The Five Arcs — implementation plan

The journey, at a glance

Thirteen stops, preflight to cutover. Click any to jump — the picture is the index.

Assumptions standing in for rulings

Nine calls nobody has ratified yet. The workflow builds on all of them — flip one and only the named build item changes.

The build, start to finish

One path, in order. Each stop opens to its build items underneath.

1. Preflight Before anything builds, the machinery checks itself: enough disk, a clean workspace, the base branch cut and pushed, the deployed code matching what’s on the branch, and a starting gate that means something. If any of that is wrong, the run stops before spending a single build.
   ↘ the mechanical checks

   Preflight (in-harness, mechanical, stops the run if red): bay doctor; disk headroom ≥ 20 GB; git worktree baseline clean; base branch five-arcs/base cut from main, the Codex branch merged (B0.2) — conflicts go to a grok repair leg — and pushed; live-source parity receipt; gate baseline timed; Plato change briefs for the eleven seam files written into the sidecars.
2. What the readers found Six readers spent the morning checking what’s actually true on the live system, not what the docs say. Nine findings changed the plan before a line of it was written.
   1. 1. The team's Slack replies are already coming from code that has never been merged.
      ↘ the receipt

      src/runtime/neo-team-control.ts (1,122 lines, 74 assertions) lives only on codex/neo-control-plane; the Mac-side dispatcher is an untracked script under ~/Documents/Codex/…/work/, running as com.robert.neo-team-dispatcher, and it delivered real Slack replies Thursday night. B0.2 lands the branch first; preflight proves the live version matches it.
   2. 2. The starting gate itself is broken.
      ↘ the receipt

      npm run gate is red on clean main (two unallowlisted person: literals in worker.ts:775-776) and typecheck fails independently (scripts/corpus-decisions-run.ts:56). B0.1 repairs the baseline so every later red means something.
   3. 3. Today's assistant runs on a different engine than the one this plan builds.
      ↘ the receipt

      The dispatcher's engine is Codex (codex app-server --stdio), authenticating with the Telegram webhook secret from eve/.dev.vars. The Claude replacement is a different engine on the same doors, with its own credential (B1.2).
   4. 4. The Slack app's own paperwork doesn't match how it actually behaves.
      ↘ the receipt

      “Writes are bot-token on every path; reads use Robert’s personal user token; the on-disk manifest still describes an app sending “as Robert” with user scopes only and no bot scopes. The event door handles only app_mention. main carries a live fetch-binding bug in slack-write.ts the unmerged branch already fixes. B2.1 proves identity by live readback and a negative test, never by the file.
   5. 5. There's no durable record of what Neo has actually told people.
      ↘ the receipt

      The Telegram ring holds 20 turns/24h; the outbound ledger keeps hashes only — no durable store of the claim text itself. B4.1 builds the claims ledger before B4.2 can audit anything against it.
   6. 6. The company's own standing rules live in an uncommitted folder.
      ↘ the receipt

      ~/.claude is a GitHub repo with 1,315 uncommitted changes including the constitution; ADVERSARY-CHARTER.md was never committed; session memory is gitignored; GENOME-ROADMAP.md is cited by ratification number in code and in every deposited fact, and exists nowhere.
   7. 7. Three claims in the spec don't hold up against what's actually on disk.
      ↘ the receipt

      “1,400+ genome events” is unverifiable (340 mirrored rows, 47 days stale). “Southtown was the accidental cold-start prototype” was a loose metaphor — it is a client-site migration. Provenance actually lives in src/runtime/world-model.ts, not src/contract/world.ts.
   8. 8. There is now a tool that can measure blast radius instead of guessing at it.
      ↘ the receipt

      plato brief <file> emits blast-radius JSON — work candidates in Arc 3 carry a real blast radius (B3.1) instead of a claim.
   9. 9. The scheduling pattern this plan needs already exists elsewhere in the system.
      ↘ the receipt

      Scheduled sessions run as cloud scheduled-tasks routines on Robert's seat, not launchd — the daily audit and weekly calibration already work this way. The Codex automations map onto it 1:1 (B1.5); the low-latency team loop still needs the Mac dispatcher (B1.2).
3. Phase 0 — foundation Two repairs, done first and serially, because everything else depends on them: fix the broken gate, and land the branch that’s already running production onto a base every court can see.
   ↘ go deeper — the build items

   ID	Build	Engine	Touches	Done when
   B0.1	Gate baseline repair — clear the two pre-existing reds	grok	src/runtime/worker.ts:775-776 (mint the ids or allowlist with a reason), scripts/corpus-decisions-run.ts:56	npm run person-id:gate, npm run typecheck, npx vitest run all exit 0 on the base
   B0.2	Land the Neo control plane — merge codex/neo-control-plane; the first non-author court it has ever had	keeper merge + grok conflict repair if needed; claude courts	the 11-file diff (worker.ts +383 is the conflict surface)	gate green on the merged base; live-source parity receipt: the deployed version id appears in the branch's DEPLOY-ANCHOR.md entries

4. Arc 1 — the brain moves to seats The brain moves off the metered API and onto seats already paid for — Telegram, the morning brief, and Neo’s own standing orders extracted from the Codex thread that’s been running it, so nothing has to be pasted by hand.
   ↘ go deeper — the build items

   ID	Build	Engine	Seam (from inventory)	Done when
   B1.1	Judgment queue — generalize NeoRequestLedger into typed judgment tasks: kind ∈ {neo_team_request, telegram_reply, one_converse, morning_brief, ambient_triage, ambient_investigate, work_candidate_review, dispatch_build}, lease/claim/receipt/verdict doors, heartbeat, fail-closed-to-silence (a lease that dies N times is failed, recorded as machinery failure — never rendered as an answer)	grok	neo-team-control.ts (claim SQL :535-549), new src/runtime/judgment-queue.ts, admin routes in worker.ts	a task enqueued by a test client is claimed, answered, receipted, and readable; expired leases reclaim; zero api.anthropic.com on the new path
   B1.2	Dispatcher in-repo, Claude engine, own credential — scripts/neo-ops/dispatcher.mjs (port of the Codex dispatcher; wake = claude -p headless session with a versioned system prompt, fresh per wake, state from ledgers never scrollback), scripts/neo-ops/door.mjs (the session's only hands: read-thread, reply, transition, world-query, receipt — all Worker doors), new secret NEO_CONTROL_ADMIN_TOKEN (Worker) + ~/.config/neo-ops/env (never .dev.vars), launchd plist com.digitalempathy.neo-ops-dispatcher, install/start/stop/status/rollback script, dead-man heartbeat (call-log pattern: Worker alarms if the dispatcher is silent > 15 min)	claude (security-domain)	the Codex dispatcher's loop (runOnce: health → claim → wake → receipt); call-log/worker.ts:52-99	the 11 ported assertions + new ones pass; --once against a local Worker round-trips a task; no Slack write when health is unsafe; the Telegram secret is not read anywhere in the new code
   B1.3	Neo Ops session definition — docs/neo-ops/RUNBOOK.md (wake contract, silence bias, the authority charter lifted verbatim from DEFAULT_NEO_TEAM_POLICIES, context-firewall rules, receipt requirements), docs/neo-ops/prompts/neo-ops.system.md with a version stamp every verdict carries	claude (team-facing words — draft until Robert earns them)	neo-team-control.ts:59-86	a fresh sonnet following the runbook answers a fixture request correctly (walk); prompt version appears in every verdict payload
   B1.4	The face on seats — ingestTelegramUpdate, diagThinkTurn, and morningBrief enqueue judgment tasks instead of calling the in-Worker brain; the Worker keeps transport, idempotency, the outbound chokepoint; flag FACE_ON_SEATS (default off) with the in-Worker path as fallback; honest “Neo is offline” system line on machinery failure, never silence	grok	worker.ts:15066 (Telegram), :16434 (One), :10370 (brief); brain.ts:940,2207	with the flag on and a stub seat answering, a Telegram update round-trips through the queue; with the flag off, byte-identical behavior to today (existing 30+ conversation tests green)
   B1.5	DailyHQ extraction — automations → docs/neo-ops/runbooks/<id>.md (prompt verbatim, schedule, status, proposed Claude routine); memory task-groups tagged Daily HQ / chief-of-staff lanes → docs/neo-ops/knowledge/; last 14 days of the thread's assistant turns → docs/neo-ops/standing-state-2026-08-23.md; obligation sheet → scripts/neo-ops/import-obligations.mjs (reads a CSV the keeper exports; posts to the ask ledger door); retirement note	claude (private data)	~/.codex/automations/*/automation.toml, ~/.codex/memories/MEMORY.md, the thread rollout JSONL, the sheet id in the morning-command-brief prompt	every automation has a runbook; the excluded personal task-groups are listed by title; the standing-state doc cites thread timestamps

5. Arc 2 — Neo faces the team Neo starts talking in the channels you name — proving who it’s speaking as, carrying receipts on every claim, and noticing things in the room without being asked, at a volume you control with one dial.
   ↘ go deeper — the build items

   ID	Build	Engine	Seam	Done when
   B2.1	Identity proof — (a) negative test: inject a user token everywhere and assert no write call ever carries it (all three post sites + the Neo reply door); (b) /admin/slack-identity secret-gated live readback (auth.test with the bot token → bot user id, team, granted scopes; never the token); (c) scripts/slack-app-manifest.neo.json — the proposed bot-scoped manifest with event subscriptions (app_mention, message.channels), for Robert to apply in Slack admin; (d) apply the fetchImpl.bind fix; (e) route the two stray posts through SlackPostExecutor so the chokepoint claim is true	claude (security-domain)	slack-write.ts:105, eve-slack-lane.ts:1705,2219, branch postNeoTeamReply, scripts/slack-app-manifest.json	the negative test goes red when a write path is sabotaged to use the user token (planted-break proof); readback route returns metadata only
   B2.2	Receipt-source registry — src/runtime/receipt-sources.ts: per source {identity it reads as, freshness semantics, trust-table fields, how to cite, what questions it can answer}; includes deploy-readback (checked Mac-side via wrangler deployments status, tier T0); conformance test: every ALL_SOURCES entry and every TRUST_TABLE field maps; the doc is a generated projection of the code	grok	sources/composite.ts:69-77, contract/trust.ts:91,133 (additive only — frozen contract)	conformance test green; docs/ai/design/receipt-source-registry-2026-08-23.md regenerates byte-identical
   B2.3	Tone contract + “@Neo check” — docs/neo-ops/tone-contract.md (receipts never people · offer-to-recheck · concede when corrected, never argue · tier-honest hedging · mute honored forever); runbook section for a check (claim → registry → door → receipt + tier); fixtures prove the mechanism (every reply carries a receipt ref and a tier label)	claude	registry (B2.2), runbook (B1.3)	fixtures green; walk: a fresh sonnet answers a fixture check with a receipt and the right hedge
   B2.4a	Ambient observe — event door handles message.channels/message.groups (eligibility: allowlisted channel + mode, human message, supported subtype, self/bot exclusion, event_id dedupe, edit/delete revisions); NeoAmbientObservation (content hash long-lived; text only for the pending window); conversation lease; delay-and-recheck window (vars NEO_AMBIENT_DELAY_MIN_MS, _MAX_MS); at window open, enqueue one ambient_triage task per thread carrying the posture dial as question framing; verdict recorded (silent / candidate{class, reasons, evidenceRefs}) — no score, no threshold	grok	slack-events-door.ts:145-168, branch control plane, B1.1 queue	fixtures: 24 golden scenarios produce the expected observation/lease/task records with every write recorded; lease races and dedupe replays hold (deterministic suite)
   B2.4b	Ambient reply — candidate → promotion into the request lifecycle → ambient_investigate task → reply through the existing idempotent door with write-time gates re-checked at write (kill switch, posture ≠ observe-only, budget NEO_AMBIENT_BUDGET_PER_CHANNEL_DAY, lease owned, trigger unchanged, a human answered first → cancel, firewall, receipt for any completion language); circuit breaker NEO_AMBIENT_DAILY_CEILING → observe-only; mute list honored forever; health readback fields	grok	B2.4a, reply door (branch :1206), hand-drain.ts:81-105 dual-post pattern	the zero-tolerance suite: 0 writes with the kill switch on, 0 duplicate replies, 0 writes over budget, 0 writes as a user token, 0 completion replies without a receipt; each proven by a planted break
   B2.5	Ambient judgment prompts — triage + investigate sections of the runbook: the three receipt-backed classes, the posture framings (what receipts-only asks vs helpful), typed JSON output with reasons and evidence refs, fail-closed-to-silent on any machinery doubt (recorded as machinery, not as an answer)	claude	B1.3, B2.3	a fresh sonnet given the golden-24 as fixtures stays silent on 6–8, 21–24 and proposes on 1–5 (walk, not a statistic)
   B2.6	Rehearsal harness — scripts/neo-ops/rehearse-ambient.ts: replays (a) the golden-24 fixtures and (b) real channel history for the allowlisted channels over N days (fetched through the Worker's reader; local only, never committed) through the REAL eligibility → observation → window → triage path, with the judgment leg executed by a local claude -p on the seat and every write seam recorded; outputs artifacts/neo-ops/rehearsal-<date>/{prevalence.json, decisions.jsonl, disagreements.md}; loads the worker's toml [vars] and fails closed on a missing key (rehearsal-config parity)	grok builds; claude runs (private data)	B2.4a/b, scripts/daily-independent-audit.ts:100-112 (config-parity pattern)	the run over the last 30 days produces the prevalence number and the disagreement list; the list is attached to the ledger for Robert
   B2.7	Digest renderer (conditional on B2.6 < 3/week) — a scheduled ambient_digest task renders the day’s candidates into one ignorable post	grok	observation ledger	one post per day, budget-counted, posture-gated

6. Arc 3 — senses meet hands A finding about the company can turn into a real piece of work — with an honest blast radius, built by the fleet, checked by an outside court, and landed back as proof, not opinion.
   ↘ go deeper — the build items

   ID	Build	Engine	Seam	Done when
   B3.1	Work candidate — src/runtime/hands/work-candidate.ts (mintWorkCandidate fail-closed like mintSignalCandidate: finding, evidence refs with verbatim-quote floor, proposed work class, blast radius from plato brief JSON or unknown: plato unavailable, authority class mapped onto NeoAction, rollback plan); candidate_json column on eve_asks via the backfill idiom; doors to mint/list; work_candidate_review judgment kind	grok	ask-ledger.ts:625-666, signals/types.ts:107-146, plato brief	a candidate with a fabricated quote is refused with a named reason; a real one lists a real blast radius
   B3.2	The warp hand — third Hand cartridge (investigate automatic read-only; build_on_branch standing-or-Robert; deploy never); act enqueues a dispatch_build task for the Neo Ops session (the only thing with hands), which dispatches through the engine-bay skill and returns {branch, ledger row, report, court verdict}; dual-post receipts; /admin/world/receipt door → writeBack({provenance:“ground”, source: receipt ref, supportingSpan}) with derivedFrom: candidate id; runbook section “dispatching to warp”	grok (Worker + hand) · claude (runbook)	hands/hand.ts:129,177, hand-drain.ts:112-190, world-model.ts:2324	EXPECTED_HANDS_COUNT = 3 and conformance green; a deploy request is refused at the actionKind gate; a receipt lands as a ground entry with a derived-from edge
   B3.3	First-loop rehearsal — a synthetic finding → candidate → dispatch_build → a Claude leg standing in for Neo Ops dispatches a trivial real bay build → one deepseek verify → receipt lands in a local WorldStore → the close-out line is composed and recorded, not posted	claude runs	B3.1, B3.2, the bay	the loop's ledger is on disk with every hop's receipt

7. Arc 4 — the truth spine Every claim Neo makes gets written down, and a weekly audit checks a sample of them against reality — replacing the honesty check that has never once passed.
   ↘ go deeper — the build items

   ID	Build	Engine	Seam	Done when
   B4.1	Claims ledger — durable outbound-claims store (text, channel, at, task id, receipt refs) hooked at the Telegram outbound chokepoint, the Neo reply door, and the morning-brief render; retention var (30 days); /admin/claims?since= readback	grok	outbound.ts:29-31, reply door, worker.ts:7199-7209	a send appears in the readback with its text and refs; retention evicts on schedule (test)
   B4.2	Weekly honesty audit (Option C) — scripts/weekly-honesty-audit.ts in the daily audit's shape: sample N claims from B4.1 (+ call-log receipts), recompute each against ground through the registry (mechanical where possible; judgment rows flagged for the session), write artifacts/audit/honesty-<date>.md, bank a v2 verdict line in docs/ai/how-calibration-history.md, post one line to #ops-infra, stamp the dead-man; schedule-calibration.ts marked superseded	grok	scripts/daily-independent-audit.ts (structure), B4.1, B2.2	a run against a fixture ledger produces a verdict line and a flagged-rows section; a planted false claim is caught
   B4.3	Wire c4-meta-gate — into groundSynthesisOrFallback with the existing EntailmentLeg; the verdict is recorded on the card; a not-entailed synthesis falls back to the raw claim (nothing hidden)	grok	workqueue/mirror-synthesizer.ts:212-273, c4-meta-gate.ts, post-gen-audit.ts (leg pattern)	Plato backward reach of c4-meta-gate.ts > 1; a real-quote/false-conclusion fixture falls back
   B4.4	Genome receipts + labels — per-call-site GenomeReceipt{fenced:false, proof} where a re-runnable proof exists (ask-ledger state events, signal receipts, call-log SF ids); honest fence reasons elsewhere; /admin/genome/labels census (fenced vs receipted, queryable); insert() counts ground entries lacking a supporting span and exposes the count on /admin/world — no new rejection	grok	genome-deposit.ts:225-229,351, world-model.ts:1147	census readable; a sampled week (fixture) shows zero unlabeled claims; insert() behavior unchanged for every existing test
   B4.5	Two SF postures + audit-cadence tile — docs/ai/design/sf-write-two-postures-2026-08-23.md; One-Surface census lane “audit cadence” (last daily audit, last weekly honesty audit)	grok (tile) · claude (doc)	sf-write-rail.ts, sf-rest-executor.ts, one-surface-html.ts:503-515	doc names both lanes with flags and justifications; the tile renders from /one/api/plate

8. Arc 5 — the germline The company’s own standing rules and skills get a home, a map of what’s actually there, and a rehearsed way to rebuild from nothing if they’re ever lost.
   ↘ go deeper — the build items

   ID	Build	Engine	Seam	Done when
   B5.1	Germline manifest + census — genome/GERMLINE.md + genome/germline.manifest.json (every heritable artifact: home, repo, VC status, owner, ratification pointer); scripts/germline-census.mjs recomputes VC status live and names absent citations (GENOME-ROADMAP.md) — legibility, never a gate	grok	inventory/arc5.md manifest	census runs clean on this machine and names the dirty files and the absent citation
   B5.2	Cold-start drill — scripts/cold-start-drill.sh: fresh temp HOME, clone the two germline repos at pinned refs, materialize what can be materialized offline (bay setup, hooks, skills, templates, charters), score categories re-grown as a percentage with named gaps, plant one gap and prove it is named; artifacts/cold-start/<date>.json	grok builds; claude runs drill #1	scripts/rebirth-drill.ts (the four-move shape)	drill #1 receipt on disk with a score and named gaps; the planted gap is named
   B5.3	Rulings index — genome/RULINGS.md seeded from the scattered ruling docs, court briefs, and memory files (pick + why + pointer + date), DRAFT for ratification; a quarterly dead-gene review routine definition	claude	inventory/arc5.md §4	every entry has a pointer that resolves; marked DRAFT

9. Courts Nothing ships on a green test alone. Every build gets an outside judge who never wrote it, and the judging itself follows a written charter so verification can’t quietly turn into another layer of rules policing people.
   ↘ the charter rules the courts are cut from

   1. Kind-(a) findings are never auto-repaired. A court finding classified as “this change adds a judgment-constraining rule” is parsed into the YOUR CALL ledger for Robert. No builder is dispatched to “fix” it.
   2. Repair briefs carry the cure-shape guardrail. A repair may add receipts, legibility, reversibility, say-it-out-loud, or a write-time invariant on facts. It may not add a threshold, a score, a filter, an auto-defer, or a new refusal on a teammate-reachable path.
   3. Two rounds, then the frame. A build gets at most two repair rounds. A second refutation on the same surface escalates the frame question to Robert; no third cleverer patch is dispatched.
   4. The drift sentinel. After all repairs, a cold opus leg reads the cumulative diff of the integration branch against base and lists every code path that decides a meaning question, tagged kind-(a) or kind-(b), plus the teammate-path refusal-site census before and after. Any kind-(a) or any census rise marks the PR HELD for Robert.
   5. Courts prove mechanism; data proves judgment (ratified 2026-08-18). Three rehearsals on real data run inside the shot: ambient over 30 days of real channel history (B2.6), the first warp loop (B3.3), cold-start drill #1 (B5.2). Their disagreement lists are products, not scores.
   6. The APEX walk outranks the suite. Every consumer-facing increment is walked on the real artifact by a fresh mind: a sonnet that has never seen the build follows the runbook, the tone contract, the registry doc, the dispatcher install script, the manifest. A green suite under a failed walk fails.
   ↘ who judges what

   Stakes	Items	Builder	Verify	Falsify	Extra lenses
   Security-domain (credentials, posting identity, admin doors, context firewall)	B0.2 courts, B1.2, B2.1	claude (sonnet)	opus	opus (a second instance)	none non-Anthropic — the guard fences them
   Trust machinery (queue, face, ambient gates, warp hand, claims, audit)	B1.1, B1.4, B2.4a, B2.4b, B3.2, B4.1, B4.2	grok	opus	deepseek-pro	on divergence: glm-5.3 + ox-alpha panel (diff under judgment is retained by those providers — Robert's named choice)
   Ordinary code (registry, candidate, meta-gate, genome labels, tile, census, drill)	B0.1, B2.2, B2.7, B3.1, B4.3, B4.4, B4.5, B5.1, B5.2	grok	deepseek-pro	— (solo, per the record)	kimi-k3 lens on any RED for a second pedigree
   Docs, runbooks, prompts (judgment-adjacent, team-facing)	B1.3, B1.5, B2.3, B2.5, B5.3	claude (sonnet)	walk by a fresh sonnet	kimi-k3 “would a stranger succeed?” lens	opus only on the tone contract and the authority charter

   Reviewer is never the author: a grok build is never judged by grok; a Claude build is judged by a different Claude instance; an opus verdict is never the leg that wrote the brief. Divergent verdicts are HELD, never averaged.

10. Walk Before anything merges, six people who never built it try to follow it cold — the runbook, the tone contract, the registry doc, the install script, the manifest — each on that item’s own branch. A passing test suite under a failed walk still fails.
11. Integrate All the separately-built, separately-judged pieces get merged together, in the order they depend on each other, gated again after every merge, until one branch holds the whole shot.
    ↘ go deeper — the workflow, phase by phase

    1. 0Preflight (in-harness, mechanical, stops the run if red): bay doctor; disk headroom ≥ 20 GB; git worktree baseline clean; base branch five-arcs/base cut from main, the Codex branch merged (B0.2) — conflicts go to a grok repair leg — and pushed; live-source parity receipt; gate baseline timed; Plato change briefs for the eleven seam files written into the sidecars.
    2. 1Build (pipeline): each item’s brief → its engine → harvest bay/<slug> → gate in a clean worktree → push five-arcs/<id>. The code chain runs serially, each item on its predecessor’s branch, so a build inherits every sibling it depends on — conflict-free and test-complete by construction: B0.1 → B0.2 → B1.1 → B2.2 → B1.2 → B2.1 → B2.4a → B2.4b → B1.4 → B2.6 → B3.1 → B3.2 → B3.3 → B4.1 → B4.2 → B4.4 → B4.5. The docs chain (B1.3 → B2.3 → B2.5) and the independents (B1.5, B4.3, B5.1 → B5.2, B5.3) fan out off the merged base in parallel. B2.7 is built on the integration branch only if B2.6’s number says so.
    3. 2Court (pipelined per item, no barrier): verify ∥ falsify per the composition table → parse verdict, altitude, Environment section, YOUR CALL entries → RED at instance/class → repair (max 2) → re-court by a non-author → divergence → panel → still divergent → HOLD.
    4. 3Walk: the six walks, each by a fresh sonnet on the item’s branch.
    5. 4Integrate: merge five-arcs/<id> branches into five-arcs/integration in dependency order, gate after each merge (conflict → a headless sonnet merge leg), full gate at the end.
    6. 5Rehearse (on the integration branch, where code and docs meet): the three real-data runs — ambient over 30 days of real history, the first warp loop, cold-start drill #1; B2.7 built, courted, and merged if prevalence < 3/week.
    7. 6Sentinel: the cold opus drift audit over the integration diff.
    8. 7Deliver: one PR five-arcs/integration → main whose body is the verdict ledger (per item: builder, engines, verdicts, altitude, repairs, walk, Environment); the HANDOFF doc; the YOUR CALL list; genome deposits for the ratifiable artifacts.

    Dispatch mechanics (verified with the warp desk, 2026-08-23): no courier agents. The orchestrator is a standalone Node script that calls the bay’s two direct doors — `durable-sealed.sh` (the sealed room on this Mac: `grok-build`, `opus-verdict`) and `sandbox-lane-worker.sh` (off-box: `grok-46-sandbox`, `deepseek-pro`, `kimi-k3-sandbox`, `glm-53`, `ox-alpha`) — and headless `claude -p` legs (sonnet builds in their own worktrees, walks, the real-data runs). Every dispatch leaves a ledger row with its `run_id`, report path, and harvested branch; briefs stay compact (~2 KB), and the full work order is the spec sidecar on the pinned branch.

    Fleet discipline baked in: every brief names `security_domain: false` explicitly or routes to Claude; every court brief carries `Brief-Class: read-only`; the filled brief stays under 5,120 bytes with the sidecar doing the work; infra failures retry once then reroute to a substitute engine of the same independence class; a report on disk is read even when the lane classed the row `NOTHING_TO_DO`; named-path `git add` only, never `-A`; nothing deletes.

    Size and cost: ~25 builds, ~50 court legs, ~10 repairs, 6 walks, 3 rehearsals, 1 sentinel, ~6 integration legs — roughly 100 agent legs, well past the session’s default “under 15” guideline, per Robert’s ask for thoroughness. Builders run on the grok seat (flat). Sandbox judges are metered on the org meter. Opus legs spend Claude usage: held to the security and trust-machinery rows, roughly 20 legs. Wall clock: 8–10 hours — the code chain is serial on purpose (17 items × 25–40 min), with the docs chain, the independents, and every court pipelined alongside it.

    What the workflow never does: deploy, merge to `main`, post to Slack or Telegram, touch the live Codex dispatcher or its plist, delete a branch or a file, change `~/.claude`, or read a secret into a network call from this session (the identity readback is executed by the keeper at cutover).

12. Rehearse Once the pieces are merged, three parts get tested against real data instead of fixtures: a month of real Slack history run through the noticing path, one real finding pushed all the way through a real build and a real outside check, and a rehearsal of rebuilding the company’s own rules from a blank machine. If the noticing path stays quiet, the daily digest gets built, judged, and merged in the same run instead.
    ↘ the three rehearsals

    Rehearse (on the integration branch, where code and docs meet): the three real-data runs — ambient over 30 days of real history, the first warp loop, cold-start drill #1; B2.7 built, courted, and merged if prevalence < 3/week.
13. Sentinel One last read, by a mind that built none of it: does anything in the finished diff quietly decide a judgment call in code instead of asking a person? If yes, or if it’s harder to say no to something than it was before, the whole thing holds for Robert instead of shipping.
    ↘ the drift sentinel, in full

    After all repairs, a cold opus leg reads the cumulative diff of the integration branch against base and lists every code path that decides a meaning question (classifier, threshold, score, filter, auto-defer, forced default), tagged kind-(a) or kind-(b), plus the teammate-path refusal-site census before and after. Any kind-(a) or any census rise marks the PR HELD for Robert. This is the report Robert asked for, as a mechanism.

14. Cutover Nothing here goes live in this shot. Cutover is a separate, later checklist the keeper runs after Robert has read the ledger — ten steps, each with its own undo.
    1. Merge. five-arcs/integration after reading the ledger; any HELD item stays on its branch. Undo: Revert merge commit.
    2. Slack app. Robert applies scripts/slack-app-manifest.neo.json in Slack admin (bot scopes + event subscriptions) and reinstalls; the keeper runs the identity readback and pastes the receipt (bot user id, scopes). Undo: The previous manifest is exported first.
    3. Secrets. NEO_CONTROL_ADMIN_TOKEN minted and set on the Worker; ~/.config/neo-ops/env written by Robert (the keeper never sees the value). Undo: Delete the secret; the old door still honors the Telegram secret until step 5.
    4. Deploy. Through scripts/deploy-gated.ts (baseline smoke → promote → post-smoke), after DEPLOY-ANCHOR.md is brought current (the anchor discipline is RED today, 26 days stale). Flags at deploy: FACE_ON_SEATS=0, NEO_AMBIENT_POSTURE=observe-only, budget 1, #bot-test only. Undo: npx wrangler rollback -y.
    5. Dispatcher swap. Install the Claude dispatcher (scripts/neo-ops/install.sh), confirm --once round-trip on a test request, then launchctl unload the Codex plist (kept on disk 30 days). Undo: launchctl load the Codex plist; stop the Claude one.
    6. Face on seats. Flip FACE_ON_SEATS=1; the keeper walks a Telegram exchange and the morning brief; Robert’s first real morning decides. Undo: Flip to 0.
    7. Social contract. Robert posts the team message (drafted in B2.3's deliverable) before the dial leaves observe-only. Undo: —
    8. Dial. After the first clean observe-only week (the observation ledger shows what Neo would have said), Robert turns receipts-only. Undo: The dial.
    9. Routines. Register the weekly honesty audit and the extracted Codex automations as scheduled routines; delete the traps-mode routine. Undo: Re-register.
    10. DailyHQ. Robert reads docs/neo-ops/standing-state-2026-08-23.md and the obligation import diff; the thread is archived, not deleted; the Codex automations are paused (they already are, mostly). Undo: —

Risks that remain after all of this

Honest, not hidden.

What I need from you

No new open questions — three decisions only you can make.

Reference strip

By type, for lookup — every link resolves to an atom above.

Arcs — 1 · 2 · 3 · 4 · 5

Items — B0.1 · B0.2 · B1.1 · B1.2 · B1.3 · B1.4 · B1.5 · B2.1 · B2.2 · B2.3 · B2.4a · B2.4b · B2.5 · B2.6 · B2.7 · B3.1 · B3.2 · B3.3 · B4.1 · B4.2 · B4.3 · B4.4 · B4.5 · B5.1 · B5.2 · B5.3

Courts — the gate · charter rules · who judges what

Cutover steps — 1 · 2 · 3 · 4 · 5 · 6 · 7 · 8 · 9 · 10

Risks — 1 · 2 · 3 · 4 · 5 · 6 · 7

Receipts and provenance

• Inventories: five-arcs-build-2026-08-23/inventory/{arc1..arc5,infra}.md — six sonnet readers, 2026-08-23, main @ c290101f0; Plato index @ 45ebff4cd (19 behind) where cited.
• Codex body: ~/.codex/automations/*/automation.toml (9), ~/.codex/memories/, ~/.codex/sessions/2026/07/29/rollout-…-019faebe-fdb3-…jsonl (804 MB, the DailyHQ thread), ~/Documents/Codex/2026-08-14/neo-control/work/neo-team-dispatcher.mjs + plist; launchctl list shows com.robert.neo-team-dispatcher pid 33101.
• Bay record: ~/.claude/engine-bay/ledger.jsonl, last 7 days (grok-46-sandbox 21 ok; grok-build 5 unreported; deepseek-pro 19 ok / 23 nothing-to-do; glm-53 20 ok; ox-alpha 5 nothing-to-do; kimi-k3-sandbox 12 ok / 30 nothing-to-do).
• Standing law applied: constitution (trust core, frame laws, code-never-decides and its 08-18 court corollary), the Adversary Charter and its 08-10 one-way valve, dispatch-economy (independence + fit), courts-prove-mechanism-data-proves-judgment (08-18), rehearsal-config-parity (08-18), capture-first-over-worst-case (08-18), live-source-parity-before-extraction (08-20), fleet-rooms-share-your-checkout (08-20).