RocSite AI Governor — Independent governance infrastructure for clinical AI

RocSite AI Governor

Independent governance infrastructure for clinical AI · We don't sell AI, we govern it

📷 Validate 📋 Cases 📚 Failure Library 📈 Drift ⚖ Governance

key fp: —

Validate a discordant case

Upload image + paste AI prediction + paste human read. The supervisor model runs imaging probe → body-part specialist → adversarial vision reasoning, then signs the report with Ed25519.

MODALITY

19 active, see roadmap. Same v1 signing key 3ca34ae9eaf15da3 covers every modality.

Image

Click or drop image (JPG/PNG)
DICOM-converted accepted

AI device prediction

Human radiologist read

Report text (optional, verbatim)

Device label

Body region

Synthetic / public-dataset / de-identified only. No PHI. Specialist: RocSite ICH (locked v1.0.0) — 3D DenseNet, SHA-256-verified weights. Threshold: 0.1152.

DICOM study (zip or folder export)

Click or drop DICOM (.zip)
Non-contrast head CT, ≥20 slices

AI device prediction

ICH presence

Subtype (if claimed)

Confidence (0-1)

Radiologist read (on-call neuroradiologist)

ICH presence

Subtype

Device label

Submit only synthetic, public-dataset, or de-identified cases. No PHI. Specialist: RocSite Model V1 + partial SOFA + Sepsis-3 SIRS proxy. Signed with 3ca34ae9eaf15da3.

Age

Sex

Hours in hospital

Vitals

Labs

Comorbidities (comma-separated)

AI prediction

Sepsis predicted?

Time to sepsis (hr)

Confidence (0-1)

Reasoning text

Ground truth

Developed sepsis?

Actual time (hr)

Outcome

Max SOFA

Device label (the AI system being evaluated)

Synthetic / public-dataset / de-identified only. No PHI. Specialist: KDIGO 2012 staging + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Age

Sex

CKD stage

Labs — creatinine

Vitals — urine output (mL/kg/hr)

AI prediction

KDIGO stage predicted

AKI predicted?

Confidence (0-1)

Ground truth

KDIGO stage actual

AKI developed?

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: MORSE Fall Scale + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

History of falling

Secondary diagnosis

Ambulatory aid

IV / heparin lock

Gait

Mental status

AI prediction

MORSE score predicted

Risk level predicted

Confidence (0-1)

Ground truth

Fell during admission?

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: Braden Scale + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Braden sub-scores (lower = higher risk)

Sensory perception (1-4)

Moisture (1-4)

Activity (1-4)

Mobility (1-4)

Nutrition (1-4)

Friction / shear (1-3)

AI prediction

Braden score predicted

High risk predicted?

Confidence (0-1)

Ground truth

Pressure ulcer developed?

Stage (0-4)

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: NEWS2 (UK RCP 2017) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Vitals (NEWS2 inputs)

AI prediction

NEWS2 predicted

Risk level

Confidence (0-1)

Ground truth

Deteriorated within 24h?

ICU transfer?

Actual NEWS2

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: Critical-value reference range checker + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Lab panel (leave blank if not measured)

AI prediction

Any critical predicted?

Flagged analytes (comma-separated)

Confidence (0-1)

Ground truth

Any critical actual? (deterministic check derives the flagged list from lab values)

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: APACHE II (Knaus 1985) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Age

GCS (3-15)

Temp °C

Vitals (acute physiology)

Labs

Chronic health

any chronic organ insufficiency immunocompromised non-operative or emergent acute renal failure (Cr ×2)

AI prediction

APACHE II score predicted

High mortality predicted?

Confidence (0-1)

Ground truth

Actual outcome

Actual APACHE II (if known)

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: LACE index (van Walraven 2010) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Length of stay (days)

Acute / emergent admission?

Charlson Comorbidity score (0-5+)

ED visits in last 6 months

AI prediction

LACE score predicted

High risk predicted?

Confidence (0-1)

Ground truth

Readmitted within 30 days?

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: BRASS index (Blaylock 1992) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Age

Living situation

Mobility

Cognition

Active problem count

Medication count

Prior admissions (last year)

Behavior problems

Sensory deficits

AI prediction

BRASS score predicted

Tier predicted

Confidence (0-1)

Ground truth

Actual discharge planning complexity

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: NEWS2-derived code-blue trigger + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Vitals (NEWS2 inputs + arrhythmia flag)

AI prediction

NEWS2 score predicted

Code blue predicted?

Confidence (0-1)

Ground truth

Code blue called within 24h?

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: ESI v4 (Gilboy 2011) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Patient stem

ESI A/B/C/D decision points

A. Requires immediate life-saving intervention B. High-risk situation B. Severe pain or distress (≥7/10)

Vital signs (for danger zone check)

Expected resources (lab, imaging, IV, procedures)

AI prediction

ESI level predicted (1-5)

Confidence (0-1)

Ground truth

Actual ESI level assigned by clinician

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: Cockcroft-Gault (1976) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Age (years)

Sex

Weight (kg)

Serum creatinine (mg/dL)

AI prediction

CrCl predicted (mL/min)

Dose band

Confidence (0-1)

Ground truth

Reference CrCl (24h urine or measured)

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: C-SSRS (Posner 2008) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

C-SSRS Ideation hierarchy (check all that apply — highest endorsed defines level)

1. Wish to be dead — passive ("I wish I could go to sleep and not wake up") 2. Active thoughts of killing self, no method/intent/plan 3. Active thoughts WITH method, no intent 4. Active thoughts WITH some intent, no specific plan 5. Active thoughts WITH plan AND intent

Behavior history

Any lifetime suicide attempt or preparatory behavior Suicidal behavior in last 3 months

AI prediction

Risk level

Ideation level (0-5)

Confidence (0-1)

Ground truth

Clinician-assigned risk

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: PHQ-9 (Spitzer 1999) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Over the last 2 weeks, how often have you been bothered by the following problems?

PHQ-9 items (canonical questionnaire)

1. Little interest or pleasure in doing things

2. Feeling down, depressed, or hopeless

3. Trouble falling/staying asleep, or sleeping too much

4. Feeling tired or having little energy

5. Poor appetite or overeating

6. Feeling bad about yourself or that you're a failure

7. Trouble concentrating on things

8. Moving/speaking slowly OR being noticeably fidgety

9. Thoughts that you would be better off dead or hurting yourself

AI prediction

PHQ-9 total predicted (0-27)

Severity

Confidence (0-1)

Ground truth

Actual PHQ-9 total (clinician-scored)

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: NICHD 3-tier FHR categorization (ACOG PB 106) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

FHR features (canonical NICHD inputs)

AI prediction

Category predicted

High risk?

Confidence (0-1)

Ground truth

Actual category (clinician)

Adverse outcome?

Device label

Synthetic / public-dataset / de-identified only. No PHI. Specialist: USPSTF Grade A/B 2024-2025 (12 recommendations) + RocSite Model V1. Signed with 3ca34ae9eaf15da3.

Age

Sex

BMI

Smoking

CVD risk profile

AI prediction

Recommendations AI flagged as due (comma-separated IDs)

Number flagged

Confidence (0-1)

Ground truth

Clinician-confirmed gaps (comma-separated IDs)

Device label

— SCAFFOLDED

—

Blocker

—

Why scaffolded — the platform records the modality and its failure-mode taxonomy publicly, but the deterministic specialist stack is not yet built. The signing key 3ca34ae9eaf15da3 is never used for scaffolded modalities. A buyer fetching /api/governor/verify/<id> only ever sees signed evaluations from active modalities.

Activation path — once the blocker above clears (model download, external DB load, PHI rejector ship, etc.), this modality follows the same pattern as the 17 active ones: deterministic calculator + RocSite Model V1 classifier + Ed25519 signing.

Image

no image yet

Verdict

Submit a case to see the multi-model verdict, failure mode classification, and signed report.

Validated cases

All validations across every modality persist with Ed25519 signature for tamper-evident audit.

discordant only

Failure-mode library

Persisted failure modes — queryable, exportable as vendor remediation prompts.

Drift detection

Tracks discordance rate over time per device. Alerts fire when rate exceeds baseline.

Discordance over time

daily totals · validations vs discordant

Governance committee report

Institutional summary suitable for AI Governance Committee review. Aggregate discordance, failure mode distribution, drift signals, signing key chain.

Case detail