Payments Intelligence
Learning System
What the controlled self-learning system has observed since the last cycle. This is OBSERVE-ONLY: every candidate below is a proposal, offline-evaluated against historical data — nothing here changes production automatically.
How it learns
One closed, observe-only loop. No candidate can reach production on its own — promotion is always a separate, explicit, human decision.
Hover a stage for what it does. Every stage names a real step in the code — nothing here is illustrative.
Learning overview
Experiences processed, candidates proposed
Last updated 27 Sept 2026.
Candidates
Proposed strategy improvements
Each candidate is a deterministic proposal, backed by a minimum evidence count, then offline-evaluated against real historical data with named metric deltas — never one opaque “better” score. PASSED_OFFLINE is the ceiling this system can reach on its own; promotion to production is always a separate, manual, human decision.
| Scope | Proposed change | Evidence | Expected effect | Offline result | Status |
|---|---|---|---|---|---|
accountingtoday.com Source deprioritization | consistently contributes duplicate coverage rather than new evidence | 36 | duplicate_rate ↓(+0.649) | n=36 | FAILED_OFFLINE |
openpr.com Source deprioritization | consistently contributes duplicate coverage rather than new evidence | 36 | duplicate_rate ↓(+0.649) | n=36 | FAILED_OFFLINE |
ecb.europa.eu Source deprioritization | consistently contributes duplicate coverage rather than new evidence | 348 | duplicate_rate ↓(-0.098) | n=348 | PASSED_OFFLINE |
fintechmagazine.com Source deprioritization | consistently contributes duplicate coverage rather than new evidence | 144 | duplicate_rate ↓(+0.107) | n=144 | FAILED_OFFLINE |
reuters.com Source deprioritization | consistently contributes duplicate coverage rather than new evidence | 36 | duplicate_rate ↓(+0.649) | n=36 | FAILED_OFFLINE |
engadget.com Source deprioritization | consistently contributes duplicate coverage rather than new evidence | 36 | duplicate_rate ↓(+0.649) | n=36 | FAILED_OFFLINE |
Failed offline is the safety gate holding a candidate back deliberately, not a broken run.
Source quality
Transparent, decomposable source-reputation memory
Every classification below is re-derivable from the counters in its own row — never an opaque, model-generated trust score. A source needs at least 3 observations before it gets anything other than “insufficient history”; a single run never blacklists a domain.
| Domain | Classification | Primary history | Publish history | Evidence count |
|---|---|---|---|---|
| ecb.europa.eu | DUPLICATE_PRONE | 348 / 348 | 240 | 348 |
| cnbc.com | CONSISTENT_SECONDARY | 0 / 180 | 144 | 180 |
| thenextweb.com | CONSISTENT_SECONDARY | 0 / 180 | 144 | 180 |
| fintechmagazine.com | DUPLICATE_PRONE | 0 / 144 | 72 | 144 |
| adyen.com | HIGH_PRIMARY_HISTORY | 72 / 72 | 72 | 72 |
| thefintechtimes.com | CONSISTENT_SECONDARY | 0 / 72 | 72 | 72 |
| accountingtoday.com | DUPLICATE_PRONE | 0 / 36 | 0 | 36 |
| engadget.com | DUPLICATE_PRONE | 0 / 36 | 0 | 36 |
| openpr.com | DUPLICATE_PRONE | 0 / 36 | 0 | 36 |
| talent.ecb.europa.eu | HIGH_PRIMARY_HISTORY | 36 / 36 | 36 | 36 |
| ft.com | CONSISTENT_SECONDARY | 0 / 36 | 36 | 36 |
| thepaypers.com | CONSISTENT_SECONDARY | 0 / 36 | 36 | 36 |
| reuters.com | DUPLICATE_PRONE | 0 / 36 | 0 | 36 |
Safety contract
No candidate can reach production on its own — promotion is always a separate, explicit, human decision, enforced at both the code and database level. See the loop at the top of this page for the exact stages every candidate passes through.
See system-wide metrics