How Digital Empathy becomes one agent — a world model that cannot lie, intelligence pressed into the seams, a genome that compounds — in five arcs starting from what is already alive.
Jump by type — Decisions: the eight locks · Contracts: judgment queue, receipt registry, work candidate · Gates: arc 1, arc 2, arc 3, arc 4, arc 5 · Repairs: watchdog · Questions: Q1–Q6
Implementation plan → five-arcs-implementation-plan-2026-08-23.html
The company is being built as one agent. Its perception is a world model that cannot lie: every fact carries where it came from, how fresh it is, and how much to trust it — claims stay labeled as claims until a receipt promotes them. Its action is intelligence pressed into the seams where companies actually stall — and seams run three deep: seams of work (handoffs, blockage, unowned questions), seams of truth (the gap between what we believe and what is real — the false green lives here), and seams of time (the gap between what happens and what the institution learns). Its learning is a genome: the ratified core of rules and judgment that compounds as models improve and can grow new bodies — and no agent ever edits its own germline.
The humane boundary: people are not components of this agent — they are its principals. Humans supply the ends; the agent-company supplies means. Robert is the fitness function. Nothing grades itself.
Measured against main as of yesterday. The surprise: the organs mostly exist. What is missing is three connections (brain→seats, eve↔warp, Neo↔team) and one repair (a broken court). And one small organ — the call-log worker — already runs the complete loop in production: it senses calls and sales email, judges privacy per message, writes Salesforce on its own, leaves Slack receipts, gets audited daily by a non-author, and alarms if its auditor goes quiet. The elephant already has one living cell.
The pattern already exists three times — the daily audit (a scheduled Claude Code session), warp's seat economy, and the dispatcher that wakes Neo's private task. Arc 1 promotes it to standing law, and retires DailyHQ properly: threads are bodies; ledgers are state. A chat thread may hold a mind's working context, never the company's durable obligations.
Judgment-queue contract (Worker side): typed task in {taskClass, evidence refs, context refs, deadline, authority class} → typed verdict out {decision, basis, receipts, confidence}; lease + timeout + fail-closed-to-silence. The Neo request ledger already implements the lifecycle — generalize, don't reinvent.
Dispatcher, versioned: Phase 1 of the 08-21 plan unchanged — dispatcher + launchd definition move into the repo; the reused Telegram secret is replaced with a purpose-specific Neo control credential; install/start/stop/rollback are deterministic; a dead-man's switch (call-log pattern) alarms on heartbeat loss.
Neo Ops (the Claude mind): a persistent/scheduled headless Claude Code session woken by the dispatcher exactly as the Codex task is today, same silence-biased continuation contract, state read from ledgers never scrollback.
DailyHQ migration: (1) extract — standing instructions → versioned runbooks, open obligations → ask ledger, operating knowledge → repo docs/genome; (2) stand up Neo Ops; (3) one parallel week, Codex thread read-only, decisions diffed; (4) retire — thread archived, never deleted; the 08-21 plan's “Daily HQ” references re-point to ask ledger + Neo Ops.
Done when: a queued judgment task is answered by a seat-backed session with zero metered tokens on the judgment path; dispatcher survives restart with durable receipts; no external write while health is unsafe; DailyHQ retired with nothing operational referencing it. Rollback: re-point the dispatcher at the dormant Codex task (kept 30 days); the queue is engine-agnostic by construction.
2a and 2b ship together (Robert's red pen, 2026-08-23, reversing Decision 2): “@Neo check” and ambient reach their end state as one build, not two stages. What's gated is not whether ambient exists but how readily Neo speaks — a posture dial Robert owns (observe-only → receipts-only → helpful), a per-channel daily budget, and a delay-and-recheck window. Identity proof (bot-only posting) stays a hard precondition — a safety fact, not a stage.
2a — the verifier, “@Neo check” (ships with 2b, not before it): trigger = existing mention path (74 baseline assertions); investigator = seat-side (Arc 1's first production proof), reads receipt sources, answers with the receipt + link + tier-honest hedging (T2 reported soft, T0 as fact); reply = existing idempotent door + firewall, unchanged.
Receipt-source registry (new, versioned): each source Neo may check — the identity it reads as, freshness semantics, trust-tier mapping. Doubles as Arc 3's dispatch toolbelt. Done when: every reply carries a source a non-author could recompute, and zero false receipts occur (a wrong receipt is zero-tolerance). The v1 draft's counts (10 checks / 2 weeks / 2 teammates) survive only as measurements reviewed, never as gates on whether the path exists (red pen 2026-08-23).
2b — ambient, built to its end state: the 08-21 plan's safety plane stands (observation ledger split, write-time gates, fail-closed silence, tool-less detector, event dedupe/revision semantics, kill-switch layers). Amendments: (1) taxonomy = the three receipt-backed classes; approval-boundary fires on agent-directed asks only. (2) budget + delay + the posture dial: the dial is the sensitivity gate Robert asked for — observe-only records what Neo would have said, receipts-only speaks only when a receipt contradicts or completes a claim in the thread, helpful speaks whenever the mind judges it would materially help; the dial changes the question the mind is asked, never a threshold applied to its answer — plus budget ≤1 reply/channel/day in code and a reply window T+10min–T+2h with silent cancel when a human answers first. (3) identity first: live readback proving bot-only posting + a negative test that turns red if a user token can reach the ambient write path — before shadow (the on-disk manifest still describes an app sending as Robert). (4) social contract: team told before shadow in any real channel; tone = receipts never people, offer-to-recheck stance, concede when corrected; mute honored forever + mute-pattern. (5) validation right-sized: golden-24+ replay incl. wrong-correction / being-corrected / quote-loop cases with negative controls; adversarial suite aimed at the gate layer (lease races, dedupe-window replay, mode-flip mid-flight); replacing the 7-day shadow, a real-history rehearsal before shipping — the real detector path runs over the last 30 days of the authorized channels with every write recorded instead of sent, producing the prevalence number and the disagreement list (the list is the product); live, the dial starts at receipts-only, budget 1/channel/day, and real interventions are rated by the humans in the thread as review triggers, never as statistics that decide whether the path exists. (6) global daily spend/reply circuit breaker auto-degrades to observe-only. (7) detector output is untrusted input to the investigator — a promotion never escalates privilege by itself.
Prevalence hypothesis (falsifiable, measured by the real-history rehearsal): authorized channels produce ≥3 genuine receipt-backed opportunities/week. Measured lower → the daily stuck-thread digest is built as a second renderer over the same observation ledger, decided by the number, in the same build — not discovered in month four. Rollback: layered per the 08-21 plan — replies off → observation off → kill switch → revision revert; every write re-checks its gate at write time.
This closes the eve↔warp seam. A world-model finding becomes a typed work candidate; warp builds and courts it; verdict and deploy receipt flow back as ground-truth facts. The trick that keeps it safe: Neo's leverage grows while Neo's authority doesn't — courts carry verification (reviewer never the author), the ruling library carries routine judgment, and everything novel batches to you.
Work-candidate schema: {finding, evidence refs, proposed work class, blast radius (Plato change brief), authority class, rollback plan} — produced by Neo Ops seat-side, stored in the request ledger.
Authority: routine classes (test-fix, known-failure-class patch, doc sync) dispatch under ratified rulings cited by name; novel/destructive/scope-changing batches to Robert; the Neo control plane's forbidden list binds unchanged. Custody: warp's existing doors (sealed room / sandbox lane); deploys only under existing deploy rules with proven rollback; court verdict + provider receipt land as world-model entries with ground provenance.
v1 scope & done: read-only investigations + builds landing on courted branches; the first ten loops end at “PR ready + court verdict, Robert taps merge.” Done when one real Slack-sensed failure flows detection → candidate → build → court → tapped merge → deploy → receipt back in the world model → Neo closes the thread. Ten clean loops graduate the first routine class into the ruling library.
The honesty watchdog has never passed — four straight scheduled runs died on a freshness gate an unattended job structurally cannot satisfy. You said you have no standing preference, so this arc proposes three fixes and asks for your pick (Q1). Separately: a finished quote-entailment verifier gets wired in or deleted, genome deposits get real receipts instead of placeholders, and the two Salesforce-write postures get named so nobody “unifies” them by accident.
Watchdog Option A — self-seeding runs: the weekly job generates its trap seed at run start (freshness measured from run start, not a human gesture); an independent seeder leg keeps the trap sealed from the subject. Unattended-satisfiable by construction; keeps synthetic trap coverage. Option B — attended monthly: simplest; cadence drops; reintroduces the depend-on-a-human failure the system exists to remove. Option C — production-derived audit (recommended): retire synthetic traps; weekly, a non-author session samples N real claims Neo made and recomputes each against ground truth. Audits reality instead of a rehearsal; no seed gate exists to fail. C+A is a valid combined pick.
Repairs 2–5: wire the quote-entailment gate into the court-brief path or tombstone it · genome deposits carry real receipt refs, and a deposit without one is queryably labeled a claim · nothing is represented as ground without a receipt (honest labeling at insert time — the world model represents, never governs) · document draft-queue-human-drain vs. call-log-autonomous as two named trust postures with their justifications.
Done when: the picked watchdog passes 4 consecutive scheduled runs; the entailment gate has a caller or a tombstone; a sampled week of genome deposits shows zero unlabeled claims; audit cadence is visible on the One-Surface.
One versioned home for the heritable core: constitution, working manual, ruling library, skills, taxonomies, tone contracts, capability maps. Ratification is the only merge gate, amendments get an adversarial read before landing, and deletion is a first-class gene edit — the genome must stay compressible. Southtown was the accidental prototype of growing a second body; this makes it a quarterly drill.
Home (Q6): a dedicated repo fronted by the existing genome ledger service is the candidate; courts run on changes. Cold-start drill (quarterly, scripted): fresh machine, genome only — measure what re-grows across named categories (harness up, bay materialized, courts run, world model cycles, a review surface ships); score is a percentage with named gaps.
Selection input beyond Robert: teammate feedback and client outcomes enter the germline as labeled evidence — guarding against overfit to one person's taste — while ratification authority stays with Robert. Done when: cold-start test #1 has a scored receipt; every standing rule can name its home; a sampled ratification shows the court-read happening.
Re-cut 2026-08-23 (“drive us to the end state”): the build is one shot; only the world's clock is phased.
observe-only · Codex dispatcher stopped, Claude dispatcher installed (old plist kept 30 days) · social contract message posted by Robert. Gate: live readbacks — bot identity, seat-backed judgment path, zero API tokens on the judgment path.receipts-only after the first clean observe-only week · DailyHQ parallel week then retired · watchdog accumulates 4 passes · first Arc 3 loops end at Robert-tapped merges. Gate: measurements reviewed as review triggers; first routine class ratified when ten loops are clean.Arcs 4 and 5 run as continuous threads; their first receipts land in the build phase — the audit's first run, cold-start test #1 in a disposable container.
Status 2026-08-23: none answered directly. The implementation plan carries each as a stated assumption Robert can reverse — Q1→Option C, Q2→extracted from disk (no paste needed), Q3→his Max seat with a concurrency budget, Q4→#bot-test first, cost ceiling = replies/day since the path is seat-backed, Q5→the rehearsal's number decides in-build, Q6→manifest first, home chosen from the inventory. See the plan's “Assumptions standing in for rulings.”
Option A (self-seeding traps), B (attended monthly), or C (production-derived audit — recommended; C+A is a valid combo)? This is the one broken court in the body and the ruling is currently orphaned.
Export or paste the thread's standing instructions and state, or grant a read-through session. What in that thread must survive verbatim?
Run Neo Ops on your Max seat, or stand up a dedicated seat? A dedicated seat cleanly bounds concurrency and separates identity; it also costs money.
Which low-sensitivity channel follows #bot-test in the ambient canary, and what monthly model-cost ceiling applies to the ambient path? (The 08-21 plan's defaults otherwise stand: public/internal only, 30-day metadata retention, Robert-only mode changes.)
If shadow measures fewer than ~3 genuine opportunities a week, do you pre-approve the pivot to a daily stuck-thread digest, or does that come back for a fresh ruling?
Dedicated repo, or extend the existing genome ledger service? And any objection to teammate/client evidence entering the germline as labeled selection input?
Canonical source: docs/ai/design/agentic-os-five-arcs-2026-08-22.md · amends the 2026-08-21 ambient plan (safety plane retained; sequencing, taxonomy, validation, and substrate changed) · inventory measured against main 3ecefc86f, 2026-08-21.
Two notes, verbatim, and what each changed.
On “Your face — Telegram, the One-Surface, the morning brief”:
“This can just be a Claude code session. I prefer to use my sub instead of api anyways.”
→ changed: Decision 5 amended; Arc 1 gains “the face on seats” as a deliverable.
On Decision 2, “Wedge before ambient”:
“Drive us to the end state, gates should be ambient reply sensitivity, not whether it happens at all.”
→ changed: Decision 2 reversed; Arc 2 re-cut (2a and 2b ship together; posture dial + budget + delay are the gates; the 7-day shadow replaced by a real-history rehearsal; counts become review triggers); the sequencing table collapsed to Build → Cutover → World; the trust-risk mitigation re-cut.