Project Phoenix · Domain Demo
PPR Agent queries government-mandated cardiac device performance reports from Abbott, Boston Scientific, and Medtronic. 92 million US implants across 18 years. The model selects a tool and formats the answer. The database supplies the facts. Swap the model, rewrite the interface — the numbers don't change. The invariants live in the substrate, not the code.
The data source
What the data is
Abbott, Boston Scientific, and Medtronic are required to publish annual PPRs disclosing cumulative US implant counts, model numbers, and device family classifications. These are not estimates. They are manufacturer disclosures under regulatory obligation.
Why this matters
Every implant count in the database came from a PPR. The model's job is to
parse intent, select a tool, and format a cited answer. It cannot guess a count,
invent a device name, or improve an extraction. If the data says
device_family: Unknown, the answer says Unknown.
The architecture
PPR data extracted from manufacturer PDFs, validated against gold-standard CSVs, and stored in a normalized SQLite database. Every answer traces to a row. No model involvement at this layer. Rewrite every layer above it — this layer does not move.
Each tool has an exact input schema, an exact output schema, and a correctness test. Tools query the substrate directly. They return JSON. They do not summarize, infer, or fill gaps.
The model parses intent, selects a tool, provides arguments, and formats the cited answer. It is not permitted to answer from memory, guess counts, or bypass tool retrieval. Two interface paths exist: deterministic CLI (no model) and local Ollama-backed shell (no external API key).
Canonical demo — Query 001
What were the top 5 pacemaker devices by US implant volume in 2025?
Natural language. No schema knowledge required from the user.
get_top_devices
Parameters: year=2025, device_category="Pacemaker", limit=5
Provenance: devices table · filter year+category · order implants DESC
| Rank | Company | Device Family | Model | US Implants |
|---|---|---|---|---|
| 1 | MDT | Azure XT DR | W1DR01 |
878,312 |
| 2 | ABT | Unknown | PM2272 |
500,885 |
| 3 | MDT | Adapta DR | ADDR01 |
454,889 |
| 4 | MDT | Advisa DR MRI | A2DR01 |
344,448 |
| 5 | BSX | ACCOLADE/PROPONENT/ESSENTIO DR | L101/L111/L201/L211/L301/L311 |
330,000 |
Row 2 device family is Unknown — an unresolved extraction gap in the source PDF. The system reports the substrate exactly. It does not substitute or infer a name.
Canonical suite
| Query | Tool | CLI | Ollama |
|---|---|---|---|
| Top 5 pacemakers by implant volume, 2025 | get_top_devices |
✓ | ✓ |
| Compare Medtronic and Boston Scientific, 2020–2024 | compare_companies |
✓ | ✓ |
| Market share trend, all companies, 2020–2024 | get_market_share_trend |
✓ | ✓ |
| Medtronic year-over-year growth, 2020–2024 | get_year_over_year_growth |
✓ | ✓ |
| Device families with 10+ years in market | get_device_longevity |
✓ | ✓ |
| Search for Quadra devices | search_devices |
✓ | ✓ |
CLI path: deterministic pattern matching, no model. Ollama path: local
qwen2.5:14b, no external API key. Both paths hit the same SQLite
substrate and return identical tool results.
The data
These charts render frozen canonical query results — exact rows from the deterministic tool layer. No model generated these numbers.
US Market Share by Company · 2014–2025 · get_market_share_trend
★ 2017: Abbott completed its $25B acquisition of St. Jude Medical in January 2017, consolidating SJM's cumulative implant counts under ABT. ABT share jumped +6 pp; MDT share dropped −9 pp. Both reflect the reclassification, not a change in device performance. The 2024 MDT dip reflects official retirement of legacy device models from the PPR report — visible in the data, sourced to the substrate.
Top 5 Pacemakers by US Implant Volume · 2025 · get_top_devices
Rank 2 device family is Unknown — an unresolved extraction gap in the source PDF. The substrate reports it. The chart renders it. The model does not substitute a name.
The boundary
Allowed
The model reads the question, decides which tool fits, provides the parameters, and formats the returned rows into a readable cited answer. This is the entire scope of its authority.
Not allowed
If the tool returns empty, the answer is "No data found." If a device family is Unknown in the substrate, it is Unknown in the answer. The model has no override on what the data says.
Related
The open-core framework behind this demo. Standards for grounded domain systems: deterministic substrate, explicit validation, clear trust boundaries.
The harness boundary finding that preceded this demo. Capture transport is part of the measurement result. When the apparatus is not neutral, the score cannot be interpreted as a model-only signal.