7.4 KiB
CORE Demo Capture Shot-List & Script (2026-06-14)
A ready-to-shoot plan for the first public-facing demo assets (flagship video, short clips, screenshots). Every route, on-screen line, and honesty card below is sourced from the live workbench — nothing here overclaims past what the engine actually proves.
Cardinal rule: never open on the Chat box. The moat is refuse / replay /
show-its-work, not "type → get text." The target reaction is the workbench's
own guiding question (docs/workbench/README.md):
"Wait… I can actually inspect and replay the cognition." — not "Cool chatbot."
0. Setup
core workbench api # read-only backend on http://127.0.0.1:8765
# in a second terminal, from workbench-ui/:
pnpm dev # React UI
Capture settings:
- 1920×1080, or 2× retina downscaled for crisper text.
- Dark theme. Hide OS chrome / notifications. Full-screen the browser.
- Slow the cursor. Pause 1–2s on every honesty card and every hash — they are the payload; let viewers read them.
- Keep the Evidence Chain Rail visible where possible (reinforces "URL = subject; everything is addressable evidence").
The verbatim thesis (from workbench/tour.py) — use as tagline / cold-open VO,
do not re-author:
"CORE is the deterministic engine the opaque transformer defines itself against. Bring a claim — from any model, including your own — and watch the engine decide it, refuse it, or replay it to the same hash. The proposing model's authority is ignored; only verified evidence promotes."
1. Flagship video — film the Tour (target 60–90s)
/tour is purpose-built for this: a 5-step narrative bound to the live demos,
where honesty cards are pulled from each demo spec and cannot claim more than
the demo proves. Film it in order.
| # | Route | Hold the camera on (real on-screen text) | Voiceover | Dur |
|---|---|---|---|---|
| 1 | /tour (intro) |
"These are not screenshots or animations. Each step runs a real demo over pinned fixtures, end to end." | Read the thesis line above. | ~12s |
| 2 | /demos/deductive_entailment_authority |
"served … only when the pinned ROBDD engine and an independent oracle agree" | "Two independent engines. They must agree, or it refuses." | ~12s |
| 3 | /demos/epistemic_truth_state |
"You are watching a wrong answer get refused rather than served" | "This is what wrong=0 looks like — a wrong proposer gets rejected." | ~12s |
| 4 | /demos/proof_carrying_promotion |
"authority lives in the evidence, not the model" | "Nothing enters memory without a verified certificate." | ~12s |
| 5 | /replay |
"Replay to the same hash… Determinism isn't a claim here; it's a deliverable." | Re-run the turn; let both hashes sit side by side. | ~15s |
Money shot: step 5 — two identical trace hashes after a re-run. Freeze-frame
- highlight box in post. This is the thumbnail and the pinned-tweet image; no transformer demo can show it.
2. Short clips (15–40s each) — one capability per clip, numbered series
Post as a thread/reel series. Each is a self-contained "they can't do this."
| # | Route | What to show | Caption |
|---|---|---|---|
| 1 | /demos/epistemic_truth_state |
Proposer smuggles an unsupported truth-state; engine rejects. | "We measure intelligence by what it refuses to say." |
| 2 | /replay |
Run, re-run, hashes match. | "Same input, same cognition, same hash. Every time." |
| 3 | /demos/{any} |
Slow zoom on the paired what_this_proves / what_this_does_not_prove cards. |
"Every claim ships with what it does NOT prove." |
| 4 | /evals |
The wrong=0 ledger; the wrong column at 0. | "0 wrong answers. Not 'low' — zero, by construction." |
| 5 | /calibration |
Gold-tether arena: Wilson floor vs θ ceiling. | "It has to earn the right to guess — with a one-sided confidence floor." |
| 6 | /trace → Field tab |
versor_condition < 1e-6, field_valid, field digest. |
"Cognition runs on a geometric field — proven coherent to 1e-6." |
| 7 | /lived-life |
Always-on heartbeat holding identity across reboot. | "Not a stateless chatbot. One continuous life." |
| 8 | /trace |
Scrub Pipeline / Grounding / Verdicts / Surfaces tabs. | "Every turn, fully inspectable: intent → graph → realizer → verdict." |
3. Screenshot set (stills for README, site hero, social cards)
High-res, dark theme, Evidence Chain Rail visible.
- Hero: the two matching trace hashes (
/replay). - Honesty-card pair (
/demos/{any}). /evalswrong=0 ledger (the 0 column)./calibrationWilson-floor-vs-θ arena./traceField tab withversor_condition.- LeftNav fully expanded — all 7 sections (Converse / Cognition / Determinism / Evidence / Discipline / Substrate / Settings) — "instrument, not chat box" at a glance.
- Command palette open (⌘K) over a route — operator-grade tooling signal.
4. Caption / VO bank (all from real on-screen text — nothing overclaims)
- "Bring a claim — from any model, including your own. The engine decides it, refuses it, or replays it to the same hash."
- "The proposing model's authority is ignored. Only verified evidence promotes."
- "These aren't screenshots or animations — each step runs a real demo over pinned fixtures, end to end."
- "Two independent engines have to agree, or it refuses to answer."
- "Determinism isn't a claim here. It's a deliverable."
- "Every claim ships with what it does not prove."
5. Docs walkthrough order (for the written launch post / deep dive)
docs/Whitepaper.md/docs/position_paper.md— the "why" (decoding, not generating). Lead here.docs/workbench/README.md— the "cognition observatory, NOT a chatbot shell" framing; the guiding-question quote is a strong pull-quote.docs/workbench/UI-UX-GUIDE.md— operator/evaluator route map + evidence grammar; the proof the UX is designed. Cite for the UI/UX-value argument.docs/runtime_contracts.md— surface/walk/articulation contract (determinism is contractual, not aspirational).CLAIMS.md+docs/claims_ledger.md— the receipts. Linking these publicly is itself a credibility move.
Technical-reader combo: Whitepaper (idea) → Tour video (demo) → CLAIMS.md (receipts).
6. Honesty guardrails — stay inside these (the whole pitch is honesty)
- Do not headline Chat answering a hard open-domain question. Verified
capability is formal/propositional decision, not broad NL reasoning — the
demos' own
what_this_does_not_provecards say so; let them. If Chat appears, it's "the conversational surface," not "look how smart it is." - GSM8K is a diagnostic, not a flagship. Never headline a math score. If math appears, it's "watch it refuse what it can't ground."
- Lived Life is real; the frontier is honest. Resume-as-same-life and idle-tick learning are built; the indefinite always-on process is the active edge. "One continuous life" (true), not "alive for months" (not yet).
- Lean on the honesty cards harder than on any capability. A system that volunteers its own limits on screen is what makes serious people trust the rest.
7. Post-production cues
- Freeze-frame + highlight box on the matching hashes (flagship step 5, clip 2).
- 1–2s holds on every honesty card and every hash.
- Lower-thirds from the caption bank only — no invented claims.
- Brand voice: humble + confident, geometric. Donations/credit: ACB Content; site core.acbcontent.org.