The CMI framework
PROWLER treats the market as a stream of observable events: a contract notice published, an approach to market opened, a ministerial announcement, a partnership confirmed in a public newsroom. Each event is captured, enriched into a structured signal, and mapped against the programmes and capability priorities that organise Australian defence investment.
The output is not a verdict. It is a structured, queryable record of activity that an analyst reads in context. The dashboard surfaces what is happening; the judgement about what it means stays with the person, not the model.
Why these sources
Sources sit in four tiers. Official Commonwealth records come first: AusTender contract notices and approaches to market, and Defence portfolio ministerial releases. These are the primary record of what the customer is buying and saying.
Industry media adds context and lead time: Australian Defence Magazine, Defence Connect, Asia Pacific Defence Reporter, the ASPI Strategist, and USNI News filtered for AUKUS and Australia. Competitor public newsrooms give organisation-level posture. Policy documents, the National Defence Strategy and the Integrated Investment Program, are the frame that every other signal maps back to.
How relevance is calibrated
Every raw signal is scored from zero to one for relevance to Australian defence market intelligence. Signals below a low threshold are skipped rather than enriched, which keeps the corpus and the enrichment spend honest. Duplicates are caught on a content hash before they reach the model.
The score weighs an Australian defence nexus, whether a known programme, capability, or tracked organisation is named, and the materiality of the event. The score is advisory. It orders the queue and powers the filters; it does not make decisions, and an analyst can always read past it.
OPSEC boundaries
This is open-source intelligence at the level of organisations, not people. There is no person-level collection, no LinkedIn scraping, and nothing behind authentication. Framing stays neutral: the dashboard reports observable activity and does not assert who is winning or losing, or infer commercial outcomes that are not public.
These limits are deliberate. A market-intelligence function inside a prime has to be demonstrably clean on ethics and on law, and the discipline of staying organisation-level and open-source is what makes the output safe to act on and safe to share.
Model routing
Enrichment runs on Ollama Cloud with cost discipline as a design goal. A small, fast model handles first-pass classification and de-duplication. A larger model handles summarisation and the extraction of organisations, programmes, capabilities, and domain. An embedding model writes each enriched signal into the vector index that powers semantic search.
The enrichment prompt is fixed: output structured JSON only, Australian English, and no speculative judgements about commercial outcomes. Routing the cheap work to a small model and reserving the larger model for genuine extraction keeps spend low without losing quality where it matters.
Real-time bid intelligence inside Thales
This is described, not demonstrated, because it needs non-public data that does not belong in an open build. Inside the company, the open-source layer shown here would be one input among several. It would be fused with the pipeline and capture plans, teaming and pricing history, win and loss records, and the customer relationships that never appear in a media release.
The same event model then becomes early warning: a recompete approaching, a partner positioning against an opportunity, a capability gap opening against an Integrated Investment Program priority. That fusion is the real value of a CMI function, and it is exactly the line this public demonstration does not cross.