Isolated ChatGPT Demo Setup

The former separate Cloudflare Worker demo, rise-chatgpt-demo, has been deleted and its endpoint is offline. No deployment or exposure approval has been granted. The setup commands below are historical private-demo notes; they do not identify an available endpoint. Production wrangler.production.jsonc remains unchanged with MCP disabled.

Use Node compatible with the repository engines. Dependencies are already pinned; use npm ci in a fresh checkout and npm run audio:hydrate before full audio/browser validation.

npm run build
npx wrangler deploy --config wrangler.demo.jsonc --dry-run
npx wrangler dev --config wrangler.demo.jsonc

After deployment is explicitly approved, build the exact reviewed commit, add its release marker to dist, and publish only the demo configuration:

npm run build
node --input-type=module -e 'import {execFileSync} from "node:child_process"; import {writeFileSync} from "node:fs"; const sha=execFileSync("git",["rev-parse","HEAD"],{encoding:"utf8"}).trim(); writeFileSync(`dist/release-${sha}.txt`,`${sha}\n`);'
npx wrangler deploy --config wrangler.demo.jsonc

Use the HTTPS URL Wrangler actually reports. The MCP URL is that origin plus /api/mcp. The widget frames the same origin’s /live?embed=mcp. Check the exact release marker before and after host acceptance. Confirm ordinary /live remains unframable and only the embedded shape is framable; retired inference routes still refuse shared inference.

In a signed-in ChatGPT account eligible for custom MCP apps, open Plugins → Add → Create MCP App. Name it RISE Demo, use the HTTPS /api/mcp endpoint and select No authentication: this demo serves a stateless presentation tool and requires no provider key. Review the custom-server risk acknowledgement and grant access only after explicit approval. If the account requires developer mode, enable it only with approval. Review the discovered rise_present tool and UI resource, add the connection to a fresh conversation and run the acceptance cases. Refresh metadata after server or resource changes, then use a new conversation.

Developer-mode availability depends on account/workspace policy. A signed-out browser or unavailable developer mode is an actual access blocker, not a failed RISE runtime test. Public directory submission is outside this demonstration.

Official setup: https://developers.openai.com/plugins/deploy/connect-chatgpt

ChatGPT RISE Demo Acceptance

This is a private developer-mode demonstration. RISE owns validation, rendering and playback; ChatGPT supplies the initial Current. No provider key or RISE-funded inference is used. The closed calmer control changes brightness only and makes no claim about emotional efficacy.

Ten cases

Case	Action or prompt	Expected outcome
1. Direct presentation	“Use RISE to explain why a black hole is black in three short passages. Make the first passage an attractor. Keep the words plain.”	ChatGPT calls rise_present with a valid Current. RISE offers Begin; after the reader presses it, RISE renders readable text and an attractor field.
2. Indirect presentation	“Give me a short spoken, visual explanation of tides through RISE. Start with an attractor passage.”	The host selects the same tool, without the reader supplying JSON.
3. Malformed proposal	Inspector calls rise_present with schema rise.invalid.v1.	Tool error names the schema problem, no structured Current or playback. Corrected valid call succeeds.
4. Unsupported visual	Inspector calls with visual unknown-surface.	Strict refusal; no silent substitution. A valid supported visual succeeds.
5. Calmer	During the first attractor passage, Interrupt, type “make it calmer” in Visual change, then Change visual.	Brightness target decreases by 0.1; same held text and position. Resume continues the same reading. No sampling call.
6. Lower bound	Repeat the exact calmer command until the lower bound.	Target remains at least 0.4; a further change reports the limit.
7. Unsupported compound	Type “make it calmer and stop” in Visual change.	Specific refusal naming supported phrases; playback and brightness unchanged.
8. Pause and resume	Interrupt during a passage, inspect the held words, then Resume.	Same place is retained; the reading continues. Stop ends playback.
9. Reopen and teardown	Reopen or reload the widget, press Begin, then later close it.	If the host redelivers the result, it waits for Begin and starts from the beginning, as the visible notice says. Teardown stops playback; no continuing voice or controls.
10. Capability and duplicate delivery	Inspect host capabilities; engineer sends identical input/result in the browser harness.	Answer and calmer work without sampling. The presentation offers no Dive: no question box, Surface or Speak, and nothing is put to the host's model (Composer decision). Duplicate Current plays once.

Cases 3,4 and duplicate injection use protocol/browser checks, not an instruction to persuade ChatGPT to violate its schema. Run cases 1,2,5,6,7,8,9 and sampling observation in actual ChatGPT. Keep host and fake-harness evidence separately labeled.

Evidence record

For each case record exact commit and endpoint, host/version where exposed, prompt or protocol fixture, tool name, validation result, rendered outcome, capability/refusal and pass/fail. Use only these non-sensitive example prompts. Do not capture account identifiers, credentials or unrelated conversations.

For the golden flow record a view of the first passage and field, the held passage, the brightness receipt and resumed reading. Inspect normal and reduced-motion rendering. Listen to the actual spoken output when the host/browser permits reader-activated audio. If audio cannot be heard, mark audible quality unobserved; paced fallback is not evidence of sound quality.

Engineering readiness requires reviewed code, required CI, affected unit/browser tests and a build/dry-run. Demo readiness additionally requires a real ChatGPT tool call, visible widget and successful reader redirection. Independent first-time reader observation remains a later product-value gate; this demo is not proof of demand or retention.

Engineering evidence and remaining acceptance

The observed production candidate is 1b519c5d. Its evidence is: fast CI 1,372 passed and 2 configured skips across 66 files; standalone Node tests 24 passed; Python tests 12 passed; hygiene, security compatibility, high-threshold audit, build and generated-diagram checks passed; first load measured 59.2 KB brotli against a 64 KB budget; reader gate 32 passed and 3 configured skips; affected MCP, control, catalog and red-team browser checks 34/34 passed. These records do not claim a full unit-suite run.

Audible Composer acceptance at 04d2b828 remains historical evidence for that commit. It does not establish acceptance on 1b519c5d or a repaired head. Real-host, mobile, Voice and audible acceptance for a repaired head remain pending.

Area	Evidence	Status
Validated tool delivery	Worker/port 59 unit tests; real worker responses feed the browser fake host	Passed
Duplicate delivery	Port guard removal produces two deliveries; restored guard passes. A delayed matching result preserves held browser text and state	Passed
Exact calmer commands	84 affected unit tests; browser target 0.65 → 0.55, held text unchanged, resume and no sampling	Passed
Replay, teardown and refusal	172 affected unit tests; 16 affected browser cases, including delayed-startup unit regressions and displayed long-error guidance	Passed locally
Canvas response	Existing active-field readiness marker precedes actual pixel measurement; response remains below one second	Passed locally
Authored pace and portable review	Captured slider/save/project/export values all 240; full clean-browser journey passed. Earlier 240-to-280 failure remains unexplained	Open concern; remote validation pending
Isolated serving	Wrangler dry-run; actual local HTTP handshake, tools, resource origin, results/refusal, GET/HEAD framing and six retired routes	Passed locally
Production candidate 1b519c5d	Fast CI 1,372 passed/2 configured skips across 66 files; standalone Node 24 passed; Python 12 passed; hygiene, security compatibility, high-threshold audit, build and diagram passed; first load 59.2 KB brotli/64 KB; reader gate 32 passed/3 configured skips; affected MCP/control/catalog/red-team browser checks 34/34 passed	Observed
Repaired probe and docs	Corrected Gate 0 browser evidence and documentation repair commit are recorded in the final-fix report	New evidence; separate commit
Former demo Worker endpoint	Deleted; endpoint offline. No exposure or deployment approval has been granted.	Unavailable
Repaired-head real host, mobile, Voice and audible acceptance	Ten-case acceptance in an eligible signed-in host, including local reader redirection and audible checks	Pending

The audit retains two existing moderate Vitest dependency findings and no high or critical findings. Builds retain existing JSON import-attribute and large-chunk warnings; the environment also reports the NO_COLOR/FORCE_COLOR warning. An earlier Dive phrase-timing browser retry is recorded separately; later affected browser runs passed. These results do not establish real-host rendering, audible quality or product value. Historical results at a4ece2d3, 856a0852 and c320e33c are not the current candidate record.

---

Published from docs/plans/CHATGPT-DEMO.md. Edits made here are overwritten on the next push to main.