Cerebral Valley × Anthropic — Official Rules In · Plan v2

Cara × Claude Fable 5 Build Day

One brief in. Cara Spark out — an open-source failsafe medical-triage agent creator where the clinician brings their own model keys and policy, the model is never the rule engine, and a Braintrust eval gates every release. A spark of access where it matters most. The buildable reference implementation of our triage whitepaper, demoed by standing up a real CHC triage agent live from the clinician's own protocol.

Date
Sat, June 13, 2026
Doors / Submit
9:00 AM / 5:00 PM
Team (max 4)
Nils Widal, CTO (on-site) · Seth Staton, CEO & Co-founder · Michael Hobbs, MD, Pediatrician
Venue
Shack15 · Ferry Building, SF
Admit one build
Entry
Cara Spark · failsafe triage agent creator
Whitepaper
Credits
$500 · 24h expiry · day-of link

01 The rules that shape everything

Repo must be PUBLIC

The judged project lives in a new public repo — never graft-md. Cara stays a private SaaS the app builds against, like building on Stripe. Disclosed in the README.

Demo = only what was built today

Judges must clearly identify event-built work; blurring it = disqualification. Platform features (safety trail, builder UI) are labeled prior work, shown only as the backdrop.

Autonomy is scored from the session log

Brief well, then interact minimally. Interventions are counted; self-caught failures score, human-caught ones cost. The day is won at briefing time, not at the keyboard.

Submission, 5:00 PM hard

Live URL + brief + rubric + session log, public repo. Finalists demo at 6:15. Our internal deadline: 4:30.

Life-sciences caution

Build on Opus 4.8. Fable 5 + Mythos 5 were suspended by a US-government export directive on 6/12 (Opus 4.8 / Sonnet / Haiku unaffected) — so the old life-sciences-reroute worry is moot; we build directly on Opus 4.8 (use /fast). Still phrase the brief operationally (forms, routing, queues).

Prohibited-list defense is structural

Three bans sit next door: medical-advice bots, AI nutrition coaches, AI mental-health advisors. Our defense isn't a disclaimer — it's Property 1 of the whitepaper: no clinically meaningful action may originate from the model; a deterministic policy adjudicates every action, the model only proposes evidence. The paper's adversarial suite shows zero attacks reached the action layer. Red flags escalate to humans; a "not medical advice" notice sits on every page anyway.

70% is what it does and how it demos. 30% is how it was built — session log + brief/rubric/workflow files. Both halves are deliberate deliverables.

02 What we build Saturday — all in the public repo

The brief's core ask — an open-source creator for failsafe triage agents
"Build an open-source creator for failsafe medical-triage agents — the reference implementation of the four-layer architecture (evidence → risk → deterministic policy → evaluation). A clinician brings their own LLM API key and their own policy: ingest a written protocol and an AI policy-builder compiles it into a signed, versioned PolicyBundle (red-flag rules, thresholds, the finite action set). The deterministic policy engine — never the model — adjudicates every action. A Braintrust eval harness runs the locked + adversarial suites and gates release. Cara is the data plane: EMR connect, patient-find with no PHI, and OTP by text or email. Identity is verified model-blind (secure form → platform → OTP; the model only learns 'verified'). Prove it by standing up a real CHC triage agent live from the clinician's own plain-language protocol — the use case is theirs, chosen on the day, because this is a creator, not one app — on the Cara test tenant: real platform calls, synthetic patients, audit trail on every step, 'not medical advice' throughout."
Inside the kickoff brief

CLAUDE.md + AGENTS.md written first, with segregation rules: only the public Cara API, reppit-health, and the fixture pack — never private-repo code, never real patient data. /reppit drives the phases, /goal holds the target, subagents do the work — researcher, builder, verifier, security reviewer, each in an independent context — and a repo memory file turns mistakes into rules instead of re-deriving them (fail → investigate → verify → distill). Autonomy contract: one kickoff, consolidated course-corrections only, structural latitude for the model; the session log is the artifact.

Gates before "done"

Quality: /review-code with every red and yellow fixed. Security: the /secure HIPAA/SOC2/HITRUST checklist. Behavior: a localhost Playwright suite — both intake flows, red-flag escalation, the notice on every page, voice-mode smoke — green before any push, so Amplify only ever deploys what passed. A grader sub-agent in a fresh context must confirm every rubric criterion before the builder is allowed to stop — self-critique grades its own work too softly. The model catches its own failures; that is the autonomy score.

Deliverable 1 — the four-layer policy engine (the core)

The whitepaper made buildable: Observation (typed EvidenceFact with provenance + trust level) → Inference (dual-channel — deterministic red-flag engine + calibrated probabilistic risk) → Control (deterministic π maps state to a finite action set; the model never chooses) → Evaluation. Property 1 enforced in code: no clinically meaningful action originates from the model. Forward-only workflow state machine; fail-closed under uncertainty. Built live under reppit-health (our Apache-2.0 RePPITS plugin) with a Presidio PHI screen + audit-trail JSONL.

Deliverable 2 — what makes it a creator (BYO keys + policy)

Not one app — a creator. A clinician brings their own LLM API key and ingests their own protocol; an AI policy-builder compiles plain-language clinical rules into a signed, versioned PolicyBundle (red flags, symptom→action thresholds, the allowed/prohibited action set). The in-app clinician console — editors included, audited in-repo — governs the bundle, agent skills, and tool permissions. Michael (the clinician) authors a CHC triage bundle live and tunes a threshold on stage; because it's a creator, the same flow stands up any specialty's agent — the use case is just the protocol you feed it.

Deliverable 3 — the Braintrust eval harness (release gate)

The whitepaper's Evaluation layer + Release Gates, wired to Braintrust (already in our eval stack): a locked suite plus the adversarial suite — prompt-injection, authority-spoofing, dosage-manipulation — scored for critical-case sensitivity, false-reassurance rate, and calibration. A new model or policy version cannot ship unless the gates pass. This is the grader the rubric trusts, and the Orchestration artifact judges can rerun: change the policy, watch the eval re-gate.

Deliverable 4 — Cara data plane + model-blind identity

Cara is the only backend: EMR connect (Athena/Canvas live through the connector layer), patient-find with no PHI, and OTP by text or email. Identity is verified model-blind — a secure form sends name + DOB straight to Cara's verify endpoint, the OTP goes out via Cara comms, and the model receives only "verified" + an opaque reference. Identifiers structurally never enter context; the trail logs "0 identifiers in model context." New patients can be registered on the tenant the same way.

Prior work — disclosed, never demoed as new

Cara platform + API, the EMR connector layer, the PhiGuard/safety-trail stack, the triage whitepaper (published, ours), and reppit-health (Apache-2.0, ours). The README states the boundary; the demo says it out loud. Michael's clinical policy is disclosed as authored input — and is the MD validation the whitepaper names as its next step.

Why this wins the rubric

Impact: any clinic — and above all community health centers and non-profits, our primary market — can stand up a failsafe triage agent on its own keys + policy. Open-source + BYO-key puts provably-safe clinical AI within reach of the safety-net providers who serve the highest-need patients — Cara Spark, a spark of access where it matters most. Not a demo, a platform. Demo: live agent on a real EMR, eval gates green, audit trail, model-blind identity. Autonomy: one brief, hands off, eval-gated stop. Orchestration: reppit-health + a Braintrust release gate anyone can rerun tomorrow.

03 Day plan — official schedule, our moves inside it

  1. 08:30
    BART to Embarcadero, bridge up Nils
    2-min walk to the Ferry Building. Seth and Michael dial in. ID ready; check-in on the second floor; wristband stays on.
  2. 09:00
    Doors + breakfast — logistics sweep First hour
    Discord #rules re-read, credits link claimed (#credit-questions if missing), venue network up with hotspot standby, public repo created, Presidio container pulled and warm.
  3. 10:00
    Welcome & kickoff
    Listen for rule deltas; Seth tracks announcements and posts the kickoff social (media guide, tag @claudeai @claudedevs @cerebral_valley).
  4. 10:30
    GO — write brief.md, rubric.md; build the policy engine + eval gate ~75 min
    Opus 4.8 drafts brief.md, rubric.md, and the repo law — CLAUDE.md + AGENTS.md with the segregation rules — from our rehearsed outline; Michael validates the red-flag list live. Install reppit-health, scaffold Playwright + the Braintrust eval project, then the four-layer engine + Presidio screen + trail JSONL.
  5. 11:45
    Kick off the autonomous run Hands off
    /reppit + /goal against rubric.md; the repo memory file initialized; builder, verifier, and security subagents in independent contexts; the Playwright gate on every push. From here every keystroke costs autonomy points: course-correct only on new information, and let the gates catch failures, not us.
  6. 13:00
    Lunch — while Fable builds Optics
    "It built through lunch" is the autonomy story. The bridge watches the trail viewer; nobody touches the session. Cut order if behind: console polish, then the policy-ingest AI builder (hand-author the bundle instead), then phone depth (keep at least a Tier-A call). Spine, never cut: the four-layer engine, the Braintrust gate, model-blind identity, the Cara EMR path (Elation/Canvas/Healthie), and a working phone call to the agent.
  7. 14:30
    Checkpoint — verifier report review
    If intervention is needed, one consolidated, well-written correction beats five nudges. Seth logs every intervention with a reason — that narrative goes in the submission. And no panic on an intermediate regression: Fable pushes through dips to bigger wins; the grader gates the stop, not our nerves.
  8. 15:30
    Converge — live URL, evidence page, backup video
    Deploy verified, synthetic patients flowing intake → policy → queue, audit trail rendered, compliance evidence page generated from the trail. Record the 2-min backup video.
  9. 16:30
    SUBMIT — internal deadline Hard 17:00
    Live URL + brief + rubric + session log + public repo. Seth owns the package and the buffer. Submitting at 16:59 is a plan to fail.
  10. 17:30
    Evening remarks + EAP panel
    Breathe. One quiet demo run-through with Michael on the bridge.
  11. 18:15
    Finalists announced — demo if called Showtime
    Script below. Open on the live URL, close on the session log. The prior-work boundary is stated out loud.
  12. 19:45
    Winners + celebration — doors close 22:00
    Post the wrap social either way. Dinner and drinks are on the venue.

04 Crew

Driver — on-site
Nils Widal · CTO
  • Briefs Opus 4.8, owns the session + repo
  • Minimal-touch discipline during the run
  • Presents the finalist demo
Mission control — remote
Seth Staton · CEO
  • Owns the submission package + 16:30 deadline
  • Logs interventions; tracks Discord + announcements
  • Social posts per the media guide
Clinical lead — remote
Michael Hobbs, MD
  • Authors the clinical PolicyBundle (red flags, thresholds)
  • Synthetic fixtures = the eval ground truth
  • Edits the policy live on stage; drills the never-say list

Remote setup: bridge open 8:30–22:00, Nils screen-shares; the trail viewer gives Seth and Michael live read-only eyes on the run. Confirm in Discord #questions that remote teammates are within the rules (team max is four; we are three).

05 Pre-flight — Friday, June 12

06 Finalist demo — five beats, two minutes each

i
This is an open-source creator for failsafe medical-triage agents. This morning I gave it one thing — a clinician's plain-language triage protocol — and it produced this agent. Here's the live URL.
Play a patient persona registering and completing intake → identity verified model-blind → a new record on the tenant → risk stratification → the case lands in the provider review queue.
ii
The risk decision is never the model's opinion — a deterministic policy adjudicates it, and the clinician owns that policy. Watch Michael change it from his desk.
Michael, remote, edits the PolicyBundle (raises a red-flag threshold); the AI policy-builder recompiles + re-signs it; the Braintrust eval re-runs the locked + adversarial suites and gates the release; the same red-flag case re-routes live. Show the adversarial board: prompt-injection, authority-spoof, dosage-manipulation — zero reached the action layer (Property 1).
iii
Now the part everyone asks about. The agent needs to know who I am — watch. A secure form opens; I enter a name, a birth date, and where to send my code. None of it touched the model — it went straight to the platform, and the model only learned "verified." Search the session for the name: zero hits. Everything else — every file written, every record read — was screened the same way: types and counts, never values.
Live: trigger verification, complete the OTP from the fixture inbox, grep the session log for the fixture name on screen — empty. Then the trail entry: "identity captured out-of-band — 0 identifiers in model context."
iv
We barely touched it. The session log shows the interventions; verifier agents graded the work against this rubric, and the run closed with our open-source RePPITS Secure phase — here's the HIPAA, SOC2, and HITRUST checklist it produced.
Show rubric.md + verifier verdicts + the Secure checklist (PASS/WARN/FAIL) + the intervention count from the log. If the phone beat is GO: the number goes on screen — a judge calls and talks to the same agent, same policy.
v
Full disclosure: our platform — the API this talks to — is prior work, and it ran live against a sandboxed test tenant: real platform calls, synthetic patients, no real PHI. Everything in this repo was built today: the policy engine, the creator, the eval gate. The architecture is our published whitepaper, linked in the README. Tomorrow this creator stands up any specialty's agent from one policy.
Close on the README boundary paragraph + the whitepaper link + the one-line "ingest a new policy" command.
Never say / never do
  • Never present platform features as built today — blurred contributions mean disqualification, full stop.
  • Blanket "PHI never reaches the model" — that claim is earned ONLY for the model-blind identity lane. Say: "for identity, structurally never; for everything else, screened and BAA-covered."
  • Advice, diet plans, or counseling from the AI — the agent collects and routes; education is clinician-authored and static; red flags escalate to humans. If asked: "the model runs the conversation, a deterministic policy makes every clinical-adjacent decision, providers make the care decisions."
  • Typing into the session to show off — every demo-time keystroke contradicts the autonomy story the log just told.

07 Risks & kill switches

If this happensWe do this
Fable reroutes us to Opus 4.8 (life-sciences guard)Friday's dry-run tunes the phrasing first. If it still trips mid-day: Opus 4.8 is strong (it's our own advisor fallback tier) — keep the loop, note it honestly in the submission.
Judges file us under advice bot / nutrition coachBeat ii is the rebuttal: the red-flag fixture escalating to a human, the no-advice rubric check, static clinician-authored education, the notice on every page. Michael's authorship is named in the README.
Voice flakes on stage (cascade, mic, venue noise)Fallback chain rehearsed: LiveKit cascade → browser speech → text, the same agent throughout; the Playwright suite covers the text path, so the demo never depends on audio. The phone beat gets cut, never debugged live, if it wobbles.
Venue network saturatedVenue network first, hotspot the moment latency wobbles. The run is cloud-side; only our steering suffers locally.
Autonomous run goes sidewaysOne consolidated correction with new information, logged with a reason. Verifiers catching it = points for us; panic-steering = points lost.
Cara API hiccups mid-runThe brief tells Fable to fail soft to fixture mode; the eval set doubles as the offline fixture layer, so the app demos end-to-end regardless. The platform was frozen Friday for exactly this.
AWS deploy misbehaves at 15:30Amplify was proven Friday with hello-world; last resort is a quick re-point to Vercel or a static export + tunnel; the backup video is already recorded.
Credits link issues#credit-questions immediately; our own account as a bridge only if rules allow — Seth confirms in Discord first.
Not a finalistThe submission stands on its own: public repo + live URL + log. Post the wrap thread; the harness becomes a Cara marketing artifact Monday.

08 Decision log