — | none | A research probe of rung 8 |
| THREAT_MODEL / CONFIDENTIALITY (studbook §10.2) | Book 5 ch6; Patterns | chapter | — |
| Subsystem / idea | Covered by | Depth | What is missing |
|---|---|---|---|
| Super, ampd, one door, bots, lanes | Book 2 | deep | — |
| Super-builds-Super gate | Book 2 ch7, Book 6 ch9 | chapter | — |
| Carrier confinement (seccomp, Landlock, descriptors, PTY possession) | Patterns carrier-confinement; Book 2 ch5 | paragraph | The x32 hole, the decoy terminal, CLOSE_RANGE_CLOEXEC, Yama (§2 story 1) |
| Ordered participant / INDETERMINATE | Patterns three-valued-outcome | pattern only | "A timeout ABANDONS the request, and the mutation can still land" |
| Cockpit transport (Tauri/wry) | — | none | Channel::send returning Ok when no JS ran |
| Super storage budgets and retention | — | none | 141,381 bytes against a 131,072 budget; the 32-review ceiling |
| Super Mobile / phone clients | — | none | Two phone clients; the Expo repo that drifted (memory-only lead) |
| Dogfood cycle | — | none | ?? repos[0] and a lane that cannot be undone |
| Subsystem / idea | Covered by | Depth | What is missing |
|---|---|---|---|
| T&R distro | Book 5 ch3 | chapter | The build traps: fontless icewm, nuageinit firstboot, 4K and the 0×0 greeter |
| RRABBIT road | Book 5 ch4 | chapter | The GL faults, the idle brake, the compositor-proxy port to FreeBSD |
| RAVIO | Book 5 ch4 | mention | Altitude as evidence rung |
| PARKVPS | Book 5 | mention | Console hazards: an escape sequence that executed itself |
| The lab (RB5009, cd-floor-01, locuchest) | Book 6 ch7 (bad RAM), Book 1 ch8 | chapter | The network as a story; failover drill; host profiles |
| body-os / body-browser (GRASP) | Book 4 ch8 | mention | — |
| Subsystem / idea | Covered by | Depth | What is missing |
|---|---|---|---|
| World Cloud, restore-as-grant | Book 5 ch5 | chapter | — |
| Cloud v1 control-plane review rounds (findings 16–123, R14–R88) | Book 5 ch5 (Hyperdrive in one paragraph) | paragraph | Some 100 measured authority defects and ten sentence-sized laws (book S2-3) |
| WORLD_SIZE | Book 5 ch2 | chapter | — |
| locu.systems | Book 5 ch7 | chapter | — |
| E analytics plane | Atlas / appendix row | none | 15 rounds, 500 falsifiers, N87/N90 |
| Workbench | Atlas row | mention | The three ingest defects, one of them a live secret leak |
amp harness (code/, 20,042-line amp.py) | Atlas "around everything" | none | Everything (book S2-1) |
| Review-bundle culture (30 root bundles, adjudications) | Book 6 ch8 (rulings) | paragraph | How GPT, Codex, Fable and Claude divide the work |
| Parallel sessions in one tree | Book 6 ch8 (numbering) | paragraph | The AGENTS.md that disabled the memory loop |
| Academy | Book 0, Book 6 ch2 | mention | The corpus that was invisible |
| alkeyword | — | none | "nobody had ever run the whole pipeline in one pass" |
| Portfolio sites, amp-nav, Cloudflare deploy | Series-plan appendix | mention | The edge that rewrote a claim; the nav with four causes of staleness |
| ToolboxHVAC | Appendix row | none | The same discipline in a product for HVAC technicians |
Cyberutensil / alphapiler (~/Projects/Travelmonstr) | — | none | Outside the tree and uncommitted. A four-arm experiment awaiting a held-out set |
The shape of the gap. Series One explains the architecture and the method. What it leaves out falls into four bodies of material, each with a large, dated, reproducible record behind it:
The counting layer: floor, FCB, attribution, E.
The adversarial review record: Cloud v1, compose admission, WEK, invariants.
How the stack is built by agents: amp, bundles, parallel sessions, the Super lane.
The machine layer's engineering: lab, distro, compositor, confinement.
Series One mentions each of these and narrates none of them.
Twenty stories, in no ranked order. None of them appears in any Series One chapter. For each one I checked that the cited words or numbers are present in the file named (2026-09-24).
Super confines each Carrier with a seccomp filter that denies 28 syscalls by number. On x86-64 a process can issue the same numbers in a second numbering space: x32 sets bit 30, and the kernel reports those calls as x86-64. The filter checked the architecture and let them through, so 23 of the 28 denied calls reached their handlers, fork and clone among them. The fix refuses every number with the x32 bit set. It closes the whole space, where extending the list would have closed one number at a time. Rung: live_local (measured against the frozen filter on this kernel). Verified in tree: super/host/src/verify.rs:3296-3311 ("Measured against the frozen filter: 23 of the 28 numbers"), super/host/src/confine.rs:702-815.
Super's host sabotage battery edits real source files to prove the checks catch the damage. It was meant to restore them on exit. It carried a comment referring to "the INT trap this file already carries", and no trap at all. An interrupted run left native_fd.ex, the module that closes every received descriptor, sabotaged in the working tree, "one git add -A away from being committed". The follow-up check kills a battery with SIGKILL, which no trap can survive, and verifies the tree anyway. The file's own words: "The comment is the worse half: it tells a reader the protection exists." Rung: in_tree. Verified in tree: super/tools/sabotage-host.sh:78-88, super/tools/check-sabotage-isolation.sh:1-20.
In the World Cloud control plane, migration 0011's comment said the API "may read membership and record last-seen". The migration granted table-level UPDATE, which reaches every column. The measured result: SET app_role='owner' succeeded, and a revoked membership un-revoked itself. It was the third round in a row in which a comment described a restriction the grant beside it did not implement ("Immutable by convention" on a table later granted UPDATE and DELETE; "Never matched against" on a column the Worker matched on). The round's law: "A comment is not an enforcement mechanism." This story and story 2 are the same failure in two unrelated codebases a month apart, which makes them a natural cross-book thread. Rung: in_tree (reproduced in a local PostgreSQL battery; fixed in 0012_column_authority.sql). Verified in tree: REVISION_REGISTER.md:1096 (finding 37), :1105-1122 ("THE LAW THIS ROUND EARNED").
Migration 0009 used ALTER DEFAULT PRIVILEGES … REVOKE EXECUTE ON FUNCTIONS FROM PUBLIC as a safety net. Measured on PostgreSQL 18.4, the IN SCHEMA form stores zero rows, so every function created afterwards was PUBLIC-executable. The sentence read as a guarantee for two rounds, and the guarantee was really carried by later migrations remembering to REVOKE by hand: four did, one did not. Earlier, a read-only role had called resolve_principal and created a principal: a read-only role performing an INSERT. When the fixes are removed and the battery is re-run, 9 of its 55 checks fail. The record's lesson: "a battery inherits the blind spots of whoever wrote it." Rung: in_tree. Verified in tree: REVISION_REGISTER.md findings 22 and 51 (around lines 356 and 388), finding 57 (R42), and l.371 ("9 of the 55 fail").
Abandoning an upload released its bytes. But abandon is a row in our database, and the presigned URL already handed out still worked. So reserve 100 → offer A → get URL → abandon A → offer B → use A's still-live URL stored 200 bytes against a 100-byte reservation. The next round found the same law broken in the byte ledger: "The provider said the bytes landed, and expiry made them free", and afterwards no path was left to charge them. Ruled as R55/R69: our clock ends our willingness to mint authority; it does not end the authority. Rung: in_tree (measured in the local battery). Verified in tree: REVISION_REGISTER.md finding 61 (≈l.427), finding 65 (l.442), R55 (l.72), R69 (l.86); worker/migrations/0022_intent_state_machine.sql:62 ("THE LOCK IS NOT THE POINT; THE RE-READ IS").
The compose layer transferred ownership of bytes with ArrayBuffer.prototype.slice, which it had captured up front so that a caller couldn't swap it. The spec, however, has slice run SpeciesConstructor(O, %ArrayBuffer%), which reads the caller's buffer's constructor. Caller code ran mid-copy, and the boundary authenticated bytes B from a frame that held A when the call began. The law meant to catch this, WIRE2, was green the whole time: it counted calls to the getters it had replaced and never asked what the copy did. "The P0 was real and my law was false." Rung: in_tree. Verified in tree: WIRE_CLOSURE.md:1-40 (quote at l.37).
In ComputeDriven's durable-authority round R0.8, the canonical encoding was not injective. A forged one-record history reproduced an honest two-record history's SHA-256 exactly and opened clean. The same record catches a hand-typed "thirteen" that was really 21. It is the cleanest illustration in the tree of why meaning-is-a-hash depends on the canon, not the hash. Rung: in_tree (FROZEN at e42cc88). Verified in tree: computedriven/receipts/STATUS.md:17-26.
A planning conversation proposed 8/16/32 KiB tiers for the ComputeDriven floor, on the understanding that it "was already known to sit somewhere in the 32–64 KiB band". It was not. The tree held a 64 KiB RAM budget in a linker script and a codec probe under 2 KB. When the census measured the floor, the production ring image came to 5,264 bytes against an 8,192 gate, at about 312 bytes per context (309 after R4-P). Rung: in_tree (size gate PASS). Verified in tree: computedriven/docs/floor-census.md:15-18, computedriven/receipts/KERNELLET-R1.md:22,33.
Experiment 2C of R7.2.1D was refused by its own guards three times. On the first attempt the sentinel read 34 % below calibration while load and idle looked normal. The second, the one authorised run, failed both guards. Travis then rebooted the host with a single terminal open, and attempt 3 ran 72 of 72 and found P4b EQUIVALENT under both policies. The episode became standing doctrine: a reboot is "a precondition and a preparation, not a gate and not evidence." Rung: live_local. Verified in tree: computedriven/receipts/R7.2.1D-2C-EXECUTED.md:48,101, computedriven/docs/battery-host-profile.md:4, FOUNDATION_LANE.md §3n–§3q.
The first official run of the floor competitive benchmark (cd-kernellet's policy against the Linux, Go, BEAM and Seastar schedulers) admitted 1 of 20 blocks and quoted no result. The median lateness was 2 µs. But a 500 µs worst-case gate applied across twelve runs per block passes about 0.76¹² ≈ 4 % of the time. Its author proposed fixes and labelled them "a second-author recommendation … (NOT a ruling)", because choosing gates after seeing the data is exactly what the design forbids. Codex's review said the evidence supports none of "faster than C", "better than Tokio" or "better actor runtime than BEAM". Rung: live_local, EXECUTED / NOT ESTABLISHED. Verified in tree: computedriven/receipts/FCB-BATTERY-R1.md:34, computedriven/receipts/FCB-BATTERY-R1-RECOMMENDATION.md:1, computedriven/docs/fcb-benchmark-review-2026-09-17.md:21-27.
An outside review of the floor-attribution run (T4b) accepted the official run and withdrew five claims written about it. The "69.3 % semantic" column turned out to be a subtraction against a constant-operation control. No measurement family had touched the production image. And "≈13.6 ns" had been computed by dividing cycles by the maximum clock, when the receipt already carried a measured ns_per_op = 14.17. The benchmarking lane's own log adds: "An outside review retracted five of T4b's claims, and I had repeated four of them." Rung: live_local (the run); the claims were withdrawn. Verified in tree: BENCHMARK_LANE.md:1202-1215 (§7.13), T4B_GATE3_REVIEW_BUNDLE.md:60-70.
The-residency was designed around one premise: a multi-agent deliberation board beats a single strong critic at verified findings. Five hand-adjudicated runs measured board minus single-critic precision at −0.33, −0.25, +0.125, 0.0, 0.0. That is never a reliable win, even on threads built to favour the board. The one thing that helped every time was a stance-free entailment audit: "does this source line entail this claim?". That check works on any producer, so it shipped as a verification gate and the board went behind a research flag. The record states its own caveat: every agent was one model debating itself. Rung: in_tree (n = 4–5, unblinded same-model judge). Verified in tree: the-residency/DIRECTION.md:53-72, the-residency/EVIDENCE/gate0*.md.
amp scores each lane by the highest evidence rung it has recorded. lane_rungs() took the maximum rung ever claimed. A lane that overstated itself once kept being scored as if it hadn't, even after a later review judged the claim unearned, and every direction proposal was fed that score. The fix added retract_rung. The docstring states it plainly: "the highest rung ever claimed can never fall." Rung: in_tree. Verified in tree: code/amp.py:11589-11593. That its first live run settled 15 of 18 findings is memory-only, unverified (project_amp_contradictions_gate.md).
AGENTS.md, the instructions file every Codex session reads, listed 12 required context files, and 11 of them did not exist. The skills had been renamed. The register's audit explains why this was worse than a typo: "an agent that cannot find its required reading concludes the memory loop is unavailable, and the policy silently stops running." The same file named the wrong database vendor, stated the EXLA setting backwards, and had a repository layout four months stale. Rung: in_tree. Verified in tree: REVISION_REGISTER.md:168-186 (§9).
"Graphonomous v2 crashes Claude Code" turned out to be the kernel OOM-killing the editor. dmesg named the editor's own MainThread at about 23.6 GB. The test suite's entire excess came from one `assert.deepEqual` on two 15 MB Buffers, which on a mismatch renders a diff of every byte. Replaced by a hash comparison, the same test peaks at 185 MB. The fix commit is titled "One assertion cost 6.4 GB, and the machine paid it with the editor process." Rung: live_local. Verified in tree: graphonomous/v2/tools/mem_probe.mjs:20-22, graphonomous/v2/tools/test_repro.mjs:32-33, commit 824b9688 (graphonomous repo).
The Periodic Table of Agent Invariants (opensentience.org/invariants.html, 46 cells) was grouped by which spec folder owned each cell, which its own data file calls "an org chart, not a structural axis". That grouping is why "the table closed ZERO cells across nine rounds while the ledger grew". It also had no colour for REFUTED, so "nine rounds of killing propositions rendered as an unbroken accumulation of achievement." The page rendered 52 cells on a table of 46. Three gates used to cut a JavaScript literal out of the HTML page and eval() it. The ledger now has 190 claims, 19 of them REFUTED, including "κ as defined in cell 01 carries the federation obstruction of cell 36", and one claim that "shipped PROVED on two thousand trials" (LED-C9) and was false. Rung: in_tree. Verified in tree: opensentience.org/_invariants/data/axes.json:2,18, opensentience.org/_invariants/build/build.mjs:11-12,668,723-725, scripts/check-claim-ledger.mjs:67-68, CLAIM_LEDGER.json (status counts computed 2026-09-24), INVARIANT_FRONTIER.md:44-48,638.
Three headless tabs of RRABBIT, left behind by earlier browser-automation runs, held eleven of twenty-four cores for hours. They were compositing at 15–30 fps for a server that had died long before. One had rendered 369,325 frames and logged 358,290 GL errors that nobody would ever read. document.hidden, "no windows" and "no new content" all failed as idle signals, so the frame loop's brake keys on input alone. Its sequel: two sessions in a row set out to fix a "frozen window" bug in Surface.js, and the freeze turned out to be the brake doing its job. Rung: live_local. Verified in tree: RRABBIT/m2/shell.js:1810-1817,1918-1933, RRABBIT/docs/RUNBOOK.md §7 (≈l.497), RRABBIT/docs/spec/README.md §30 (≈l.2412).
"Black edges on resize" and "the window is black" were filed as two faults. The runbook's verdict: "They are one number." The texture mapping assumed the picture sat at the far corner, so padding of 14 produced exactly 14 black columns. A related fault, a diagonal half-window, came from three.js leaving CULL_FACE enabled for Greenfield's six-vertex triangle strip, which alternates winding, so exactly one triangle was culled. "Every report of this fault named native applications, for four weeks", and that pattern turned out to be the diagnosis. Rung: live_local. Verified in tree: RRABBIT/docs/RUNBOOK.md:571-595, RRABBIT/docs/spec/README.md:1402-1440.
Two stories from porting the machines:
FreeBSD with no GPU. Running compositor-proxy inside FreeBSD with no /dev/dri hit six faults,
every one FreeBSD-specific and none needing a GPU to fix. libepoll-shim's "epoll" is really a kqueue, which rejects uv_poll_init with ENOTTY. Nothing checked the return value, so the process segfaulted. The end state was a live foot terminal on software rendering.
PARKVPS consoles. Replaying a recorded boot into a live xterm.js answered a ESC[6n query from
minutes earlier. The answer arrived at a shell prompt as input and was executed (-sh: 80R: not found). The same fleet has a "reseed and reboot" command that is silently a no-op, because nuageinit runs on first boot only and appends to its log, so you end up diagnosing a crash that is no longer happening.
Rung: live_local. Verified in tree: RRABBIT/docs/spec/README.md:2196-2255,2309-2360, PARKVPS/docs/spec/README.md:103-109,462-467.
Cloudflare's Email Obfuscation took [email protected], an npm install name, for an email address and served it as [email protected]. It also ate retraction quotes. The deploy was correct, the repository was correct, and the published claim was wrong. A second edge story: the docs atlas was hash-routed, so 373 documents shared one indexable URL that served 6 extractable words, about 458k words invisible to every search engine and to the stack's own crawler. Rung: live_deployed (both were observed on live surfaces). Verified in tree: CLOSURE_BRIEF_R18_REVIEW.md:436 (§2.13), scripts/cf-email-obfuscation.sh, ACADEMY.md:26-34.
| Story | One line | Rung | Source |
|---|---|---|---|
| A kill test struck out its author | SIGKILL during a DETS save left the table dirty in 15/36 kills on ext4 and 29/36 on tmpfs, starting below 100 effects. The "small near an empty journal" inference is kept, marked superseded | live_local | wek/b2/trvm/RESIDENT_TIMING.md:335-354 |
?? repos[0] | The first dogfood run opened its lane on the wrong repository. Lanes cannot be closed by design, so recovery meant re-planning on a new lane | live_local | super/tools/dogfood-setup.mjs:136, super/tools/dogfood-recover-lane.mjs:2-8 |
Ok means nothing ran | wry drops the async result, so Channel::send returned Ok when no JavaScript ran, and flow control stayed shut (wry#1644) | in_tree | super/cockpit/src/worker.rs:134-140 |
| A badge that cost a second | Asking a terminal whether it was active waited 1 s per Worker: 16 terminals took 16,015 ms against a 15 s budget | live_local | SUPER_D13C2C1B_REVIEW_BUNDLE.md:141,245 |
| A timeout abandons the request, not the work | "may be about to" is part of INDETERMINATE, and :bind_basis, the one op that mints source authority, was classed a read, with its gate red and unrun for weeks | in_tree | super/ampd/probes/ordered_participant.exs:171, super/docs/reviews/R0B_R_R7_VALIDATION_VOCABULARY.md:522-538 |
| Super filled its own store | development_attempts reached 141,381 bytes against a 131,072 budget; a world that has taken 32 reviews can never take another | live_local | SUPER_BUILDS_LANE.md:80-83,501 |
-march=native is znver5 | The build id hashed the flag text, so two machines could share an id over different machine code | in_tree | TRVM/compiled/FLAGS_POLICY.md:41-47 |
| A checker that stayed GREEN | WEK R3's checker exited 1 and kept its previous GREEN receipt. A second reviewer then built 11 falsifying traces across four premise families | in_tree | WEK_R3_REVIEW_4.md:21, wek/r3/HANDOFF.md:58-86 |
| A verifier with an invented setting | E round 15: a hand-copied query in a verifier named a ClickHouse setting that exists "in no family at all". Also: a ROLE is not a USER | in_tree | e.computedriven.com/FINDINGS.md:2404-2428,2619-2627 |
| A live secret leak in the recorder | Workbench's PII profile never redacted tool_call.arguments; one segment carried 86 key-shaped strings | in_tree | workbench/docs/INGEST_SPIKE.md:131-164 |
| Fabricated modesty | A handoff called two live products "not built" (71 and 157 tests) — "still false" | in_tree | HANDOFF-2026-08-10.md:30 |
| Cut off mid-rename | A worker killed by a limit committed a half-renamed ic32.c on an early-commit rule, and it survived two more kills | in_tree | DOCTRINE.md:145-152 |
Memory-only, unverified (do not use without finding the record):
A git plumbing step returned the empty tree and a push wiped main to zero files; recovered by
force-push (project_nav_stale_three_causes.md).
The wrangler token has no DNS scope, so API-attached domains stay "pending" forever
(project_cloudflare_pages_new_site.md).
"12 of 14 sharpening rounds moved odds DOWN". The note's own body says 14 rounds, 2 gained, which is
weaker. The MEMORY.md index overstates its note.
Eight books. They share Series One's shape: 6–9 chapters of 800–1,500 words, a hook question per chapter, the rung stated, what this does not establish, a slide block, and a video script. Each cites Series One and the Patterns catalog instead of re-teaching them.
How the stack is built by agents, and what keeps them honest.
Thesis: Agents write most of the code. The honesty comes from a record that cannot be walked
back: refusals, outside review, and a single evidence ladder.
Audience: people who orchestrate coding agents (Claude Code, Codex, Cursor) across real repos.
This is the largest audience the Library has.
Serves: computedriven.com door two (Super); the YouTube series "How this was built"; Academy's
Prove layer.
Chapters:
One tree, many sessions. No worktrees, many concurrent sessions, ruling numbers that
collided three times; read the high-water mark from the table, not your own file (REVISION_REGISTER.md:129-132).
The instructions are code. DOCTRINE's core is injected into every prompt and CLAUDE.md is
prepended to every session. So the AGENTS.md that listed 11 missing files switched the memory loop off (story 14).
A lead-manager harness. amp: one chat window, lanes, a board; apply refuses to merge
because other sessions are editing the checkout (code/amp.py:5768-5769); the PR gate reads GitHub's check rollup, never the harness agreeing with itself.
A rung that could only go up. lane_rungs and retract_rung; sharpening that stops on
measured gain rather than a count (amp.py:11202-11207); "drifted into sameness" was the wrong premise (amp.py:2636-2640).
Write it to be attacked. Review bundles are generated, never hand-written; the adjudication
governs; self-graded GO comes back REVISE (the WEK R0→R3 chain, about ten rounds).
When the reviewer is also the author. The WEK R3 checker kept its GREEN receipt, and Review 5
was needed because Codex wrote both the patch and the controls. T4b's five retractions, four of them repeated by the next author.
A board that never beat one critic. The-residency's Gate 0 and the entailment audit that
shipped instead (story 12).
Recording the work. Workbench turns transcripts into bundles; three redaction defects, one a
live secret leak; and "verify deploys by the route table, never by grepping HTML".
The workshop takes the pen. The Super lane: tasks become development_tasks in the live
world; gate 3; the store that filled itself. This book hands the reader to Book 2 ch7.
Key sources: code/docs/spec/README.md, code/docs/HARDENING.md, code/amp.py,
WORKSPACE_DIRECTION.md §0, agents/README.md, REVISION_REGISTER.md §9 and the numbering notes, WEK_R*_REVIEW_BUNDLE.md / _ADJUDICATED.md, WEK_R3_REVIEW_4.md, WEK_R3_REVIEW_5.md, BENCHMARK_LANE.md §7.13–7.14, the-residency/DIRECTION.md, workbench/docs/ROADMAP.md, workbench/docs/INGEST_SPIKE.md, SUPER_BUILDS_LANE.md, superlane/gate3/.
Must not claim: that the dark factory closes; that Super builds Super (still spec, R98); that
review by another model is external (it is outside the author, not outside the stack); that the residency result covers heterogeneous multi-model boards (it was one model debating itself); any productivity multiple.
ComputeDriven Edge: counting admitted steps, and a benchmark that refused to crown itself.
Thesis: A scheduler that bills for progress has to count honestly. Here is how to size it,
attack it and benchmark it, and why "not established" is a result.
Audience: systems programmers, scheduler and runtime people, benchmark skeptics.
Serves: computedriven.com (the floor under World Cloud and locu.systems); a "benchmarks we
refused to publish" video.
Chapters:
What a floor is. Locus, carrier, admitted transition. The floor evaluates no WRL law; every
check is locus, mark, generation, carrier, fence or evidence (BENCHMARK_LANE.md §4).
Nobody had sized it. From a 32–64 KiB assumption to 5.3 KB and 312 B per context (story 8).
A history that lied with the right hash. Canonical encodings must be injective (story 7).
The host is part of the experiment. Three refusals, a reboot, and "preparation, not
evidence" (story 9).
The rendezvous was hiding a bug. F163: a mutant never exercised on synchronous handoff,
caught on every asynchronous one (computedriven/receipts/R7.2.1B-OPEN.md §11.3).
Where one step spends its cycles. Floor attribution; ptrace and perf disagreeing by
construction on rep (computedriven/docs/floor-attribution.md:118-129); story 11.
One in twenty. FCB: five adapters, six controls, and the arithmetic of a worst-case gate
(story 10).
Five benchmarks, two of which are benchmarks. B1–B5; B2 is an admission oracle; the printer
that was the real bottleneck (4,424× for the step alone, 297× with render) (BENCHMARK_LANE.md §1, B2·perf).
The money never flows back. A short close on the count record and the settlement boundary
(R89–R91), handing off to Book 5 ch7.
Key sources: computedriven/docs/{floor-census,fcb-model,floor-attribution,battery-host-profile,settlement-boundary,sflr-r2}.md,
computedriven/receipts/{STATUS,KERNELLET-R1,FCB-BATTERY-R1,FCB-BATTERY-R1-RECOMMENDATION,R7.2.1D-2C-EXECUTED,R7.2.1B-OPEN}.md, BENCHMARK_LANE.md, T4B_GATE3_REVIEW_BUNDLE.md, FOUNDATION_LANE.md §3b–§3r, §3af–§3an, lab/BENCH.md.
Must not claim: that ComputeDriven is faster than Linux, Go, BEAM or Seastar; that the FCB
adapter is the production kernellet (it is computedriven-policy); "69.3 % semantic"; that the production image has been attributed; any FCB result at all (B1 is EXECUTED, NOT ESTABLISHED); that the §4 fixes are ruled.
Authority bugs found by adversarial review, and the laws they left behind.
Thesis: Authority defects live in the gap between what the prose says and what the grant
does. Every one of them here was reproduced before it was fixed.
Audience: backend and security engineers; anyone writing PostgreSQL RLS, presigned URLs,
queues or capability code.
Serves: cloud.computedriven.com; a "security laws" short-video segment; the stack's
credibility with technical buyers.
Chapters:
Read the grant, not the sentence. Viewer→owner and the three comments (story 3). Paired
with story 2 from Super.
The default you didn't revoke. SECURITY DEFINER, PUBLIC EXECUTE, and the safety net that
stored zero rows (story 4). A check means nothing until it has been shown to fail before the fix.
The database is not the wire. Hyperdrive refuses advisory locks, so a green battery can run
on a mechanism production refuses (CLOUD_V1.md §6.4). Book 5 gives this one paragraph.
A key is a promise, not a name. Idempotency that bound nothing; replay before authority
(findings 48–50, 56; R43).
Authority already issued. 200 bytes on a 100-byte hold; expiry that made bytes free
(story 5; R55, R69).
A grant without a policy returns nothing. The whole observer path was unreachable and looked
like absence (finding 59).
Two true facts and no join. Event identity ≠ delivery identity ≠ object state; a projection
must not depend on arrival order (R57, R71, R73, R75).
Evidence for absence. A completed sweep that could be forged by a literal; protocol
provenance versus provider provenance; set difference needs both operands complete (R81–R87, findings 113, 122, 123).
Capturing a function guarantees nothing. SpeciesConstructor and the law that was green while
false (story 6), plus the admission round's WeakSet brands and module-instance authority (R27, R48).
Key sources: REVISION_REGISTER.md (the Cloud v1 and M2.0 sections, l.337–1330), CLOUD_V1.md
§6, worker/migrations/, worker/CLOUD_V1_REVIEW_BUNDLE.md, COMPOSE_ADMISSION_ROUND.md, WIRE_ROUND.md, WIRE_CLOSURE.md.
Must not claim: that anything is deployed (M2 is local); that quota is enforced at the
provider (R31 is OPEN, and CLOUD_V1.md forbids saying it); an external security audit; the reviewer's identity beyond what the register records; that the list of defects is complete.
The lab, the distro, the road and the fleet: what it took to make a machine tell the truth.
Thesis: Every layer above assumes a machine. This is the machine: a router with no radio, an
OS assembled from parts, a compositor with no GPU, and a RAM stick that flipped one bit.
Audience: homelab and systems hobbyists, FreeBSD and Wayland people, the T&R download audience.
It is the most visual book in the Library.
Serves: tr.computedriven.com, and a YouTube "lab log" series with screen-recorded B-roll that
already exists.
Chapters:
A router with no radio. RB5009, the DD-WRT client bridge ("the best radio in the house by
2–2.5×"), the failover drill, the routing loop that looked like NAT for twenty minutes (lab/NETWORK.md, lab/BENCH.md:150).
Assembled, not installed: the traps. A fontless icewm that stops moving windows; nuageinit
first-boot only; /firstboot, schg, pkg -r (travel-and-rrabbit/README.md:60-89, PARKVPS/docs/spec §3).
A 4K screen and a greeter that sees 0×0. efi_max_resolution is a ceiling; xrdb -symbols;
Firefox with no window manager (travel-and-rrabbit/build-image.sh:176-185, overlay-desktop/…/Xsetup:17-38). The "three causes" framing is memory-only; re-derive it.
A fleet with no root. PARKVPS: serial over SSE, readiness probes that flood the consoles they
measure, and the escape sequence that executed itself (story 19).
A compositor on an OS without a GPU. Six FreeBSD faults (story 19).
One number, two faults. The padding band, the culled triangle, the dead texture handle
(story 18).
The brake. Eleven of twenty-four cores, and the "frozen window" that was the brake (story 17).
The machine that flipped a bit, from the operator's chair. Hashing every image against a
clean pull; the backup and restore drill (lab/backup.sh). Cross-links Book 6 ch7 rather than retelling it.
Measuring on a quiet host. cd-floor-01 against the laptop; host profiles; "a 'does not
scale' result must be reproduced outside the harness" (lab/BENCH.md §2).
Key sources: lab/NETWORK.md, lab/BENCH.md, lab/backup.sh, PARKVPS/README.md,
PARKVPS/docs/spec/README.md, travel-and-rrabbit/README.md, travel-and-rrabbit/build-image.sh, travel-and-rrabbit/overlay-desktop/, RRABBIT/docs/RUNBOOK.md, RRABBIT/docs/spec/README.md §18–§30, RRABBIT/m2/shell.js, RAVIO/README.md.
Must not claim: that T&R installs on bare metal (USB and internal disk are untested on the
download page); GPU acceleration in the guest; that the road-shell product is implemented (Tier 1 is spec, with its proofs in_tree); that benchmarks taken on locuchest before the RAM fault stand; that tr.computedriven.com resolves (as of 2026-08-25 it had no DNS record).
A research loop that audits itself: invariants, the claim ledger, and memory as an evidence graph.
Thesis: A research program that grades itself will improve its bookkeeping first. Here is how
this one noticed, and what nineteen refuted claims are worth.
Audience: researchers in AI evaluation and formal methods; people building "AI scientist"
loops.
Serves: the opensentience.org invariants and patterns pages; Academy; a "what we got wrong"
video series.
Chapters:
A periodic table for agents. 46 cells, what an invariant is, and why table was the right
metaphor and the wrong axis.
An org chart wearing a periodic table's clothes. Zero cells in nine rounds; regrouping by
evidence register (story 16).
Nineteen refuted. The ledger's statuses (190 claims: 54 PROVED, 40 OPEN, 19 REFUTED…); LED-C9
"shipped PROVED on two thousand trials"; property test versus proof (rule R22).
287,496 federations under the wrong hypothesis. An exhaustive search can be exhaustively
wrong (INVARIANT_FRONTIER.md:44-48).
Three gates that eval'd a web page. 52 cells on a table of 46; *"refusal that cannot fire is
decoration"* (build.mjs:435).
One unchecked line, seven copies. A bhyve row in every package for a machine running QEMU on
Arch (PRIOR_ART_0.md:8, invariant-r10 commit 0983119).
Did it say so, did it happen. F38: sixteen "did it say so" gates failing while every "did it
happen" gate passed (invariant-r10/CHANGELOG.md:9).
Memory that relates facts without owning them. Graphonomous v2: a 3,928-object projection;
one snapshot reading one file at two blobs; a spike that came out byte-identical to the real kernel (D-048); the 6.4 GB assertion (story 15).
Discipline, or bookkeeping? The program asks this of itself (INVARIANT_REVIEW_BUNDLE.md:184)
and the book should leave it open.
Key sources: CLAIM_LEDGER.json, INVARIANT_FRONTIER.md, INVARIANT_CURRENT_BRIEF.md,
ROUND4_RESEARCH_BRIEF.md, ROUND5*–ROUND7_8_MOSAIC_BRIEF.md (15 briefs), mosaic/, invariant-r10/, opensentience.org/_invariants/, CURRENT_FRONTIER.json, graphonomous/v2/handoff/{STATUS,DECISION_LOG}.md, graphonomous/v2/tools/mem_probe.mjs.
Must not claim: any cell "closed" that the ledger doesn't mark; that a property test is a
proof; that the three-model loop (Fable researches, Opus implements, GPT audits) is outside review; that Graphonomous v2 has a write surface or that the MCP works (Travis reports it broken; memory-only); anything from ~/.invariant-factory that isn't mirrored in the tree.
Not a model plus a harness: an active locus in a world.
Thesis: "Agent = model + harness" can't tell apart two identical systems when only one holds
authority, and it can't ask whether an agent survived a model swap. An active locus can do both.
Audience: general readers, AI policy people, philosophers of mind and action, product people.
This is the most "idea" book in the Library.
Serves: the opensentience.org front door; the channel's "big idea" episode; the Patterns
locus and agency families.
Chapters:
The wrong equation. Why model + harness fails (AGENCY §2).
An active locus in a world. The definition and its five properties (§1, §5).
The chain. World → Observation → Established state → (with policy and grant) admissible
actions. The model appears once, at "proposes" (§3).
Establish, never know. The vocabulary: position deprecated; agent ≠ agent record (§4).
State does not grant authority. The ambient-authority trap (§4), tied to Book 2 ch4.
Continuity through replacement. Carriers swap, loci persist; and the Codex-swap claim that
was walked back (AGENCY.md:532-535).
Models become machinery. §5b, the factory, and the two corrections the tree has already made.
Established or a reading? §6: what is witnessed (46/46 authority fields) and what is not
(heredity: "the engine runs and has nothing to run on").
Prior art, honestly. Hewitt's actors, Anthropic's PSM "locus of agency": not a coinage
(patterns.json novelty_not_claimed).
Key sources: AGENCY.md, DOCTRINE.md, opensentience.org/_patterns/data/patterns.json
(locus and agency families), UNBOXED_PATTERNS.md, scripts/check-messaging-language.mjs.
Must not claim: that the definition is ratified (AGENCY.md is proposed, not entrenched; only
Travis ratifies); novelty of the word or the concept; that heredity is witnessed; any claim about consciousness or moral status.
Twenty-four anti-patterns, each caught in the wild, each with its witness.
Thesis: The same dozen mistakes turn up in databases, compositors, benchmarks and research
loops. Learn to recognise them before they cost a week.
Audience: every engineer. Built to be skimmed, with one page per species.
Serves: YouTube Shorts (one species per short); the Patterns catalog's anti-pattern section,
which it extends; Academy practice prompts.
Chapters (families; each entry gives field marks, habitat, a sighting with its source, and the
antidote):
Numbers: A Number in Two Places; the self-staling header (e.computedriven.com/FINDINGS.md:19);
the hand-typed "thirteen" that was 21.
Instruments: the harness that serialised the benchmark; probes that flood the console they
measure; ptrace and rep; the cached module; the 6.4 GB diff.
Gates: the gate that can't fire; refusal that cannot fire is decoration; the bundle guard
whose falsifier found two of three holes (scripts/sabotage-super-d1-bundle.sh:4-6); --quick silently narrowing the admission record.
Authority: Ambient Authority; a comment is not an enforcement mechanism; a grant without a
policy; authority already issued.
Identity: a key is a name, not a promise; the non-injective canon; Carrier Identity;
-march=native; zip-by-index.
Scope: measuring the wrong thing's boundary (WORLD_SIZE.md §11); the bhyve line; Invisible
State.
Silence: a swallowed QuotaExceededError; False Quiescence; the rejected reply that leaves no
evidence; Ok with nothing run.
Correction: The Quiet Correction; fabricated modesty; the inference kept and struck out.
Key sources: opensentience.org/_patterns/data/patterns.json (anti_patterns) plus §2 of this
file.
Must not claim: that the catalog is complete or the families are a taxonomy (they are editorial);
a rung higher than each sighting's own. Each entry states its rung.
A newcomer's handbook: how to read this stack, run it, and add one honest claim to it.
Thesis: You can contribute in an afternoon if you learn three habits: derive, don't type;
state your rung; propose, don't rule.
Audience: developers who liked a Series One book and want to touch the code; Academy's B1
outsider.
Serves: Academy (Read → Practice → Prove); every public repo's CONTRIBUTING; the "start here"
video.
Chapters:
The map you're allowed to see. Which repos are public (box-and-box, TRVM, WRL, PARKVPS,
travel-and-rrabbit, the sites) and which are private, and why the private ones are "named and hashed, never linked".
Run the laws yourself. node test/laws.mjs && node test/compose-laws.mjs; the playground
button; measure on a fresh origin (the cached-module trap).
Replay a film. One TRVM or trvs run end to end, and its byte-exact receipt.
Boot T&R. The download, the verifier in the box, and what "untested" means on that page.
Write a claim with its rung. The five words; "when in doubt, downgrade"; a claim-ledger entry.
Write something to be attacked. A small review bundle: self-verdict, severity counts,
reproduce commands.
What you may not decide. Rule 6: propose under DOCTRINE: proposed; rulings are Travis's.
How a page gets published. Built, never hand-edited; amp-nav fan-out; the edge that rewrote a
claim (story 20).
Key sources: DOCTRINE.md, ACADEMY.md §6, academy/B1_PROTOCOL.md,
AmpersandBoxDesign/box-and-box/README.md, TRVM/README.md, travel-and-rrabbit/README.md, opensentience.org/CLAUDE.md, ampersand-nav/, STACK_COMPLETION.md §8, §10.
Must not claim: that any outsider has contributed or completed B1 (Lane B has never run); that
ProjectAmp2 is public; that a contribution path exists for Super (no installer, a private lane). Write it last among these three repos' docs. It is only true once one outsider has walked it.
Material deliberately left out of Series Two (a chapter or a Patterns entry, not a book):
The E analytics plane gets one chapter in S2-2 or a standalone long article: 15 rounds and 500
falsifiers, but nothing deployed and no written reason for keeping it private.
alkeyword and Academy go into S2-8.
ToolboxHVAC gets a single essay, "The same rule, for a furnace" (toolboxhvac.com/README.md:7-8).
Cyberutensil and alphapiler should wait until the pilot runs. It is outside the tree, uncommitted,
and has no held-out result.
weave and FPLA get a Patterns realization each, when they have a witness.
The score multiplies three things. Reader value is the size of the audience and how badly it needs the book. Evidence strength is how much of the book rests on reproduced, dated records. R97 fit asks whether the book helps Super, the MVP, or at least doesn't pull attention away from it.
| Book | Reader value | Evidence | R97 fit | Verdict |
|---|---|---|---|---|
| S2-1 Many Hands, One Tree | very high | high (in_tree records throughout) | direct: it is how Super is being built, and it ends in the Super lane | 1st |
| S2-7 The Field Guide | very high | high per entry, since each carries its own witness | neutral to good: its antidotes are Super's disciplines | 2nd |
| S2-3 A Comment Is Not an Enforcement Mechanism | high (security engineers) | the strongest in the tree: every finding was reproduced before it was fixed, and removing the fixes turns 9 of 55 checks red | indirect: stage 2 code, but the laws (authority already issued, grant ≠ policy) are the same ones ampd enforces | 3rd |
| S2-4 The Machine Under the Desk | high, and the most filmable | live_local throughout | stage-2/OS layer | 4th; write it first if video comes before text |
| S2-2 The Floor | medium (specialist) | high, and honest ("not established") | stage 1 (Fabric) | 5th, after FCB §4 is ruled |
| S2-5 The Table That Closed Zero Cells | medium | high | low, since it is research | 6th |
| S2-6 What an Agent Is | high, broad | low to medium (AGENCY is proposed) | framing | 7th, blocked on ratification |
| S2-8 Your First Rung | high, later | low until B1 runs | enabling | last, and only after one outsider has walked it |
Why this order.
S2-1 is the only candidate whose subject is stage 0 itself. Its audience, people running
agents on real code, is the same audience Book 2 targets, so the two books feed each other.
S2-7 costs the least to write. Most sightings are already verified in §2. It also yields the
most short-form video.
S2-3 has the best evidence-to-drama ratio in the tree. Its chapters are sentence-sized laws,
each with a measured counterexample. It still reads correctly while Cloud stays undeployed, because it only claims in_tree.
Interactive figures (these would extend the steps / ladder / compare / numbers blocks in books/FIGURES.md):
The x32 door. Toggle bit 30 on a syscall number and watch the deny-list miss it while the
allow-list catches it.
The 200-byte reservation. A step-through of reserve → offer → URL → abandon → offer → PUT with
the byte ledger drawn beside it. It pairs with an "authority already issued" slider showing the clock versus the credential.
0.76¹². A slider for the per-run pass rate and the number of runs per block, showing the
block admit rate. It turns FCB's lottery into one picture.
The black band. Drag the encoder padding and watch the black columns match it pixel for pixel.
Board versus critic. The five Gate 0 runs as a dot plot around zero, with the entailment
audit's lift overlaid.
The ledger over time. CLAIM_LEDGER.json statuses stacked by round, with REFUTED in its own
colour, the colour the table once lacked.
A number in two places, live. A page that reads a count from two sources and turns red when
they diverge, built on the derived law manifest.
The floor's bytes. A treemap of the 5,264-byte kernellet image against its 8,192-byte gate.
Recurring YouTube segments:
"Retracted": one banded retraction per episode (the RETRACTED lines, the register's
supersessions), read out loud with what it cost.
"One Number": one derived number, re-derived on camera (the law count, the 7.1×, the 5.3 KB).
"Field Marks": one anti-pattern in 60 seconds (feeds S2-7).
"Lab Log": the machine layer, framebuffer screenshots straight from vps.py screenshot.
"Who Ruled?": one register row per episode: who decided, and why a recommendation isn't a
ruling.
"Refused": a gate refusing on camera (FCB's preflight, refuse_narrowing, a sabotage battery
turning red).
Cross-book threads worth making explicit:
"A comment is not an enforcement mechanism" appears three times: Cloud v1 migrations (story
3), Super's sabotage battery (story 2), and AGENTS.md's required reading (story 14). Tag it across S2-1, S2-3 and S2-7.
"Authority already issued" links Cloud R55/R69, ampd's INDETERMINATE, and Book 5's
restore-as-grant.
"Measure the effect" runs through Book 6 ch6, S2-2 ch6 and S2-4 ch9.
"The instrument, not the engine" covers TRVM's six rounds, the serialising harness, probes
flooding consoles, and ptrace with rep.
Formats and series:
"What We Got Wrong," a numbered series built directly from the 19 REFUTED ledger claims, the
RETRACTED/CORRECTED bands in CLAUDE.md, and REVISION_REGISTER.md's superseded rows. Each entry states what was claimed, what refuted it, and what replaced it.
**A Library "errata" page generated from the books' What this chapter does not establish
callouts.** It would show each chapter's rung and last re-derivation date. build.mjs already parses those callouts.
An "evidence ribbon" on every chapter showing the rung distribution of the claims it makes,
derived and never typed.
Audio edition of S2-7, since the anti-patterns read well as two-minute episodes.
A "Traps" appendix for S2-4, harvested from RRABBIT/docs/RUNBOOK.md and the CLAUDE.md T&R
section (the 8911/8913 CORS trap; restart the proxy between reloads).
A companion essay, "The same rule, for a furnace": ToolboxHVAC's rule that *"the strength of
the claim on screen never exceeds the strength of the evidence behind it"*. This shows the method isn't only for AI.
A "reproduce this chapter" box on the public edition: the one command a reader can run
against a public repo, or an honest "not reproducible from public sources".
A multi-model credits page for S2-1. Which model wrote, reviewed, researched and ruled each
round (Claude, GPT-5.6, Codex, Fable), from the bundles' own headers.
A slow-TV episode: one full FCB preflight on the dedicated host, uncut, ending in a refusal or
a pass.
Book 6 ch7's source note is out of date. It says the misdiagnosis and timings are "NOT IN THE
TREE". They are now in lab/NETWORK.md:58-79 ("The first diagnosis was wrong, and is recorded so"), but that edit is uncommitted (git status: M lab/NETWORK.md). Commit it, then update the chapter's source list.
`BENCHMARK_LANE.md` has two §7.10 sections (l.962 and l.1025).
`computedriven/receipts/KERNELLET-R1.md` now reports 309 B per context after R4-P, beside the
historical 312 B. Quote the one that matches the date.
Four 0-byte files at the root: undeclared, id, ID and d, all created 2026-08-22 15:53
and committed in c1c7d82. They look like shell-redirect accidents. This is a guess, so check before removing them.
Three memory notes are out of date:
The Cyberutensil note says 52 laws, but cu.py laws reports 54.
The docs-redirect note implies the redirect list was being applied, but docs/REDIRECTS.md:3
says "NOT ADOPTED — decided against 2026-08-13".
The MEMORY.md index line for sharpening ("12 of 14 moved odds DOWN") is stronger than its own
note's body.
`e.computedriven.com` is labelled "private" only in the series plan's appendix. Its own brief
calls it the "Canonical public product/domain". Rule on this before any book describes it.