NeuroAssist × CI — The Nervous System Interface

AI can read language. We're teaching it to investigate neurological signals.

A decentralized scientific collective for camera-based eye-movement measurement. Five agents, each with its own evidence access and authority, inspect a new measurement, compare it with the person's history held in Contextual Intelligence (CI), challenge each other, and return an evidence-backed observation or an abstention. Built for MIT ScienceClaw 2026.

Research prototype. It measures eye movement and reasons about measurement quality. It does not diagnose stroke, vestibular disorders or any condition, and it is not a medical device. All benchmark data is synthetic (NeuroAssist's own simulator). Nothing has been tested on people.

Three products

	what	how it is used here
NeuroAssist (parent A)	webcam eye-movement measurement with a quality gate (neurocore-0.4.0)	its signal engine is vendored unchanged into lib/neuroassist/ (checksums in VENDORED.sha256); a Python port in naci/neurocore.py agrees with it exactly on all 270 held-out traces (tests/test_collective.py)
Contextual Intelligence (parent B)	user-owned memory: bitemporal, trust-ranked, consent-gated, hash-chained audit	installed from its public repository at commit 2e50658 (requirements.txt) and used through its own classes: TemporalMemory, ConsentLedger, AuditLog, Store (naci/memory.py)
NeuroAssist × CI (child, this repo)	the collective, the benchmark, the website, the Evidence Auditor	independently deployable; neither parent is modified

Integration boundary: naci.session/1 (lib/contract.ts, naci/contract.py). Agent messages: naci.agent-message/1, which maps onto CI's ci.agent-message/1. Details: docs/ARCHITECTURE.md.

Results (held-out, 270 synthetic cases, thresholds frozen in commit dbb7baa before the run)

system	correct	unsupported conclusions	harmful artifacts caught
NeuroAssist gate + previous visit	58.5%	41.5%	81.3%
Single agent, one pass, full history	91.9%	8.1%	81.3%
Collective, raw history (no CI)	98.1%	1.9%	100%
Collective + CI	98.9%	1.1%	100%
Ablation: same checks in one function + CI	98.9%	1.1%	100%

What this does and does not show:

• The accuracy gain comes from the checks (independent recomputation from raw frames, a head-motion counterfactual, trust-aware baselines), not from organising them as agents: the single-function ablation matches it.
• What the organisation adds is fault isolation. Corrupt one component: the single function's unsupported-conclusion rate rises to 33.7% (memory), 32.6% (analysis) or 7.4% (robustness test); the collective stays at or below 1.1% for every agent, silent or faulty.
• CI beats raw history only on conflicting history (97% vs 90%), where unverified records are numerous enough to move a median.
• Finding about NeuroAssist itself: between 0.2 and 0.6 deg/s its gate passes 80–100% of recordings while the 95% interval contains the true value only 5–38% of the time (docs/ENGINE_BLINDSPOT.md). The benchmark stays above 2.6 deg/s.

Protocol, metrics and every deviation: docs/BENCHMARK.md. Results: benchmark/results/heldout.json.

Run it

npm ci
pip install -r requirements.txt pytest uvicorn
python3 -m uvicorn api.index:app --port 8000 &   # the collective (Python, FastAPI)
npm run dev                                       # the site, http://localhost:3000 (proxies /api/py to :8000)

Tests and benchmark:

python3 -m pytest -q                                 # 29 tests: parity, access control, faults, CI, auditor
npx vitest run                                       # vendored engine checksums, scenarios
npx tsx scripts/gen-cases.ts                         # regenerate cases (deterministic)
python3 -m naci.benchmark --split heldout --faults   # ~45 s
node e2e/smoke.mjs https://<deployment>              # smoke test a deployment

Layout

lib/neuroassist/     NeuroAssist engine, vendored unchanged
lib/scenarios.ts     synthetic case generator (NeuroAssist simulator + corruptions + ground truth)
naci/                the collective: contract, CI memory adapter, agents, baselines, benchmark
evidence_auditor/    reusable, domain-neutral Scientific Evidence Auditor (stdlib only)
api/index.py         HTTP API (Vercel Python function)
app/, components/    Next.js site and dashboard
benchmark/           cases, results, freeze hashes
docs/                architecture, benchmark protocol, engine blind spot, deployment

Reuse the Evidence Auditor

evidence_auditor/ checks that every claim a pipeline wants to report cites evidence that exists, is intact (sha256), is of adequate quality and allowed provenance, and has survived every challenge. Copy the folder or call POST /api/py/v1/auditor/audit. See evidence_auditor/README.md.

Team

Nishanth Revuri and Jayashabari Shankar.