Dispute queue
p(win)
⚡ Live triage — Razorpay dispute webhook
🟢 Connected to your Razorpay account (test mode) ·
Razorpay fires a dispute.created webhook → RokdaDaav
auto-triages each one (no human) and prepares a submit-ready contest packet in
Razorpay's format. (dispute events simulated — test mode has no dispute API; connection & submission are real)
·
Click “Simulate incoming
Razorpay disputes” to watch them stream in and get handled live.
⏳ receiving dispute.created events from Razorpay…
Approve & send queue — FIGHT cases
with a drafted letter (human-in-the-loop)
·
· EV
🕵️ Abuse-ring detective (defense-only)
Clusters disputes by shared signature — device fingerprint,
card BIN, shipping city — and flags coordinated friendly-fraud rings. Validated
on planted rings (a known-answer eval).
Scanned disputes ·
rings detected ·
precision ·
recall ·
F1
Recall under 100% is honest: one
stealth ring (residential IPs, slow burn) evaded the velocity/device
signals — a documented limitation.
⏳ clustering the stream and assessing each cluster with AI…
Click “Scan the dispute stream” to detect coordinated abuse rings.
Ask RokdaDaav — live AI agent (calls the tools itself)
The AI decides which tools to call — you watch it happen. Try:
“Should I fight DSP001561?” · “Is DSP000864 worth fighting?” · “Prove RokdaDaav beats the baselines.”
🔧 called ()
→
⏳ the AI is deciding which tools to call…
Select a dispute
from the queue to see its evidence, decision, and letter.
from the queue to see its evidence, decision, and letter.
calibrated p(win)
escalated → human (base: )
agent-initiated
· · filed days after txn
true outcome (held-out): — shown to illustrate calibration; never a model input
Decision — expected-value arithmetic (₹)
| EV fight = p(win)·amount − cost_to_fight | |
| EV accept | |
| EV refund = −amount + ratio_benefit |
cost_to_fight =
· ratio_benefit =
· chosen EV =
Uncertain & expensive: ranked in the top 10% by uncertainty×amount, so it is routed to a human with evidence pre-assembled rather than auto-decided.
💡 How to win this — evidence-gap advisor
The model re-scores this case under counterfactual evidence
to find the cheapest documents that would make it winnable.
⏳ running counterfactual re-scoring…
Current:
win ·
Documents you can still gather (ranked by impact)
—
+ pts →
All together →
win
Evidence agent — fixed 8-step sequence
no artifact on file for this step
Rebuttal letter — the document RokdaDaav would submit
Representment — Chargeback Rebuttal
To the issuing bank / card network,
We respectfully contest this chargeback on the following documented evidence:
On the strength of the evidence above, we request that the chargeback be reversed and the transaction upheld.
Sincerely,
Merchant Risk Operations · RokdaDaav
Audit trail — of
AI-drafted claims kept;
the verifier stripped the rest. Click a line to see its source evidence.
✓
✗