← Sketches · the native app program

Theatre mode

c1442026-11-12

Three kinds of tour your agent can give; in theatre, a story of AI going wrong is played on synthetic data, between a switch turned on and off together, and every beat is a test of the dashboard.

Source: c144 initial-prompt.md (Mark, 2026-11-12, verbatim); plan.md §1–§9; scenarios.md (wave 1, S1–S6; wave 2; seven gaps); c125 mcp.ts (qpoint-guide v0.3.0, fifteen tools; actions stay with the person); c123 DRILLS; qnative c2 new-to-qpoint.json, this-machine.json, app::guide; qnative c3 worlds, conditions, streams; c133 trainings; c137 messages.

Ask your agent for a tour
▭New here

Show me around, at the level I pick: a slider from ELI5 to “I have networking degrees”. Your data where there is any. From start-here and new-to-qpoint.

▣AI, on your own machine

The model is the brain, the harness the body, tools the hands, MCP servers borrowed hands, skills the speciality, and where each metaphor breaks. Your data. From this-machine.

▤The doomsday scenarios

Phishing for financial records, the server that calls Pyongyang, the secrets stealer, the public link, card numbers to a stranger, the runaway loop. Synthetic data, always, in theatre.

any harness, through the qpoint-guide MCP server (web lab, then the native app's guide): “give me the phishing tour”
A theatre tour
Theatre · synthetic data · nothing here is happening on your machine · Phishing for financial records · Leave theatre
1Curtain up

The agent says what is about to happen and points at the switch; the person flips it on. The tour waits until it is on.

2Beat by beat

Each beat writes only what the sensor would see, on the story's clock. The dashboard works out findings, alerts and notices by its own rules. The agent says what changed and points at it.

3Replay at Enforce

Rewind with one policy at Enforce: the story stops at that policy's beat, the person gets a notice, a training is assigned. Changes stay in the theatre's world.

4Curtain down

The agent points at the switch again; the person flips it off. Back on your own data: what was real (nothing) and what to look at next. Any other exit ends theatre too.

Inside theatre: the draft, Publish, Connect and every outward action are disabled; every capture carries the band.

A scenario is a test
The scenario · data, beside the tours
▤Beats

at · say (at three registers) · sensor events · expect lines

▤Replay

which policy stops it, at which beat

▤Gaps

where no policy names what happened: a destination's country, a link anyone can open …

world at beat N = base world + events 1…N, folded by the real rules

What the dashboard shows · never written by the scenario
▭The operator's seat

Findings, alerts, the policy's page, the delivery log (web).

▣The person's seat

Status, Activity, the inbox's notice, the training (native).

each expect line checked by region, at Observe and at Enforce

The walk · the demo is the test
◇theatre-walk

lab: every scenario, every beat, both modes; a miss names the beat and the region.

⬚Stories

native: a story per beat, a region-text test per line.

●A harness run

Claude Code plays the tour from one prompt.

Two axes the person sets
The register: ELI5 ↔ networking degrees

Five stops; 2, “I'm new to this”, by default. Authored at 1, 3 and 5; the agent adapts the rest in chat. In the panel's header; the agent reads and sets it.

The marks: Cute ↔ Plain ↔ Swiss

Each part (harness, model, provider, MCP server, tool, skill, workspace) keeps one mark everywhere, taught by the primer. The person picks the style at its end: pixel pawns with faces, today's glyphs, or a geometric set on a grid. A lasting preference.

tours and the guidetheatre: synthetic data, labelledwhat the story writesEnforce stops itgaps: no policy yet

Qpoint Brand Style Guide