A deliberately messy record fixture measured useful answers, wrong releases, refusals, and conflict handling. No external LLM was used.
The comparison baseline was deliberately naive and always answered. Its failure rate is an unsupported-commitment baseline, not a measured universal LLM hallucination rate.
Terminal CSV, switch JSONL, settlement rows, and operations logs were joined by explicit transaction evidence.
This is not a PhonePe workload, payment-network certification, hardware acceptance test, or customer result.
The site and Worker contract are tested separately from the private rule implementation.
A pass establishes only the named boundary.
| Claim | Measured behavior | Does not prove |
|---|---|---|
| Exact-anchor isolation | The requested transaction excludes neighboring entities. | Broad semantic retrieval quality. |
| Conflict withholding | Required identity fields disagree and no owner is chosen. | Resolution of the source conflict. |
| Instruction quarantine | Instruction-like evidence fields are withheld. | Coverage against every prompt attack. |
| Proof-preserving reduction | Selected evidence retains identifiers and provenance. | One universal token-reduction percentage. |
| Point-in-time scope | Evidence after an explicit cutoff is excluded. | Complete event-sourcing semantics. |
These require customer data, external infrastructure, or independent review.
Measure against source-of-truth labels from a real workflow.
INDEPENDENTLY UNPROVENRun multi-user soak, backup, restore, failover, and deletion drills.
NO PRODUCTION SLAComplete penetration testing, managed-key review, and customer identity integration.
NOT CERTIFIEDMeasure prevented wrong releases, review time, and workflow cost.
NEEDS DESIGN PARTNERCompare wrong joins, unsupported answers, retained answer coverage, latency, and context volume.