Skip to content

Chapter 27 — Checking this book yourself

Status: INTERNAL (handbook-only) · Prerequisites: all prior chapters

Mental model

A book about honesty must be falsifiable. When observation disagrees with text, precedence is: executable code wins, then the Wave-4 correction layer, then canonical audit ledgers, then this handbook.

Explanation

Why this chapter exists

Every handbook chapter carries a CHECK — a runnable observation that can refute a load-bearing claim. This chapter collects them into a protocol and states what to do when code and prose diverge.

Precedence rule

  1. Code at cited paths/lines.
  2. Wave-4 engineering findings (docs-audit/ENGINEERING-FINDINGS.md, EF-*).
  3. Canonical ledgers (TRUST-MODEL.md, ENTRYPOINT-MODEL.md, etc.).
  4. This handbook.

A disagreement is first a book bug until re-reading the cited source decides otherwise. "The tool is broken" is not the default conclusion.

Register of checks (by chapter)

Ch Tag Check summary
3 LOCAL OCCAM_RECEIPTS=off — key still in ~/.occam/keys/
4 LOCAL OCCAM_LOG: transcode vs probe stderr work differs
5 NETWORK 404 no browser; public-ref no browser; SPA has browser attempt
6 LOCAL Same session_profile: transcode vs probe storageState
7 NETWORK Two max_tokens → different contentHash + compile.omitted
8 NETWORK cache_ttl_s: fragment #a vs #b collision risk
9 NETWORK search default is DuckDuckGo; off fails closed; map ≤64 links
10 NETWORK Digest with one 404 — per-item failure, reduced receipt
11 NETWORK playbook auto vs off diff
12 LOCAL Edit unsigned v1 provenance.verify.score — still "verifies"
13 NETWORK extract ignores max_tokens; telemetry confidence 0.0
14 LOCAL receipts off — hash, key, playbook save still signs
15 LOCAL Foreign receipt via MCP without pubkey; unsigned watch chain
16 NETWORK Paraphrase claim_check miss
17 LOCAL crosscheck visible under reader + consensus flag
18 LOCAL tools/list per profile
19 LOCAL keys export empty dir mints key
20 NETWORK OCCAM_LOG browser automatics
21 LOCAL Filesystem diff vs state inventory
22 NETWORK proxy ignored by probe
23 LOCAL extract_knowledge vs transcode on private IP
24 NETWORK Three rejected chains
25 LOCAL unset OCCAM_HOME → workers_unavailable
26 LOCAL dead-register grep
27 meta Refute one sentence without reading C#

Tag NETWORK: checks against live third-party sites decay — pair with LOCAL alternatives where listed.

Tag SOURCE-PROVEN: some claims were code-proven but not runtime-reproduced in audit (e.g. WS pool kill, fragment cache collision, Docker health scope, certain tamper constructions). Treat as hypothesis until you reproduce locally.

Running checks as one session

  1. Stand up install (Chapter 3).
  2. Walk Task R thread: transcode → probe → digest → trust tools → operator verbs.
  3. Record each CHECK outcome in a log: PASS / FAIL / NETWORK-SKIP / SOURCE-PROVEN-PENDING.
  4. On FAIL: cite handbook sentence, observation, and code path you looked up.

Reading a disagreement

Observation Likely meaning
CHECK fails consistently Handbook or public docs bug — fix docs after code confirm
CHECK flaky on NETWORK Site changed — update example URL or mark NETWORK-deprecated
Code changed since audit Expected drift — re-run gate, update handbook
CHECK passes but contradicts older docs Older docs wrong — handbook wins if code-aligned

What is not checkable (honest limits)

  • Tokenizer error bounds — unmeasured; no check asserts token counts.
  • Truth / origin / identity — no check can prove these; forbidden claims stay forbidden.
  • Third-party site behavior forever — NETWORK checks expire.
  • GitHub org branch protection — marketplace trust (OD-1) requires external evidence.

Success criterion for this chapter

You can refute at least one handbook sentence using only observations and cited code paths — without accepting prose on faith.

CHECK

META. Pick one load-bearing claim from Chapter 14 or Chapter 20. Design an observation that would falsify it. Run the observation. If it passes, state exactly what the claim still does not license you to say.

Common misconception

"If the book and the tool disagree, the tool is broken." The handbook is downstream of code. Disagreement is first a book bug; only re-reading cited path:line decides.

Limitations

  • Falsification protocol does not replace the integration gate for release claims.
  • Some CHECKs require operator care (refresh kill, connect mutations).
  • Audit incompleteness is documented in docs-audit/CANONICAL-AUDIT-INDEX.md — not every EF has a CHECK yet.
  • All handbook chapters — per-chapter CHECK fields
  • docs-audit/DISCOVERABILITY-GATE.md §3 — mechanical doc checks
  • docs-audit/CANONICAL-AUDIT-INDEX.md — known incompleteness
  • User docs lint: node scripts/check-docs.mjs