Machines read widely. The bar decides what ships.

What is watched

Prognyx reads the primary record, not the coverage of it. The standing sources are:

• ClinicalTrials.gov — registrations, status changes, enrolment, arms, endpoints, termination reasons and completion dates.
• FDA and EMA — approvals, designations, complete response letters, label changes, advisory-committee calendars, PDUFA goal dates.
• SEC EDGAR — 8-Ks, 10-Ks and exhibit press releases, which is where a US-listed company's real numbers land first and where the language is legally constrained.
• Europe PMC and the peer-reviewed literature — the published data behind a claim, with the DOI.
• Congress abstracts — ASCO, ESMO, ASH, AACR: often the first place a competitor's data becomes public.
• Trade press — used only to know that something happened and to find the primary document. Press reporting is never the evidence for a factual claim; it is treated as unconfirmed until a primary record lands.

The two rules that throw work away

1. One primary source per fact, or it doesn't ship

Every factual sentence in a Prognyx briefing traces to a document you can open. Not a summary of a document — the document. When a number exists in the market but not in a primary record, the briefing says so explicitly: "this cannot be confirmed from public sources."

That sentence is the most valuable one in the deliverable, and it is the sentence a data platform will never write, because a platform's job is to look complete. Publishing what you do not know is the only way a reader can calibrate what you do.

2. The generator never grades its own homework

The analysis is written first, then handed to a separate audit pass whose default posture is disbelief. It re-opens each cited source and asks one question per claim: does the document actually say this? Claims that fail are corrected or cut. If the audit and the draft cannot be reconciled within two rounds, the briefing escalates to me for manual review rather than shipping on a tie.

The freshness gate

Competitive intelligence has a failure mode that ordinary fact-checking misses entirely: a claim that was true when the source was written and is false today. A trial listed as recruiting was terminated last week. A PDUFA date moved. A registry record lags the company's own 8-K by months.

So before publication, every cited trial is re-verified against the live registry and cross-checked against the press record at that moment. A briefing that cites a trial whose status changed underneath it does not publish; it goes back. This gate exists because it failed once — an earlier version of the pipeline shipped a stale trial status, which is exactly the kind of error that destroys a research provider's credibility in one email.

Where the machines are, precisely

Prognyx runs on language models, and hiding that would be a worse position than explaining it. Here is the line, exactly where it sits — including the part most vendors leave out, which is what publishes on its own:

So the honest summary is this: the blog is the engine's output, published under a rule and a threshold, and the threshold is the only thing standing between a draft and this page. What a client buys is not that — a client brief is generated the same way and then goes out only when a person sends it.

The distinction that matters to a buyer is not "did software touch this" — software touches everything now, including at every enterprise platform. It is what rule the software had to obey, and who answers when a claim is wrong. The rule is published above. The person is Rali Filali, and there is no committee to absorb the error.

What Prognyx will not do

• Won't invent a number to complete a table. A gap is reported as a gap.
• Won't launder press reporting into fact. Secondary coverage is flagged as unconfirmed until a primary record lands, even when it is almost certainly right.
• Won't dress a post-hoc subgroup as a result. A positive subgroup inside a failed intent-to-treat analysis is the oldest hopeful story in oncology, and it gets labelled as what it is.
• Won't give medical or investment advice. Prognyx analyses public information about competitive positioning.
• Won't quietly revise a published briefing. Briefings are dated artefacts. Errors get a visible correction; developments get a new briefing that supersedes the old one.

What you actually receive

Every briefing follows the same structure, because it maps onto how a decision gets made rather than how a news story gets told:

1. What happened — the event, sourced.
2. The number that matters — including, when relevant, the number that has not been disclosed.
3. Why it matters — the mechanism of the threat or the opening.
4. Who is exposed, who gained — named, with the exposure quantified where public data allows.
5. What to watch, with dates — the calendar, so the read stays useful next month.
6. The now-what — the two or three moves actually on the table for you.

Read five of them on The Watchtower before you talk to anyone here. They are the same briefings clients receive, on public assets instead of yours.