Decision-OS surfaces patterns in public FDA records — inspections, enforcement actions, Drug Master Files, Orange Book — that manufacturers' internal teams don't aggregate. Every finding traces to a primary source.
Decision-OS ingests directly from FDA public data: the Inspection Classification database, Warning Letters, the Drug Master File registry, the Orange Book, and enforcement records. No third-party aggregators.
Statistical and time-series methods applied across facilities, inspection histories, and enforcement timelines — surfacing patterns that only become visible when public records are aggregated systematically.
Every claim resolves to the primary record that supports it: source, retrieval timestamp, and content hash, anchored in a Merkle audit chain. Any finding can be independently verified against FDA's published data.
Regulatory findings are only useful if they hold up under scrutiny — from your own quality team, from procurement, from counsel. Decision-OS is built so every claim can be checked, not trusted.
Each cited record carries its primary-source URL, retrieval timestamp, a content hash of the source document, and its position in a sealed Merkle audit chain. If a finding cannot be traced to a primary FDA record, it does not appear in the report.
Open the URL, run sha256 on the returned PDF, and compare it to content_hash above — it matches, byte for byte. The document is a public filing in the National Stock Exchange of India archive.
Paste a content_hash from any Decision-OS report — or the example above — and query the public endpoint directly. A hash that is in no sealed tree returns a plain "not found", which is the answer that makes the others worth anything.
The prefilled example is a public stock-exchange disclosure, used solely to demonstrate live verification. It is not a finding. No Decision-OS analysis of any company appears on this site — the checker above returns provenance for a hash you already hold, and nothing can be browsed or enumerated through it. All Decision-OS analyses are delivered privately.
Reports are private and delivered 1:1. No company is named on this site. Below is the actual structure a report follows — its section spine, its inline source-tier attribution, and its provenance log — with every identifying value blanked.
This is generated from the report pipeline's own definitions, not mocked up: the sections, the tier markers, and the footnote format are the ones the generator emits. The blanks were never populated — there is nothing to recover from them.
DECISION-OS — PRIVATE REPORT (STRUCTURE SAMPLE, REDACTED)
report_reference_id: DOS-YYYY-XXXXXX-NNNNNN
Every identifying value is blanked. This sample is generated from the
live report pipeline's own section and citation definitions.
CONTENTS
Section 1: Facility Network and Regulatory Posture
Section 2: Current and Recent Enforcement Events
Section 3: Competitive Pipeline Analysis
Section 4: Supply Chain Exposure
Section 5: Financial Risk Quantification
Section 6: Historical Pattern Analysis
Section 7: What This Analysis Cannot Determine
Section 8: Platform Architecture and Future Capabilities
Section 9: Methodology Appendix
Section 10: Provenance Log
SAMPLE CLAIM LINES — every claim carries its source tier inline
FDA classified facility [FACILITY] (FEI [FEI]) as VAI following the inspection ending [DATE]. [CLM-S1-002][FDA-primary]
The issuer disclosed receipt of the [DATE] letter in a stock-exchange filing dated [DATE]. [CLM-S2-004][issuer-primary]
Facility [FACILITY] appears in the establishment registration file retrieved [DATE]. [CLM-S1-001][secondary-industry]
Inspection cadence at [FACILITY] is consistent with a routine surveillance interval. [CLM-S6-002][inferred]
SECTION 10: PROVENANCE LOG — one resolvable footnote per cited record
EV-DOS-YYYY-XXXXXX-NNNNNN-00347
FDA primary source: https://www.fda.gov/[...redacted...]
content_hash: [64-hex sha256 of the source, redacted] (sha256)
retrieved_at: [DATE]
Merkle proof: tree [redacted] leaf N/M root [redacted] [VERIFIED] sealed [DATE]
Every citation above is re-resolved live at generation time: the URL is
re-fetched, the content hash recomputed, and the Merkle proof re-checked.
A citation that fails is excluded; if a claim loses all of its supporting
records, the report does not generate.
Claims are descriptive record statements — what the record says, and when. Every citation is re-resolved live at generation time; one that no longer resolves is excluded, and a claim that loses all of its supporting records stops the report from generating at all.