Capability Brief · Unclassified · Public Data Only

ORDN Engine

An owned, append-only intelligence data engine for the defense and ordnance domain. It collects the public record while it is still collectable, carries provenance on every row, and is built so the dataset is still usable in five years.

ordn-engine · self-hosted

Open sources only · Rights and licence recorded per source · Nothing duty-derived registered as a source

The Problem

Parts of the public record expire daily

Sanctions and screening lists serve what is listed today. A timeline of designations and delistings has to be assembled from dated captures, and a claim built on it has to show whose capture it was, when it was taken, and how it was parsed.
Analysis built on rented platforms is borrowed. When the subscription ends or the provider changes terms, the dataset and the workflow go with it.
Claims without provenance die under scrutiny. An assessment is only as strong as the ability to show where each fact came from and when it was captured.
Sources rot silently. Feeds move, endpoints break, formats drift. A pipeline that does not measure its own freshness reports stale data as current.

51,729

designation events in the store, count as of 4 August 2026

0

external dependencies: Python standard library only

8

active sources, each with licence and rights recorded, plus one registered as pending until its terms are read

100%

of records carry provenance back to a collection run

ORDN EngineUnclassified

The Capability

A spine that streams plug into

Source registry with rights built in. Every source carries its licence and two flags, commercial use and machine-learning use, so downstream questions are answered before they are asked. United States public domain, United Kingdom Open Government Licence, United Nations public releases.
Freshness is measured, not assumed. Each source has a staleness threshold; the status board flags any feed past it. A killed run is recorded as abandoned, because a killed run must not read as a clean one.
Append-only store. Nothing is overwritten. Re-collecting the same data is idempotent and verified by test; history accumulates, it never mutates.
Gaps are registered, never omitted. A source that cannot be collected yet is recorded as pending with the reason. A gap that is written down is not a gap that was forgotten.
No dependencies to rot. The entire engine is standard-library Python, so there is no dependency tree to rot.

init

create store, apply schema, register sources

collect

run all active collectors, one command

status

heartbeat, freshness, rights, store counts

delta

what changed, per list, per day

ORDN EngineUnclassified

Status Board

The engine reports on itself

SOURCES ofac_sdn active US-PUBLIC-DOMAIN commercial/ml-ok ofac_cons active US-PUBLIC-DOMAIN commercial/ml-ok trade_csl active US-PUBLIC-DOMAIN commercial/ml-ok uk_ofsi active UK-OGL-3 commercial/ml-ok un_cons active UN-PUBLIC commercial/ml-ok fedreg active US-PUBLIC-DOMAIN commercial/ml-ok eu_fsd pending PENDING-REVIEW NON-COMMERCIAL/NO-ML FRESHNESS ofac_sdn last ok 144h ago <-- STALE (threshold 72h) ofac_cons last ok 144h ago <-- STALE (threshold 96h) trade_csl last ok 144h ago <-- STALE (threshold 72h) uk_ofsi last ok 144h ago <-- STALE (threshold 96h) ABANDONED RUNS (opened, never closed) 2026-08-02T08:47:56 uk_ofsi open for 145h STORE collection_run 22 provenance 28 designation_event 51,729 RECENT RUNS 2026-08-04 fedreg ok http=200 n= 133 +0 ~0 -0 2026-08-02 trade_csl ok http=200 n=25,921 +0 ~0 -0

Representative status output, condensed. Layout and fields as printed by the status command; store counts from a local copy of the store as of 4 August 2026; two active sources omitted from this view.

Staleness shows itself. A source past its threshold is flagged, and a run that was opened and never closed is listed as abandoned. The engine does not present a quiet source as a current one.
Every run logs its result: what was fetched, how many records, and what changed.
The pending row is deliberate. The European Union feed stays flagged non-commercial until its terms are reviewed, and its data stays out of anything sold or trained on.

ORDN EngineUnclassified

First Stream

Designations first, to prove the spine on clean data

What this store adds is its own record of observation. Sanctions history is available from other publishers. This store records its own capture time on every row, keeps the raw capture, and can replay every capture after a parsing rule is found wrong, so a timeline built from it can be audited and re-derived.
Five lists, one schema: the two United States Treasury lists, the consolidated screening list, the United Kingdom consolidated targets, and the United Nations Security Council list.
Clean rights on purpose. The first stream was chosen partly because its licensing is unambiguous, the right place to prove a schema before harder sources.
Delta as a first-class product. One command answers what changed, per list, per day, which is the question every downstream assessment starts with.

51,729

designation events in the store, count as of 4 August 2026

5

lists active in the stream

15/15

engine tests passing, idempotent re-ingest verified

1

request per daily snapshot on the bulk surface: no keys, no pagination, no rate limit

ORDN EngineUnclassified

Growing Streams

The spine already feeds wider coverage

Federal Register notices, the regulatory half of the defense-industrial picture.
Global port and trade activity from public multilateral monitoring data.
Regional defense ministry releases from open government publications in the Pacific.
One source retired, on the record. A contract-announcement feed was retired after a reporting threshold was misparsed as an award value. Its rows were superseded rather than deleted, the raw captures were kept, and the store was rebuilt from them.
Planned analytic streams build on the same spine: drone warfare, emerging defense technology, conflict economics, and arms transfers, each as a versioned, provenance-carrying dataset.

1

source retired after a misparse, rows superseded, raw captures kept

5

analytic streams planned on the spine

daily

collection cadence on the perishable feeds

versioned

every dataset traceable to source, run, and day

ORDN EngineUnclassified

Collection Discipline

Silence is reported as silence

The honesty rule is structural. A quiet receiver is reported as a receiver gap, never as an empty sea. Zero signal is a statement about the sensor, not the water. Search-and-rescue data is treated as the authoritative check.
Scheduled agents, visible heartbeats. Collection jobs run on timers and report status; an idle agent between firings is normal and shown as such, so silence is never mistaken for failure or success.
Products follow the same standard as the engine: dated, sourced, and labeled as intelligence assessments with the confidence of each claim marked.

gap

a quiet receiver is logged as a receiver gap, not as an empty area

idle

a timer job between firings is shown as idle, not as failed

dated

every product carries its date, its sources, and its confidence marks

rule

no duty-derived or non-public material is registered as a source

ORDN EngineUnclassified

Principles and Status

The record is the asset

Deltas and documented timelines rather than commentary. Analysis is downstream of a dataset that can be re-derived and audited.
Public data only, rights first. Every source's licence is recorded before its data is used; anything unreviewed is fenced off from commercial and model-training use automatically.
Strict separation. The engine touches nothing from military duty: open publications, public feeds, and licensed data only.
Status. Spine built and tested. Designation stream collecting daily. Regulatory and port streams collecting. One source retired and documented. Analytic streams staged on top.

Built

spine: registry, heartbeat, provenance, freshness

Collecting

designations, regulatory, port activity

Scheduled

daily collection with visible heartbeats

Staged

five analytic streams on the same spine

ORDN EngineUnclassified