Spaces:
Running
A newer version of the Gradio SDK is available: 6.15.2
HTML Report Dashboard
Version: 1.8.0 Status: Active HTML/PDF surface note
STEM BIO-AI ships a self-contained interactive HTML dashboard as the primary human-readable output format. In the current 1.8.0 line, the dashboard reflects the post-C4/C5/C6 code-integrity split, AIRI detector-to-risk reasoning, policy-surface metadata, and audit-freshness summaries. The report is a single .html file with inline CSS, SVG, and JavaScript — zero external dependencies, fully offline-capable.
Generating the Report
stem scan /path/to/repo --format html
stem scan /path/to/repo --format all # JSON + MD + HTML + PDF in one pass
Output:
- default root:
stem_output/<repo_slug>/<repo_slug>_report.html - custom output directory:
<output_dir>/<repo_slug>_report.html
Example:
- browser preview: https://htmlpreview.github.io/?https://raw.githubusercontent.com/flamehaven01/STEM-BIO-AI/main/docs/assets/report-preview/yorkeccak_bio_report.html
- raw artifact:
docs/assets/report-preview/yorkeccak_bio_report.html
5-Section Structure
| Section | ID | Content |
|---|---|---|
| Executive Summary | #s1 |
Score gauge, 5-metric stat grid, notable risks, T0 alert banner |
| Decision Path | #s2 |
Stage bars (S1/S2R/S3/S4), score formula, and policy/configuration explainer |
| Code Integrity & Contract | #s3 |
C1–C6 + CC1–CC3 expandable cards, with C4/C5/C6 split visible |
| AIRI Risk Coverage | #s4 |
MIT AI Risk Repository donut + covered/gaps toggle table + domain filter cards |
| Evidence Detail | #s5 |
Full evidence ledger with severity filter chips |
Navigation between sections via a sticky top navbar with scroll-spy active state.
Interactive Features
Score Gauge
SVG semicircle speedometer (0–100). Arc color matches the tier:
- T0/T1: crimson (
#9B2335) - T2: amber (
#C07B10) - T3: teal (
#1A6FA8) - T4: green (
#1A7A47)
Expandable Cards (Section 3)
Each C1–C6 / CC1–CC3 card is clickable. Click expands the full evidence list (up to 8 entries). The caret rotates 180° on expand. Keyboard accessible via Enter/Space.
[C5 Compliance Boundary Integrity] WARN ▾
unsupported compliance claim in README.md ← collapsed
──────────────────────────────────────────────
- README.md:262 - HIPAA-compliant architecture (when self-hosted)
- Clinical-adjacent surfaces exist without an explicit non-diagnostic/non-clinical boundary. ← expanded
AIRI 7-Domain Shortcut Cards (Section 4)
The AIRI explorer starts with an All Domains card, followed by seven domain cards. Each card shows covered / gaps counts for that domain. Colors match the MIT AI Risk Navigator palette:
| # | Domain | Badge color |
|---|---|---|
| All | All Domains | primary navy |
| 1 | Discrimination & Toxicity | #B5272A crimson |
| 2 | Privacy & Security | #1E3A6E deep navy |
| 3 | Misinformation | #C07020 amber |
| 4 | Malicious Actors & Misuse | #6B3FA0 deep purple |
| 5 | Human-Computer Interaction | #1B7A4E deep green |
| 6 | Socioeconomic & Environmental | #7A6520 olive |
| 7 | AI System Safety, Failures & Limitations | #1A5568 dark teal |
Click a domain card → risk table filters to that domain only and the Covered (n) / Gaps (n) toggle labels update to that domain's counts. Click All Domains → full table restored. Cards with 0/0 are dimmed but still document bundle coverage boundaries.
AIRI Toggle (Section 4)
Two-button toggle: [Covered (n)] / [Gaps (n)]. Composes with the domain filter — both conditions must be satisfied for a row to be visible. The counts shown on the toggle buttons are dynamic and reflect the current domain selection.
Covered AIRI rows surface bounded why reasoning derived from:
- the triggered local detector ID,
- the local detector-to-risk mapping justification, and
- the trigger reason surfaced by the scan.
This is a review aid only. It does not mean AIRI independently audited the repository.
Evidence Filter Chips (Section 5)
One-click severity filter: All FAIL WARN PASS INFO. Filters ev-row elements by CSS class; active chip highlighted in navy.
Tooltip ? Icons
Every metric header carries a ? icon with a data-tooltip attribute. CSS ::after pseudo-element renders the tooltip on hover/focus — no JavaScript required.
[data-tooltip]::after {
content: attr(data-tooltip);
position: absolute; bottom: calc(100% + 8px);
/* ... navy background, 240px max-width, fade-in transition */
}
Scroll Spy Navigation
updateNav() listens to window.scroll (passive) and toggles .active on the nav link whose section is at the top of the viewport.
Architecture
The HTML renderer is split into three modules:
| File | Role |
|---|---|
stem_ai/render_html_components.py |
_C palette, _AIRI_DOMAIN_COLORS/NAMES, SVG generators (svg_gauge, svg_donut, svg_hbar), card builders (integrity_card, domain_card, airi_row, evidence_row) |
stem_ai/render_html_styles.py |
build_css(tc) f-string (tier-color parameterized), .domain-card/.domain-active CSS, JS constant with filterDomain() + _applyAiriFilter() |
stem_ai/render_html.py |
Five _section*() assemblers + domain-count computation + render_html(result) entry point |
The matching PDF surfaces now track the packet tiers:
level 1→brief_1p.pdf(legacy quick brief)level 2→detailed_5p.pdf(standard review packet)level 3→detailed_7p.pdf(full evidence packet)
The 5-page and 7-page packets both include AIRI coverage and bounded why: explanations, while the 7-page packet adds the deeper closeout and metadata pages.
Report Preview
Screenshot: current 1.8.0 HTML dashboard surface with the post-C4/C5/C6 integrity split and AIRI domain filtering.
Color Palette
Refined across the 1.7.x line to align with AIRI Navigator's visual language — deeper saturation, cooler backgrounds, AIRI-standard crimson for risk indicators.
| Role | Token | Value |
|---|---|---|
| Primary navy | _C["navy"] |
#0D1F3C |
| Link teal | _C["teal"] |
#1A6FA8 |
| Pass green | _C["green"] |
#1A7A47 |
| Warn amber | _C["amber"] |
#C07B10 |
| Fail / T0-T1 | _C["red"] |
#9B2335 |
| Accent purple | _C["purple"] |
#5B3D8A |
| Background | _C["lgray"] |
#F8FAFC |
| Divider | _C["mgray"] |
#E4EBF5 |
| Muted text | _C["dgray"] |
#5A6B80 |
AIRI Risk Coverage Section
The AIRI coverage section maps STEM-BIO-AI detectors to a governed local AIRI layer derived from the MIT AI Risk Repository V4:
- full local registry snapshot
- curated runtime bundle used by deterministic scans
- detector-to-risk mapping registry
The runtime HTML report reflects the curated bundle, not the entire AIRI universe.
Covered rows now show a short bounded explanation built from:
- the local detector ID,
- the detector-to-risk mapping justification, and
- the trigger reason surfaced by the scan.
This helps reviewers understand why a risk appears in AIRI coverage without implying that AIRI independently audited the repository.
Coverage rate = covered risk IDs / total IDs in detector scope.
Representative active mapping rows in the 1.8.0 line:
| Detector | AIRI Risk IDs Covered |
|---|---|
CC1_clinical_zero_default |
62.15.02, 02.09.00, 30.01.02, 69.01.01, 16.03.01 |
CC3_shallow_validator |
65.03.03, 02.01.03, 16.02.01, 47.03.01, 70.02.01 |
T0_hard_floor |
41.04.00, 61.02.28, 18.05.03, 56.14.00, 47.02.12, 16.05.02, 43.01.00 |
C2_dependency_pinning |
33.01.05, 24.04.01 |
S1_R2_unsupported_legal_or_compliance_claim |
69.01.00, 39.25.00 |
R2R_D5_single_external_service_dependency |
60.02.01, 72.04.02 |
C5_compliance_boundary_integrity |
24.01.03, 69.01.00 |
C4_exception_handling_clinical_adjacent_paths remains reserved for executable fail-open exception behavior in code.
C6_mock_auth_or_fail_open_boundary is currently a report-layer/code-integrity surface only. It does not yet carry an AIRI mapping row.
Five known gaps are reported for risks that require dynamic evaluation or extend beyond static scan scope.
File Size Reference
| Repository | HTML size |
|---|---|
maziyarpanahi/openmed (T1, 39/100) |
~164 KB |
| Typical T3 repository | ~40–80 KB |
Size scales with evidence ledger length. The renderer caps the evidence table at 200 rows.