01
Discover
Tool-using research agents review regulators, companies, specialist feeds, search, and original reporting.
- Produces
- Candidate evidence
- Blocks on
- Inaccessible or generic source
Editorial system · v2026.10
Autonomous by default.Auditable by design.
A multi-stage research and editorial system discovers payments events, resolves competing reports, selects canonical evidence, and releases only what passes explicit controls.
Operating model
Five bounded stages.
Agents are allowed to explore, retrieve, compare, and draft. They are not allowed to redefine evidence, silently repair conflicts, or bypass release contracts.
02
Resolve
Reports are grouped into underlying events so five rewrites do not become five stories.
- Produces
- Event clusters
- Blocks on
- Duplicate or stale event
03
Rank
Source authority, originality, specificity, and factual agreement determine the canonical evidence.
- Produces
- Source decision
- Blocks on
- Material source conflict
04
Compose
Specialised writing and editing agents turn verified events into concise payments intelligence.
- Produces
- Source-bound copy
- Blocks on
- Unsupported implication
05
Release
Deterministic checks validate structure, source specificity, and publishability before distribution.
- Produces
- Release record
- Blocks on
- Blocking source or schema failure
The release contract
Fluency never outranks evidence.
- R1Every story resolves to a specific article or primary document.
- R2Canonical sources survive writing and editing unchanged.
- R3Candidates are checked against recent coverage before they reach the writer.
- R4Conflicting high-authority evidence blocks publication rather than forcing a winner.
- R5Unverified caveats, generic landing pages, and unsupported specificity do not ship.
Visible assurance
What readers should see.
- Direct source links and permanent issue URLs
- Clear distinctions between fact and interpretation
- Public editorial rules and correction channels
- Versioned improvements to the publication system
Protected advantage
What remains proprietary.
- Prompts, agent topology, and model-routing policy
- Source-discovery maps, rankings, weights, and thresholds
- Event fingerprints, evaluation sets, and regression fixtures
- Failure data, cost controls, and internal operating telemetry
Technology stance
Models change. The system compounds.
Frontier models are replaceable components. Durable advantage comes from payments-specific memory: entities, events, source performance, editorial failures, corrections, and tests accumulated across every edition.
When the system makes a mistake, the objective is not merely to repair the issue. The failure becomes a regression case so the same class of defect is harder to publish again.