How Defense Signals produces intelligence

We collect from open, attributable sources only. In practice that means four families:

• Primary government records — solicitations and award notices from SAM.gov, funding notices from Grants.gov, budget justification documents, and official service and agency releases.
• Company disclosures — earnings materials, filings, and official announcements from prime and mid-tier contractors.
• Association and event sources — published programmes and industry-day notices from named organisations such as NDIA, AUSA, AFCEA, and SAME, ingested into the Atlas™ mission calendar.
• Reported and on-the-record commentary — statements made publicly by officials and executives, always attributed to the person and the venue.

We do not use anonymous leaks of non-public government material, and we do not publish classified or export-controlled information.

Structured claims carry an explicit evidence status. The hierarchy is ordered, and it is the same hierarchy used internally by our data layer:

• MISSING — the field is unknown. We say so rather than estimating.
• INFERRED — derived from adjacent evidence by a documented rule, not asserted by any source.
• CLAIMED — asserted by a single identifiable source that has an interest in the claim.
• VERIFIED — confirmed by a primary record, or by two independent sources that do not share an interest.

A claim never silently upgrades. Promotion to VERIFIED requires the primary record or the second independent source to be recorded alongside it. Full detail is in our source and evidence standards.

Defense Signals separates what is known from what we conclude. Analytical judgments are written as assessments and use calibrated language:

Where an assessment depends on an assumption, the assumption is stated in the text rather than left implicit.

Every Signal opens with an Executive Brief™ so a reader gets the decision-relevant content in the first screen. The structure is fixed:

1. What happened — the verifiable event, dated.
2. Why it matters — the market or mission consequence.
3. Our assessment — the Defense Signals judgment, with confidence stated.
4. What to watch — the next observable indicator that would confirm or break the assessment.

We use large language models for defined, supervised tasks: drafting candidate summaries from material we have already collected, generating embeddings that power semantic search, and retrieval-augmented answering in Ask Signal™, which answers only from Defense Signals content and links the underlying pieces.

AI does not decide what we publish. No AI-generated text reaches a published Signal without a human editor reviewing it against the source material, and the editor named on the piece is accountable for it. AI is never used to fabricate quotes, figures, sources, or citations.

Where Defense Signals ranks or scores something — event significance in Atlas™, GridLink™ relationship scoring — the score is produced by a deterministic, versioned formula over recorded inputs, not by a model's opinion. Scores are reproducible: the same inputs and the same configuration version always produce the same score.

When we get something wrong we correct it in place, date the correction, and describe what changed. The process is set out in our editorial and corrections policy.