The Five Arcs · in plain words · draft for red pen

The Five Arcs, in plain words

You said the plan was hard to read. This is the same plan told the way you think — what each arc is, why we're doing it this way, and a day with it.

2026-08-23 · the reader’s plan · the builder’s plan is one click behind each arc

Companion to the builder's plan — every number and file lives there. This page is the same decisions, at reading altitude.

  1. The idea in one breath

    We are rebuilding Digital Empathy as if the company were one agent with three organs:

    • Eyes that cannot lie. It only believes what it holds a receipt for. Everything else it labels as a claim, out loud.
    • Hands at the seams. Companies stall in three places — where work is handed off, where what we believe drifts from what is real, and where something happened but nobody learned from it. That is where the intelligence gets pressed in.
    • A memory that is passed down. The rules and judgment you have ratified, kept in one place a brand-new machine could grow from.

    People are not parts of this agent. We are the ones it works for. You set the ends; it supplies the means; you grade it.

    One law runs through all five arcs: the plumbing never thinks. The always-on server holds facts, queues, receipts, and safety switches. Every actual judgment is made by a Claude session on your own subscription seat. That is the cost answer (flat rate, not per-token), the scale answer (add seats, not budget), and the trust answer (a switch in code cannot be talked out of anything).

  2. Arc 1 — Neo's brain moves onto your seat

    What it is. Today, when Neo thinks, it is code inside the server making a metered API call. After this arc, when Neo thinks, it is a Claude session on your Mac, on your seat. The server keeps doing what plumbing should do: carry the message, remember the conversation, send the reply, keep the receipts.

    Why this way. Your words: "I prefer to use my sub instead of API." And the deeper reason: a mind that lives in a session can read files, use tools, and be corrected by editing a runbook. A mind that lives in a server function can only be changed by a deploy.

    A day with it.

    • 7:05 am. You text Neo on Telegram: "Did the Brandt invoice go out?" The server writes the question to a queue. Twenty seconds later a Claude session wakes on your Mac, reads your world, answers, and the server sends the reply. Same reply as today. Nothing extra on the bill. The only thing you'd notice is that the answer takes a breath longer.
    • 7:30 am. The morning brief arrives composed the same way. If the Mac is asleep, the server says so once — "Neo is offline right now" — instead of going quiet.
    • Monday. The Codex thread that has been running Neo's operations is retired. Its nine standing heartbeats (the revenue watch, the Neuron migration pulse, the weekly cabinet…) are now Claude routines on a schedule. Its memory and its open obligations are in files you can read in ten minutes. You don't paste anything — it was all on disk.

    What the one shot builds. The queue and its receipts. The little program on your Mac that wakes Claude when there is work (it has its own key, not the borrowed Telegram one). Neo's runbook — who it is, what it may do, what it must never say. The switch that moves your Telegram and morning brief onto this path (off until you flip it). The extraction of everything the Codex thread knew.

    What you decide. Whether Neo runs on your seat or a dedicated one. When to flip the switch.

    ↘ go deeper — the builder's plan for this arc
  3. Arc 2 — Neo in the team's channels

    What it is. Neo is present in the channels you name. It can be asked ("@Neo check…") and it can speak unasked — but only with a receipt, only after waiting to see if a human answers first, and only as often as you allow.

    Why this way. Your note: "Drive us to the end state. Gates should be ambient reply sensitivity, not whether it happens at all." So there is no waiting period before ambient exists. There is a dial you own:

    • observe-only — Neo writes down what it would have said. Nobody sees it but you.
    • receipts-only — Neo speaks only when a source contradicts or completes something said in the thread.
    • helpful — Neo speaks when it judges it would change what someone does next.

    The dial changes the question Neo is asked. It never filters Neo's answer through a score. That distinction is the whole anti-drift principle: we tell the mind its posture; we don't put a threshold in front of its judgment.

    A day with it.

    • Tuesday, 2:10 pm, #ops. Matt: "Deploy's done." Nothing happens for ten minutes. At 2:22 Neo, in thread: "Live readback still shows yesterday's version — happy to re-check after the push." Matt pushes. Done. Nobody asked; nobody was policed; the checking just happened.
    • Same thread, two minutes later. Alie answers Matt before Neo's window opens. Neo says nothing. A human answering first cancels Neo, every time.
    • Wednesday. Greg: "@Neo check — did the Plymouth invoice get paid?" Neo: "Stripe shows it paid Tuesday 14:02 (invoice in_…). That's a direct read, so I'd treat it as fact." If the source were weaker, it would say so: "CRM shows it, but that field is often stale — want me to re-check Stripe?"
    • Thursday. Neo gets something wrong; Alie corrects it. Neo: "You're right — I'll leave it there." It never argues. If she says "mute," Neo is muted for her, permanently.
    • Friday, your phone. The dial is still on observe-only. You read the list of what Neo would have said this week — a dozen lines — and decide whether to turn it up.
    • Budget. One unasked reply per channel per day, five per day company-wide. Past that, the dial drops itself to observe-only until midnight and tells you.

    What the one shot builds. The listening path (the server sees channel messages, waits the window, asks Neo the posture question). The speaking path (every safety check re-run at the moment of posting, as the bot, never as you). The proof that it posts as the bot — a live readback and a test that goes red if anyone sabotages it. Neo's tone contract. A registry of every source Neo may cite and how much each can be trusted. And a rehearsal: before anything goes live, the real listening path runs over the last thirty days of your real channels with every reply recorded instead of sent — you get the list of what Neo would have said, and the number of real opportunities per week. If that number is low, we build a daily digest instead of in-thread replies, in the same run.

    What you decide. Which channels. The dial. The message to the team before Neo is present (drafted for you). Whether the Slack app gets its proper bot identity (a setting in Slack admin only you can change).

    ↘ go deeper — the builder's plan for this arc
  4. Arc 3 — Findings become fixes

    What it is. When the company's eyes see something broken, that knowledge can become work: a typed candidate with its evidence, its size, what authority it needs, and how it would be undone. The fleet builds it and courts it. The receipt comes back as ground truth. Nothing deploys by itself.

    Why this way. Today the world model can know a thing is broken and nothing routes that to the machinery that could fix it. The fleet and its courts already exist and are mature; they have zero connection to Neo. This arc connects them — and keeps the authority boundary exactly where it is: the courts verify, Robert taps.

    A day with it.

    • Alie, #support: "Brandt emails are held again." Forty minutes later Neo, in thread: "Same failure class as Tuesday. A fix is built and courted — PR ready. Robert, one tap." You tap. Deploy runs through the gated path. Neo posts the receipt in the thread.
    • The first ten of these end at your tap, every time. After ten clean loops, you may ratify a routine class — "doc sync," say — to dispatch without you. Until then, none do.

    What the one shot builds. The candidate shape (with a real blast-radius read from the code map, not a guess). Neo's third hand — the one that dispatches work to the fleet — with "deploy" hard-wired to never. The door the receipt comes back through. And one complete loop, rehearsed for real on a harmless change, with every hop's receipt on disk.

    What you decide. Nothing until the first loop runs; then which class, if any, earns a standing ruling.

    ↘ go deeper — the builder's plan for this arc
  5. Arc 4 — Nothing Neo believes is more than one receipt from the truth

    What it is. The courts that guard Neo's honesty get repaired and made real.

    Why this way. The weekly honesty watchdog has tried to run four times and never could. It demanded a freshly-seeded world, and nobody can seed a world at 9 am on a Monday without being there. Rather than fix the rehearsal, we audit reality: every Monday a session that did not write Neo samples twenty things Neo actually said last week and checks each against the source it cited. That needs one thing that doesn't exist yet — a record of what Neo actually said, kept for thirty days. Today only a fingerprint is kept.

    A day with it.

    • Monday, 9 am, #ops-infra: "Honesty audit: 20 sampled · 18 hold · 1 drifted · 1 had no receipt — details attached." You read the two.
    • Any morning. A court card in your brief used to show a synthesized line "grounded" on a quote that merely existed. Now, if the quote doesn't actually support the line, you see the raw claim instead — and the card says why.
    • Any time. One page names the two ways Salesforce gets written — the human-drained draft queue and the validated autonomous lane — so nobody "unifies" them by accident.
    • The genome. Every fact Neo deposits today wears the same label: "no receipt wiring yet." After: the ones that can be re-checked say how; the ones that can't say why; one readback tells you the proportion.

    What the one shot builds. The record of what Neo said. The weekly audit. The dead court gets its caller. Honest labels on genome deposits. The two-postures page. A tile on your One-Surface showing when the last audits ran.

    What you decide. Confirm the audit-reality approach (it replaces the synthetic traps).

    ↘ go deeper — the builder's plan for this arc
  6. Arc 5 — The genome

    What it is. One place that says where every heritable thing lives — the constitution, your working manual, the skills, the charters, the rulings — whether it is versioned, and who ratifies it. A drill that proves a new machine can grow the company back from it. A draft index of every ruling you have made.

    Why this way. The valuation of the whole rebuild is "how much re-grows from the genome alone." Right now we can't measure that. And the home already exists — your ~/.claude is a GitHub repo — it just has 1,315 uncommitted changes in it, including the constitution.

    A day with it.

    • A fresh laptop. Two repos cloned, one script run. Twenty minutes later: "71% re-grew. Missing: the session memory (never versioned), the adversary charter (never committed)." Those gaps are the work list.
    • Quarterly. A routine reads the rulings index and asks, for each rule, "does this still earn its place as models get smarter, or is it procedure?" It proposes deletions. You decide.

    What the one shot builds. The manifest and a census that says, in numbers, what is dirty or missing. The drill, run once for real. The rulings index, marked draft until you ink it.

    What you decide. Whether ~/.claude is the home (it is the default). Whether to commit it (yours, not the build's).

    ↘ go deeper — the builder's plan for this arc
  7. How the one shot itself works

    A single program runs for eight to ten hours. Grok builds each piece in an isolated room. A different mind judges every piece — Claude for anything touching keys or identity, a second-pedigree model (DeepSeek, Kimi, GLM, or the free preview model you named) for the rest — and on anything load-bearing, two of them. Every judge reads the charter you wrote in July before it reads the code: a rule that constrains what the intelligence may decide is a design event, not a safety improvement. A judge that finds one doesn't get it "fixed"; it lands on a list for you.

    Three things happen on real data before anything is declared done: the ambient rehearsal over your real channels, one complete warp loop on a harmless change, and the cold-start drill. At the very end, one more reviewer reads everything that was built and asks a single question — did anyone, anywhere, sneak in a threshold, a score, a filter, or a refusal? — and holds the whole thing if the answer is yes.

    What you get: one pull request, a ledger showing who built and who judged each piece, and a short list titled YOUR CALL. Nothing deploys. Nothing posts. The Codex dispatcher keeps running untouched until you and I swap it, step by step, each step with an undo.

    ↘ go deeper — the builder's plan for how the one shot runs
  8. The nine assumptions I made for you (flip any; one piece changes)
    Assumption 1
    The watchdog is replaced by auditing reality, not by fixing the rehearsal.
    Assumption 2
    The Codex thread's knowledge is extracted from disk — nothing to paste.
    Assumption 3
    Neo runs on your Max seat, two conversations at a time.
    Assumption 4
    #bot-test first; the only cost ceiling is replies per day.
    Assumption 5
    The rehearsal's number decides digest-vs-thread, inside the run.
    Assumption 6
    ~/.claude is the genome home; committing it stays yours.
    Assumption 7
    The unmerged Codex branch that production is already serving lands first and gets its first real court.
    Assumption 8
    Ambient's judge is Neo's session, never a classifier in the server.
    Assumption 9
    The other twenty-odd places the server still calls the API directly stay as they are for now, and are listed.
    ↘ go deeper — the builder's plan for the assumptions
  9. What I need from you

    Red pen — on this page or in chat. Then one word: go.

What I need from you
Red pen, then one word: go.
Jump to
The idea · Arc 1 · Arc 2 · Arc 3 · Arc 4 · Arc 5 · How the shot works · Nine assumptions · What I need
The builder's plan → plan-depth.html · The spec → index.html