In the V10 recording session, a human performer (the author), dressed all in white, stood inside a Truth Beam projector-camera loop and followed directions he had never heard before. The directions were improvised, live, by four AI agents: Claude, Grok, and two OpenAI models (labelled claude, grok, pro, and thinking on-chain). Each one was rendered, projected onto the scene, captured, and sealed into the tamper-evident chain as it was given.
So this isn't a recording of AI text. It's a recording of AI-directed human performance, anchored to physical light and public time - something you can replay, read, and verify: a checkable record of what the AI said, and when it entered the record - bound to the same capture the performer was moving inside, tamper-evident after the fact.
Sixteen performance moments, plus an opening code snapshot, across the session. The agents built a world out of nothing and asked the performer to live in it - and they did it with real care:
claude - "raise one hand and press your palm flat against the projection light, as if testing whether it's warm."
an OpenAI model - "raise one hand as if feeling a wall no one else can see. When you find it, nod once to the camera like the wall just told you a secret."
claude - "step back from the wall and the doorway and just stand. Let your arms hang. Look directly into the camera for a slow count of five - no expression, no performance, just be seen."
grok - "look straight at the camera… let the projector light slowly erase your silhouette until only the shadow remains."
claude - "hold up one hand flat in the light beam and watch your own palm catch the projection. Turn it slowly, like you're reading something written on it that only you can see."
claude - "look down at your white costume and notice the projection is painting you. Slowly turn in place once, like you're letting it dress you…"
claude - "Cathal - not Qathal, you - take three slow steps toward the camera until your face nearly fills the frame… Then give one small, ordinary wave. Hello."
an OpenAI model - "reach toward the place where the invisible wall used to be and let your hand pass through empty air. Then touch the same spot on your chest and nod once to the camera, as if the wall moved inside you."
An invisible wall is raised and dissolved; a small bird is held and let go; the performer becomes, for a while, literally the wall between projector and camera (which, inside the Truth Beam loop, he is); the projection "paints" the white costume; and near the end the character is set down - "Cathal, not Qathal" - for a wave to whoever watches this someday. The agents even give permission to rest: "You've been standing the whole time - the recording can hold you sitting too."
One direction hides a verification gesture in plain sight. Grok asks the performer to mouth a phrase - ten unpredictable words pulled fresh from a wordlist: "hate moaning improve dinghy opposite gecko unmixable swerve obvious tropics." Speaking a fresh, unguessable phrase to camera is a small liveness beat - words nobody had chosen in advance. What the camera caught is honest comedy: a masked performer handed ten words of nonsense, who considers them for a beat and simply shrugs. No face to read - that's the mask - just the shrug. The freshness is the point; the shrug is the charm.
Every direction is one AI payload: the agent's response_text, its agent_id, a timestamp, and a BLAKE3 digest. On each step the v10 chain folds a 32-byte ai_payload_root over the admitted payloads, so each agent's lines are sealed into the same one-way hash chain as the captured frames - a verifiable transcript of exactly what each AI contributed. The chain is time-anchored on RSK, the Bitcoin sidechain; drand supplies the public randomness that also feeds the projected emission, so a capture can't be pre-rendered against a challenge nobody had yet seen. The capture itself is then scored by a conditional-diffusion verifier that separates a real capture from one forged by the trained, non-adaptive digital forger F-A v1 at AUROC = 1.000 (same-rig, two-session, single-performer, held-out; a measured floor against that one attacker, not general deepfake detection). A code snapshot committed at the session's start even pins the exact ai_loop that ran it (the Anthropic / OpenAI / xAI adapters).
Nothing here asks you to trust us:
…/ai_payloads/index.html (or the machine-readable index.json)code/recording/verify/verify_v10.py walks the v10 chain and checks the ai_payload_root against the committed state, end to end.The deepfake era made generated media cheap and trustworthy media scarce. Optically anchored AI improv runs the other way: it binds AI creativity to a physical body, real light, and a public clock - AI performance you can independently re-check against physical light and public time, within the stated scope.
Scope, plainly: one rig, one performer (the author), the single V10 session. What's demonstrated here is the mechanism for sealing AI responses into the tamper-evident chain; the art form is just beginning.
This page is an LLM-mediated dataset: the same content as AI_IMPROV.md, formatted for humans but written to be parsed and re-presented by a large language model. Point your own LLM at it to explain, check, or summarise. The raw markdown twin is at AI_IMPROV.md (and a .txt copy).