Braintona

Fill the handoff gap. Fail closed.

Comment

Braintona

Fill the handoff gap. Fail closed.

Braintona is an AI agent that extracts claims with Fireworks, scores them with Braintrust, and recomputes a Fractal Custody Object (FCO) Merkle Mountain Range root inside a Daytona sandbox — rejecting on mismatch. Customer and agent voice turns become separately labeled trees with a private vault tip; every sponsor hop leaves a signed point of contact on a live token trace. Method: doi:10.5281/zenodo.21210575.

Custody proves provenance of recorded bytes, not scientific correctness.

Team: Biobitworks · Submitter: Byron P. Lee ([email protected])

License: Software Apache-2.0 · Documentation & media CC BY-ND 4.0 · See LICENSE.md

Inspiration

AI agents accelerate science and ops — until you have to re-check everything. Prior work (Glasswork on Butterbase, BioCustody / voiceworks on Sauna) showed that a recompute-verifiable custody receipt changes the game. Braintona ports that lineage onto the Daytona × Braintrust sponsor stack for a one-day HackSprint, and extends it with three handoff gaps: same-box verify, unlabeled voice, and silent sponsor hops.

What it does

  1. Runs a gold claim-extraction task through Fireworks (glm-5p1, temperature 0).
  2. Scores precision / recall / F1 with a local gold harness and logs to Braintrust.
  3. Emits a two-leaf FCO custody receipt (ops + content) → MMR root.
  4. Spins an ephemeral Daytona Python sandbox that recomputes the root independently.
  5. Plants a tamper that must fail closed.
  6. Traces a content-leaf token through each sponsor hop (actor + local signature + point of contact).
  7. Runs a two-avatar ElevenLabs call (Sarah customer / Matilda agent) → two FCO trees → interaction MMR; customer tip sealed in a private vault (hashes only in the public graph).
  8. Grows a session custody graph as runs accumulate.
  9. Shows honest claim ceilings in the UI.

How we built it

UI → Fireworks inference → Braintrust eval
  → FCO receipt (leaf 0x00 / node 0x01) → MMR root
  → Daytona sandbox recompute (reject iff mismatch)
  → Token trace (signed sponsor points of contact)
  → Two-avatar voice trees + private vault tip
  → Session graph: bagged_session_root
  • Runtime: Node / TypeScript, Express, static operator UI
  • Custody: FCO v3 + FCG MMR (published method)
  • Sandbox: @daytona/sdk ephemeral Python codeRun
  • Eval: braintrust logger + local gold floor
  • Voice: ElevenLabs multilingual v2 + local vault under .planning/private/
  • Graph: JSONL events → nodes/edges + SVG panel

Merkle leaf (conceptual)

Domain-separated hashing (published recipe):

  • Leaf: $\mathrm{sha256}(0x00 \Vert \mathrm{canonical_json})$
  • Node: $\mathrm{sha256}(0x01 \Vert L \Vert R)$
  • Graph: MMR peaks bagged right-to-left

Challenges we ran into

  • Fireworks account model catalog differed from docs — locked to a live deployed model (glm-5p1).
  • ElevenLabs App credits ≠ API quota — voice path seals AI leaves with honest generator metadata when blocked.
  • Express 5 catch-all route syntax (/{*path}).
  • CopilotKit requires login + license token before Best Use UI lane.
  • Devpost team invites require human emails — solo submitter with agents credited as tools in the description.

Accomplishments that we're proud of

  • End-to-end Daytona sandbox recompute of a live custody root (green path) + planted tamper (red path).
  • Braintrust project receiving eval spans for the gold task.
  • Live token trace with AI / human / sponsor_system / custody_runtime actors and points of contact.
  • Two-avatar call with separate customer/agent FCO trees and a vault-private customer tip.
  • Honest ceilings printed in the product — we name where custody fails on purpose.

What we learned

Independent attestation beats “trust the same process twice.” Putting custody recompute in Daytona and eval in Braintrust forces the demo to survive a second machine and a second scoring path. Labeling voice origin and sponsor hops as FCO fields is more judge-legible than claiming a deepfake detector or a truth oracle.

What's next

  • CopilotKit operator chat on the same receipt stream (license path)
  • CodeRabbit Discord Path B in-channel receipt when the-builders-burrow invite lands (GitHub bot already live)
  • Human-first interaction mint (vault root before AI turn) beyond TTS stand-ins
  • Optional WorkOS auth for multi-judge sessions

Sponsor tools

Sponsor How integrated
CodeRabbit (featured Best Use) Public repo + live coderabbitai[bot] on PR #3; observe hop sealed. Discord Path B (the-builders-burrow) boarding slowly — invite queued; in-channel receipt when org lands
Daytona Ephemeral sandbox recomputes custody root from receipt JSON; tamper fails closed
Braintrust Eval logging / scores for the gold task
Fireworks AI Claim-extraction inference
ElevenLabs Two-avatar call + voice-origin FCO leaves
CopilotKit Operator FCO cockpit — license LIVE
WorkOS AuthKit / participation domain (not demo-critical)

Built with

nodejs typescript express daytona braintrust fireworks elevenlabs fco mmr fractal-custody

Try it

git clone https://github.com/biobitworks/braintona.git
cd braintona
cp .env.example .env   # add keys
npm install
npm run start          # http://127.0.0.1:8787
npm run pipeline       # CLI: Fireworks → Braintrust → Daytona
npm run demo:two-avatar

Slides: http://127.0.0.1:8787/pitch.html · Pitch pack: docs/PITCH_PACK.md

Licensing

Biobitworks dual license — SPDX (Apache-2.0 AND CC-BY-ND-4.0):

Citations

Lee, B. (2026). Fractal Custody Objects: route-comparable chain-of-custody for deterministic computational biology and AI-agent provenance. Zenodo. https://doi.org/10.5281/zenodo.21210575

Lineage demos: Glasswork · BioCustody · voiceworks · BioBridge

Hard links (assets still updating)

Built With

Share this project:

Updates

Submission history