Platform

The machinery behind the numbers

One event store, six free sources, a rate-limited HTTP layer and generated artefacts. Every boundary below is deliberate, including the parts that are absent.

375,378 events6 sources1.3 GiB SQLitedataset 2026-10-31

Architecture

Solid edges ship. Dashed nodes do not.

Sources → HTTP layer → event store → registries, analysis, statistics, bitemporal store, forecast registry → reports and dashboard. Roadmap drafts are drawn as unconnected dashed nodes and are not wired in.

MedRadar platform architecture: six public sources feed a rate-limited HTTP layer and one SQLite event store with an append-only revision ledger; company, track and product-code registries, analysis, statistics, the bitemporal store and the forecast registry read from it; generated reports and the dashboard publish from it. Roadmap drafts are drawn as unconnected dashed nodes.
Light and dark variants ship together; the page selects one by colour scheme. The figure can be panned horizontally on narrow screens.

source: docs/assets/architecture/architecture-en-light.svg; README.md → Architecture

Sources

Six free public sources, no credentials.

Row counts and event-date windows are from the committed data audit for the 2026-10-31 snapshot. The SQLite snapshot itself is not committed.

Source inventory — dataset snapshot 2026-10-31
SourceKindRowsEvent-date window
openFDA device 510(k)510k176,4661976-07-15 → 2026-10-18
openFDA device recalldevice_recall59,6462002-11-01 → 2026-10-28
openFDA device PMApma57,3721968-10-09 → 2026-10-16
ClinicalTrials.gov v2trial55,5921999-10-19 → 2026-10-30
SEC EDGAR 8-K corpus8k_catalyst20,9762015-01-02 → 2026-10-30
openFDA drugsfdadrug_approval5,3261966-12-14 → 2026-10-22
Total—375,3781966-12-14 → 2026-10-30

source: docs/DATA.md → sources and coverage windows; docs/metrics.json → meta.counts (source counts sum to the event total)

Source inventory figure: rows per source for the 2026-10-31 snapshot, with openFDA device 510(k) the largest at 176,466 rows and openFDA drugsfda the smallest at 5,326.
Committed snapshot — dataset 2026-10-31, package v0.5.0. Generated from the same counts as the table.

source: docs/assets/figures/sources-en.svg; docs/DATA.md

Coverage quality

What is joined, and what is honestly missing.

Coverage is reported as a share of events with the plain-language meaning next to it. Missing is reported as missing, never as zero.

Coverage quality — dataset snapshot 2026-10-31
SignalRowsShareWhat it means
Company key resolved132,29335.2 %About one event in three can be tied to a company. The rest are counted as unmatched.
Technology track resolved196,54252.4 %About half of events carry a track. Track-level aggregates are therefore half-blind.
Product code present290,90977.5 %Most device rows carry a product code, which is what the product families are built from.
Timestamp quality date375,378100 %Day resolution everywhere; no row carries a time-of-day publication instant.
Independent public time00 %published_at = available_at = event_date on every row. This is the constraint behind every return caveat.

source: docs/DATA.md → 覆盖质量; docs/metrics.json

Coverage figure: company key resolved on 35.2 percent of events, technology track on 52.4 percent, product code on 77.5 percent, and independent public time on 0 rows out of 375,378.
Coverage shares, including the zero. Dataset 2026-10-31, package v0.5.0.

source: docs/assets/figures/coverage-en.svg; docs/DATA.md

Registry and price layers

Companies
89 — 81 with a ticker, 43 in the focus set
Technology tracks
60
Product codes
7,097 — 1,958 mapped to a track (27.6 %)
Prices
1,898,409 rows · 1,017 tickers · 2016-10-31 → 2026-10-30
Signals
127

source: docs/DATA.md; docs/metrics.json

Capability matrix

Status, evidence path, and the caveat that travels with it.

The site vocabulary maps to the repository's generated one: implemented = working; limited = working with a measured gap or mechanism only; draft = branch-only; planned = not started.

  • implementedRuns on main and is exercised by the offline test suite.
  • limitedWorks, but a named gap limits what it can support (partial resolution, descriptive only, mechanism only).
  • draftExists only on a draft branch or PR. Not shipped and not wired in.
  • plannedRoadmap item with an explicit blocker. No code path claimed.
Capability matrix — repository evidence and validation note
CapabilityStatusEvidenceValidation note
Ingest: openFDA device/drug, ClinicalTrials.gov, SEC 8-Kimplementedmedradar/collect_*.py, intelligence/Counts reproduce from live sources.
Event store with bitemporal fieldsimplementedmedradar/db.pyEvery row carries event date, availability and quality; 0 rows of independent public time.
Append-only revision ledgerimplementedmedradar/evidence_core/Records the application write path; explicitly not tamper-proofing.
Company / track / product-code resolutionlimitedmedradar/config.py, config/*.json35.2 % company key, 52.4 % track; unmatched counted, never guessed.
Descriptive analytics: heat, mix shift, noveltylimitedmedradar/analyze.pyDescriptive, in-sample scores; no out-of-sample validation.
Event study, replay, simulated long-only bookimplementedmedradar/eventstudy.py, replay.py, paper.pyWorking as simulation: in-sample, 38 trades, not investment advice.
Bitemporal knowledge storelimitedmedradar/tkg/Mechanism only: the imported history has no independent public clock.
Forecast registry with as-of scoringlimitedmedradar/forecast_registry.pyMechanism only: 25 registered forecasts, 0 matured; calibration state insufficient_matured_forecasts.
Deterministic evidence-grounded research agentimplementedmedradar/research_agent.pyRetrieval-and-citation workflow; establishes no causality.
Reports and zero-dependency dashboardimplementedmedradar/report.py, docs/dashboard.htmlGenerated artefacts, not a live service.
FDA AdCom collectorlimitedmedradar/adcom.pyParses zero rows (client-side calendar); returns empty / source_changed.
CLI, offline demo and selftestimplementedmedradar/__main__.py1,160 offline tests, no network, no API key.
Market-to-R&D lanedraftdocs/MARKET_DATA.md, experiments/Recorded as a draft capability; its frozen artefacts and negative result are public.
Publication / patent lanedraftdocs/PUBLICATION_PATENT_DATA.mdOpenAlex counts evaluated; the US patent path is key-gated and not evaluated.
Early-signal study (5 firms)draftresearch/studies/early_signal_5firms/Phase 0 CONDITIONAL-GO; lead-time gate NO-GO; no public claim.
Paper drafts with claim mapsdraftpapers/Not submitted, not peer reviewed.
Weekly forecast-cycle workflowdraft.github/workflows/forecast-cycle.ymlCommitted; not executed in the authoring environment.
Live observatory lanedraftmedradar/live.pyPaper trading only; not investment advice.
Re-acquire independent public timestampsplannedREADME.md → RoadmapThe prerequisite for any point-in-time claim. EDGAR first.
Graph features in the scoring modelsplannedREADME.md → RoadmapThe relation graph informs the agent only, not the models.
Difference-in-differences / synthetic controlplannedREADME.md → RoadmapNo causal claim is made anywhere in the repository.

source: README.md → What is implemented; PROJECT_STATE.yaml → capabilities (7 working · 6 draft · 1 retired); docs/DATA.md

Deliberately absent

The most important part of the matrix is the hole.

A capability that is missing because the data cannot support it is a design decision. Each hole below has a stated reason.

Drafts, not features

Six capabilities are recorded as draft in PROJECT_STATE.yaml, and the lanes tracked in PRs #12–#30 are branch-only. None is served by this site and none is presented as a capability.

source: PROJECT_STATE.yaml; README.md → Not in main, not claimed

Retired on purpose

The monthly-resolution deployment detector was retired by decision D-045: at roughly 0.2 events per cell-month, a monthly detector cannot resolve a change. The deployment resolution is quarterly.

source: PROJECT_STATE.yaml → capabilities (retired); research/DECISIONS.md

Not evaluated or excluded

  • Patents — USPTO Open Data Portal requires a key; not evaluated.
  • Procurement — excluded: the snapshot is not point-in-time usable.
  • NMPA (China) — deliberately not scraped; values are null, never zero.
  • AdCom — parses zero rows because the calendar is client-side.

source: README.md → Limitations 5, 6, 9; docs/PUBLICATION_PATENT_DATA.md

What does work, end to end

The offline path works from a clean clone with no network: install, run 1,160 tests, build the demo artefact, and read the generated report and dashboard. That is the reproducible claim this project stands behind.

source: README.md → Reproduce; PROJECT_STATE.yaml → tests

Reproduction

Reproduce the procedure, not the numbers.

The 1.3 GiB SQLite snapshot is intentionally not committed, so a clean clone rebuilds the pipeline and the artefact formats, not the exact counts.

source: README.md → Reproduce; PROJECT_STATE.yaml → tests

Frozen studies rebuild with the artefacts in experiments/reverse_rd_v1/ and experiments/reverse_rd_v2/. Every metric definition is in docs/METHODS.md and the data contract is docs/SCHEMA.md; change those before changing a number.
shellverified commands
git clone https://github.com/ChazY11/medradar.git
cd medradar && pip install -e .

# offline: no network, no API key
python -m pytest -q -o addopts=""
medradar demo

# live radar run: network needed, no API key
medradar run --months 24 --with-trade --with-dashboard

# frozen market-to-R&D study
python experiments/reverse_rd_v2/run_v2.py   --db data/medradar.sqlite --holdouts 3