Status: REWORK — the canned version was rejected as dishonest (it ran a hardcoded animation before the model even loaded; the AI did nothing). Owner call: make it fully agentic — the 0.5B CHOOSES each action, the harness validates + applies. "Give the AI a script (a goal), not a choreography."

Agentic design (the honest version)

Non-negotiable: load the model first, gate the button on it. No movement before the AI exists.

The AI drives via a tight TOOL protocol — one call per turn, valid targets enumerated each turn so it chooses from a list (never invents a route). The harness validates every call before applying it (law #2: the AI proposes, code guards). MPA-safe: the agent's state (goal + what it's done) rides sessionStorage across each navigation and the loop re-hydrates on arrival.

Tools (one per turn, one line):

• GO <route> — navigate. Valid routes enumerated from the live catalog; rejected otherwise.
• HIGHLIGHT <id> — spotlight a section on THIS page. Valid ids = this page's rendered headings.
• NOTE <text> — append an AI-authored line to the notepad (only where a notepad exists).
• DRAFT <text> — prefill the contact message, AI-authored (only on /mail; never sends).
• SAY <text> — one short sentence to the visitor.
• DONE — the demo is complete.

Loop (per page, in the reasoner — it has the engine): build the turn prompt (goal + current page

• allowed routes/anchors + what's been done) → stream one line → parse → validate → on reject, feed the

reason back and let it retry (bounded); on accept, apply the real op + record it → GO/DONE end the page (stash state + navigate, or finish); other tools loop on-page. A per-page + total step cap and the "type anything to stop" cancel bound it.

Guardrails: routes/anchors validated against live data (never invented); NOTE/DRAFT are AI-authored text (the B1/note-write precedent — the model may compose, code targets + never submits); step caps; graceful stop.

Risk (owner accepts): the 0.5B is weak; this needs prompt iteration against the real model. Unit tests pin the loop/validation on a mocked engine; a live e2e drives the actual 0.5B to tune wording.

Tuning log (live 0.5B, via a WebGPU probe)

• v1 — parroted the user-turn question ("What is your next action?") back as SAY chatter, looped,

never advanced. Root cause: an echoable question turn + SAY escape hatch.

• v2 — dropped SAY, made the user turn a non-question cue, numbered the anchors, added a computed

next-step hint. Result: drove GO→HIGHLIGHT, but re-GO'd the current page and malformed the next hop.

• v3 — reject GO-to-current-page; the one-shot example = the actual suggested command; reject

placeholder text. Result: GO→HIGHLIGHT→authored its own NOTE, then fumbled the mechanical GO /mail.

• v4 (current) — two-pass turn: the model chooses freely first; if that's invalid, a hardened pass

forces the mechanical command. Result: full run completes — GO→HIGHLIGHT→authored NOTE→(forced GO /mail)→authored DRAFT→DONE. The model genuinely authors the note/message; the mechanical hops have a forced floor so a fumble can't wedge the demo.

Honest framing: the AI is loaded first and does the real generative work (authoring the takeaway + the message from the page it reads), emitting validated tool calls. The mechanical hops (which page, which section) are guided/floored because a 0.5B fumbles them — but they were always mechanical, and the harness only ever applies what validates (law #2). The valuable, visible AI contribution — the writing — is the model's own.

_Superseded below: the original canned Option-A storyboard, kept for reference only._

Status (superseded): SHIPPED — Option A (desk pilots the real site, homepage-staged), full sweep (read → note → highlight → save → /mail draft → choose). Built on the A2 tour relay.

What shipped

• src/ai/showcase.ts — the 3-step drive as plain data + cursor codec (mirrors tour.ts).
• actions.ts — showcase-start action; "watch me work" / "watch the AI act" / "run the demo" route

to it; "stop the demo/showcase" folds into tour-stop (shared cancel).

• desk-reasoner.ts — showcase-start handler (drives leg 0, stashes the cursor, navigates);

"type anything to stop" cancel via showcaseClear.

• desk-door.ts — runShowcaseLeg in the arrival chain (after runContactTask): performs each

stop's note.append + closing choicesOp, then advances. Reuses the existing ARRIVE_KEY (announce + anchor highlight) and CONTACT_TASK_KEY (compose prefill) relays.

• scripts/site.js — "Watch me work" is the lead chip on the hero "try this" rail.
• Tests: showcase.test.ts + reasoner/actions unit tests + e2e/desk-showcase.e2e.ts (full drive,

stop affordance, stale-cursor abandon). Full suite green; e2e green against a live browser.

Follow-up (not done): the "Watch the AI act" welcome walk card still links to /grain — could be rewired to trigger the homepage pilot instead of navigating away. Goal: make GRAIN's one thesis visible — one AI, one op vocabulary, every surface — in a single tap. Today a visitor only sees chat text + a link; the interface-driving magic never shows.

What already exists (reuse, don't rebuild)

• demo.run verb → grain stub reasoner drives scripted RenderOps over the door

(../grain/packages/grain/ai/reasoner.ts:219). Zero model, re-runnable, mutates no storage.

• A /grain scenario already runs 6 beats: read rail → type in Ask → reply in thread → complete a

task → draft a follow-up → hand back.

• Trigger points today: the buried "See what's new" button; the /grain chip "Watch the AI act" is

just Q&A (routes to grounded chat — it explains instead of doing).

The gap

1. The chip that says "Watch the AI act" doesn't run the act.
2. The existing scenario stops short of the full vocabulary — it never touches the notepad

(note.append), a form field (field.set), or an interactive choice (choicesOp). Those three are the most "an AI is operating my UI" moments.

Where it stages — the homepage is the stage, not /grain

First-time visitors land on /, so the demo earns its keep there. But / is a launcher (welcome cards + the desk + the "try this" rail) — it has no grain sandbox surfaces (grain-rail, grain-ask, grain-task, notepad). Those live only on /grain. So "on the homepage" means one of:

• Option A — homepage pilots the REAL site (recommended). The desk drives the actual portfolio:

spotlight a walk card → travel to the flagship note → highlight a passage → hop to /mail → prefill a message to TJ → offer a choice. Strongest first impression: the AI operates production surfaces the visitor is really looking at, not a toy board. Reuses the A2 tour's arrival-stash relay (each leg carries the next scripted step across the page load) — NOT the /grain demo.run sandbox. Cost: multi-leg, more moving parts.

• Option B — homepage triggers, /grain stages. The "Watch the AI act" card already links to

/grain; make it auto-start demo.run on arrival. One tap on / → lands on /grain and plays the sandbox run below. Cheapest, reuses everything — but the show happens on /grain, not /.

Option A beats — the desk pilots the real site (homepage-staged)

Closing reply: "That was one AI on one op vocabulary — it opened a note, highlighted a passage, saved a takeaway, and drafted a message to TJ. Every step drove the real site, through the same door you use."

Implementation note: Option A extends the tour relay (tourSet/arrive), teaching the arrival stash to carry an action to perform on arrival (spotlight-anchor / prefill), not just an announce. That machinery already exists for A2 (it carries announce + anchor today).

Alternative — /grain sandbox run (Option B's stage)

Extend the existing 6-beat scenario to 9 beats so it exercises every op type. Backdrop stays up across the run; the spotlight follows what it touches; the terminal narrates every beat so it never reads as stuck. Timing per the existing HOLD_MS/SETTLE_MS cadence.

Closing reply: "That was one AI, one op vocabulary — it read, replied, completed a task, drafted one, saved a note, prefilled a message, and asked what's next. Every step went through the same door you use."

Honesty guardrails (kept from the existing design)

• Never submits anything (beat 7 fills, Send stays the visitor's — no submit verb exists).
• Graceful stop: any message mid-run hands back cleanly (stopped() check between beats).
• Mutates no storage: notepad/field writes are demo surfaces, re-runnable.

Wiring change

• Homepage entry point (Option A): the "Watch the AI act" walk card on / (currently a plain

link to /grain) becomes the trigger — tap it and the desk starts piloting the real site in place. Also expose it on the "try this" rail as "▷ Watch me work".

• The /grain chip "Watch the AI act" → route to demo.run (make the chip act, not answer) —

keeps the sandbox run reachable on its own page regardless of which stage we pick.

• Keep "See what's new" as-is (it's the /notes variant).

Open questions for you

1. Stage (the one that matters): Option A — homepage pilots the real site (recommended, multi-leg

relay) — or Option B — homepage tap launches the /grain sandbox run. A is the stronger landing moment; B is a fraction of the work.

1. Length: Option A is ~8 beats across 2 hops (≈20s). Trim (skip the /mail leg, end on the note

• a choice) or keep the full read→note→message→choose sweep?

1. Interruptible: confirm any keystroke hands back cleanly mid-run (the stopped() guard) — same

as the tour's "type anything to stop".