Payments Intelligence
Changelog
How the product has evolved, milestone by milestone. Dates are when each capability was completed; some were switched on in production a little later. Live figures are on System Metrics, never here.
How it works today
One request, end to end. Every stage below is part of the running system.
Hover a stage for what it does. Method and limitations are on the Methodology page.
Milestones
From first pipeline to controlled learning
13 milestones since 28 Aug 2026, oldest first.
Intelligence platform foundation
Automated payments research that publishes only what the evidence supports.
- Automated research and publication pipeline for the payments industry
- Source-backed, evidence-scored intelligence briefs behind a strict publication gate
- Conservative story deduplication so the same story is never published twice
- System metrics dashboard and the merchant payment performance analyzer
Monitoring and product experience
Scheduled research became observable, and the product got its own identity.
- Execution checkups for the daily monitoring run
- Redesigned product experience across the intelligence feed and brief pages
Verified market facts and production hardening
A verified-facts layer, a hardened public surface and production health checks.
- Verified market facts foundation: numeric claims checked against their sources
- Security hardening of the public web surface
- Privacy, legal and launch-readiness foundation
- Automated production health checks
- Refined motion and art direction across the site
Autonomous research architecture
The architecture for a new research engine that plans, retrieves and writes on its own.
- Autonomous research, intelligence generation, distribution and operations foundations, designed and tested
- Command-center redesign of the product overview
Shadow evaluation in production
The new engine started running alongside the established pipeline, observing without publishing.
- Shadow evaluation of every daily monitoring run, one comparison each
- Bounded read windows and safe handling of evidence that cannot be reconstructed
- Production workflow baseline brought under version control
Primary research engine authors briefs end to end
The new engine researched, analysed and wrote its first published brief from scratch.
- End-to-end research: objective, retrieval, evidence ranking, structured analysis and drafting
- Escalation search that looks specifically for primary and official sources
- Explicit insufficient-evidence contract with named withhold reasons
- Numeric validation of drafted claims against the evidence
Controlled production evaluation
The primary engine began publishing on a schedule under tight, durable limits.
- Scheduled evaluation with deterministic topic rotation
- Durable, race-safe daily publication limit enforced in the database
- Accurate attribution of how every run was started
Resilient primary architecture
Requests go to the primary research engine first, with an automatic, tracked fallback.
- Primary research engine with an automatic resilient fallback
- Authenticated ingress with request validation
- Atomic daily request-budget protection
- Correlated execution tracking: one request id across both engines
- Engine comparison metrics and operator health signals
Controlled learning system
The system started learning from its own history — observe-only, with a human in the loop.
- Experience extraction from completed research runs
- Transparent, decomposable source-reputation memory
- Failure-pattern detection
- Offline evaluation of improvement candidates against real history
- Human-gated promotion: nothing changes production automatically
Scale and impact metrics
A reproducible, historical view of what the autonomous system has actually done.
- Historical metrics reconstructed from the raw run records
- Source, autonomy and fallback telemetry with request-level deduplication
- Offline evaluation volume counted from real replays
- Every figure classified as observed, derived or estimated, with a public methodology
Primary engine capacity
Authenticated requests separated from the limit that governs controlled evaluation.
- Production requests separated from the controlled-evaluation daily limit
- The daily request budget remains the traffic and cost guard
Metrics and learning experience
A new metrics experience and a visual walkthrough of how the system learns.
- Redesigned system metrics experience
- Learning-process visualization and candidate tables
- Latest
Unified product presentation
One product, one vocabulary — and an honest model for measuring analyst time saved.
- Unified Payments Intelligence terminology across the site
- Public changelog
- Benchmark model for analyst hours automated, ready for a measured baseline
- Mobile layout improvements on brief pages