The thinking behind the platform, in public. Where enterprise AI is heading, what we are learning from real deployments, and the research that shapes what we build.
Every benchmark ships with the corpus, the setup, and the parts we did not test, so the result can be checked rather than taken on faith. We refresh these as models and datasets move.
Fixed three-model catalog, comparator configured by an adversarial reviewer, every answer graded blind.
74 files in production shape, vs a leading data-security platform. Coverage diverges where the file type hides the data.
PII detection & classification across prompt and output surfaces, on a mixed India & ANZ dataset.
Recall on multi-turn jailbreak campaigns, measured against single-prompt classifiers. Publishing with the attack set.
How often plan-level checks stop a poisoned instruction before it executes, versus content filtering on the output.
Compute and latency per million documents as the cascade routes cases, benchmarked against a single frontier pass.
On a fixed three-model catalog, with the comparator configured by an independent adversarial reviewer and every answer graded blind, Cyvia won all three lanes enterprises buy on. Preregistered, hash-traceable, and reproducible on your own approved catalog.
Fixed three-model catalog, so availability and price are held equal and only decision quality remains. The comparator was configured by an independent adversarial reviewer from public documentation, given no Cyvia material. The judge saw no system names, model identities, costs, or latency.
Policy fidelity (25/40 vs 19/40) is a consistent but directional win, expected to widen when the semantic matcher ships. Outside-window notices trail at 8/15 vs 10/15, a scoped gap. The cost figure is weighted token units labelled to Round 4b; it is reported as a consumption measure, never as dollars.
Margins of six answers or more over 40 scenarios are decisive; sensitivity and sufficiency clear that bar. Cost basis: 403,608 vs 1,654,818 weighted token units per 1,000 sufficient answers, Round 4b, fixed catalog.
Detection and classification across prompt and output surfaces on a mixed India and ANZ dataset, scored against a leading platform. The cascade reads the messy, local entities other tools skip, and does it faster.
74 files in production shape, four classifiers, scored on unique matches. Coverage was close on most formats and diverged sharply where the file type hides the data. We take Data Lens output as ground truth here, so read this as coverage overlap, not a scored win, the numbers are what a customer can reproduce on their own estate.
Scored on unique matches, the conservative comparison; occurrence totals widen the gap further. Driving licence omitted as negligible (3 hits in the set). GSTIN and Aadhaar rows in amber mark the format blind spots.
Three benchmarks live, more in flight. We publish each with its dataset and its limits as it lands.
View all benchmarks →Why a four-layer cascade beats brute-forcing every document through a frontier model, on both accuracy and cost.
A single prompt looks benign. Read the lineage across a session and the campaign gives itself away.
A tainted document, a helpful agent, and the plan-level check that caught the exfiltration before it ran.
New research, benchmarks, and field notes, in your inbox when they land.
✓ Thanks — you're on the list.