01
Expected range
93.9% expectedBrazil card approval on NovaPay fell from 93.9% expected to 62.4% across four consecutive windows.
ev-br-baselineA signal is not a diagnosis.
PHAROS puts the comparison beside the claim, so a person can verify the reasoning before deciding what to do.
02
Counterfactual
AuroraPay · 94.1%AuroraPay held 94.1% approval for the same Brazil card traffic and issuer mix.
ev-br-control03
Decline pattern
do_not_honor · 71%do_not_honor increased to 71% of observed declines, up from 18% in the baseline.
ev-br-declinesSee the investigation, not a black box.
One incident, one evidence trail, one human decision. The system narrows the scope; it does not make the change.
Worked example / Brazil · NovaPayProbable · human review required
Anomaly detected62.4%
Brazil card approval on NovaPay
−31.4 pp below 93.9% expected · four sustained windowsBest-supported explanation
NovaPay × Brazil × Card × Itaú
Brazil card approvals degraded on NovaPay while a provider control with the same issuer mix stayed healthy.
Evidence support0.91
Evidence board
ev-br-baselineBrazil card approval on NovaPay fell from 93.9% expected to 62.4% across four consecutive windows.
ev-br-controlAuroraPay held 94.1% approval for the same Brazil card traffic and issuer mix.
ev-br-declinesdo_not_honor increased to 71% of observed declines, up from 18% in the baseline.
Recommendation only — no traffic was rerouted
Ask NovaPay to investigate the Brazil card decline-code spike and validate an issuer-side response with Itaú.
The chain ends at a person.
A confirmed drop opens an investigation, and the investigation reaches whoever runs the business. PHAROS never takes the next step itself — no traffic is rerouted, no provider is switched, nothing is executed.
What it observes
Live payment streamEvery authorisation, as it happens
Learned baselineBeta-Binomial, per segment
Counterfactual controlThe healthy twin that rules causes out
Decline-code mixHow the rejections shift
PHAROS
Detects · investigates · cites evidence
How it notifies
In-appAlways
EmailHigh & critical
WhatsAppHigh & critical
Business owner
Decides — nothing is executed
Email and WhatsApp report as send accepted once the provider answers. Delivery is only claimed when a provider confirms it.
We don't know the answer either.
A scenario is injected with its dimensions sealed from the investigator and from this interface. The detector only ever sees the payment stream. Here is one recorded run — and the console that lets you set up your own.
Recorded run · 30 Aug 2026
chaos_8166f08054Ground truth injected
merchant = VuelaYa−25 pp degradation · sealed until revealWhat the system reported
merchant = VuelaYaProbable · human review required- Time to report
- 3 min 32 s
- Confirmed after
- 3 sustained windows
- Incident
inc_anom_7bc545b7
The trial is bound to an episode before the reveal, so no report can be chosen after the fact for resembling the answer. The Chaos Lab scores each run against the truth it was never shown.
PHAROS / ready for reviewDetect the drop.
Detect the drop.
Prove the cause.
Give payment teams a defensible investigation before they make a human decision.