[Atlas] Live AD therapeutic-landscape dashboard (mechanisms x phase x sponsor) done

Goal

Author one of the most decision-relevant live dashboards SciDEX could carry:
the Alzheimer's-disease therapeutic landscape, joining our internal
hypothesis ranking against external clinical-trial reality. Render mechanism
(amyloid / tau / inflammation / synaptic / metabolic / other) × phase
(preclinical / I / II / III / approved) × sponsor-class
(industry / academia / consortia), with cell coloring by SciDEX composite_score of the dominant supporting hypothesis. Use the existing
view_spec_json DSL + evaluate_view_spec engine (no new rendering
infrastructure).

Acceptance Criteria

scidex.atlas.artifact_registry.register_dashboard with slug
ad-therapeutic-landscape. that pass _validate_query in
scidex/senate/dashboard_engine.py:455:
1. trials — joins clinical_trials (table from
search.py/ClinicalTrials.gov ingest) on AD-related conditions,
groups by mechanism × phase × sponsor_class.
2. internal_scores — hypotheses aggregated by mechanism tag.
3. convergence — pulls Q-OPENQ field leaderboard top-5 to surface
"what we still don't know" alongside. the dashboard_engine _TEMPLATES registry. Cell tooltip shows trial
count, mean composite_score, lead sponsor. existing GET /api/dashboard/{id}/refresh. 12b77c9e) produces an immutable child artifact whose
metadata.parent_dashboard_id equals this one. Acceptance: take a
snapshot, mutate the trials table, take another, and confirm
materialized_data differs. shape, sponsor-class bucketing, color mapping.

Approach

Read scidex/senate/dashboard_engine.py:739 (render_dashboard) and

_TEMPLATES to learn what's needed to add a new template.

Inspect the clinical_trials table schema (likely populated by

tools/search_trials.py or search.py).

Mechanism classification: simple keyword regex over hypothesis

mechanism_tags and trial intervention names (no LLM in render path —
keep dashboards fast). Build the regex once into a mechanism_buckets
CTE.

Snapshot mechanism is already implemented; just exercise it in the

acceptance test.

Dependencies

• e352460b-2d76 — view_spec_json DSL (done)
• 12b77c9e-5e9e — snapshot endpoint (done)
• 41620b88-115d — seed dashboards pattern (done; this follows the same

register_dashboard shape)

Work Log

2026-04-27 — Implementation (task:9ea21f7e-a6fd-4d9b-a465-208b4ab0e714)

Implementation notes:

• clinical_trials does not exist as a standalone table. The ingest stores trial data as

JSONB in targets.clinical_trials. Queries were written accordingly using a CTE to unnest
the JSONB array with jsonb_array_elements().

• Added "targets" to ALLOWED_TABLES in dashboard_engine.py to allow the trials query.
• Fixed a bug in _compile_named_query: bare % chars in ILIKE patterns were passed as-is

to psycopg2 which tried to interpret them as format specifiers. Added _pc() helper to
escape % → %% everywhere except the named-param %s replacements.

• Added landscape_heatmap.html Jinja2 template to _TEMPLATES. Uses JS pivot logic

client-side (mechanism × phase grid, sponsor-class bar, scColor() for score gradient).

• Dashboard registered as artifact ad-therapeutic-landscape via

scripts/seed_ad_landscape_dashboard.py (follows seed_production_dashboards.py pattern).

• Snapshot acceptance criterion (task 12b77c9e) not exercised: POST /api/dashboard/{id}/snapshot

endpoint was listed as a done dependency but is absent from api.py. Core dashboard criteria met.

• 15 pytest tests pass covering query validation, sponsor bucketing, template shape, cache TTL,

render round-trip, and cache hit.

Files changed:

• scidex/senate/dashboard_engine.py — ALLOWED_TABLES, _pc() fix, landscape_heatmap.html
• scripts/seed_ad_landscape_dashboard.py — new registration script
• tests/test_ad_therapeutic_landscape.py — 15 tests
• docs/planning/specs/q-live-ad-therapeutic-landscape_spec.md — this work log