CUSTOMERS · CASE STUDIES

Numbers, not adjectives.

Six programs, six readings. Attributed by company type. Specific on the work, the artifact, and the moment a decision moved — kept together on one record.

Talk to AI Labs Back to customers

STUDIES

Six readings

Each one a program, a workflow, a metric that actually moved.

OUTCOMES

Specific

The artifact, the number, and the audit window the buyer set.

ANONYMIZED

By type

Attributed by company type, not by name. Reference path on request.

ON THE RECORD · ANONYMIZED BY TYPE

The work, the artifact, the metric that moved.

Six programs, attributed by company type — not by name. Numbers are real, buyer-defined windows, customer-owned baselines. Deeper review available under a private reference path.

CASE 01 · POST-TRAINING

↳ EVIDENCE PACKET

A frontier AI lab

RLHF adjudication that compounds, instead of resetting.

Specialist review, regression capture, and rubric versioning stayed attached to the same packet across releases. Eval findings carried forward into the next gate without being re-stated, and override chains became inspectable instead of folkloric.

11,408

cases reviewed in one release

98.6%

regression pass rate

rubric versions in flight

Reading · aura trace

CASE 01 · EVIDENCE PACKET

A frontier AI lab · reading kept private under reference path.

CASE 02 · RELEASE GOVERNANCE

↳ SIGNED PROOF

A regulated decisioning program

Release review became a repeatable gate.

Sign-off moved from a screenshot debate to one inspectable record per run. Risk, engineering, and change-management teams worked from the same approval surface, and examiner questions earned one answer per audit cycle.

incidents caught at the gate

4.2h

reviewer pickup to sign-off

record per release

Reading · aura trace

CASE 02 · SIGNED PROOF

A regulated decisioning program · reading kept private under reference path.

CASE 03 · CLINICAL OPS

↳ REVIEW RECORD

An academic medical center

Second reads went to the right reader.

Uncertain inferences routed back to the radiologist qualified for the body region, with the rubric and override chain attached to the case. Calibration drift surfaced from real escalation data, not from anecdotal feedback.

7,210

uncertain reads routed

92%

specialist accept rate

12m

median pickup time

Reading · aura trace

CASE 03 · REVIEW RECORD

An academic medical center · reading kept private under reference path.

CASE 04 · ROBOTICS

↳ REPLAY SUITE

A top-three humanoid program

Capture, review, export — one chain.

Demonstration capture, teleop review, and export-format lineage stayed attached to the same workflow that shipped the policy. Replay suites turned escaped behaviors into permanent gates that the next checkpoint had to pass.

3,402

episodes captured

98.6%

replay pass rate

behaviors gated forever

Reading · aura trace

CASE 04 · REPLAY SUITE

A top-three humanoid program · reading kept private under reference path.

CASE 05 · EVAL PROGRAM

↳ EVAL LEDGER

A frontier AI lab

Evaluation stopped starting from zero.

Rubric studio and regression bank kept every reviewed example, every override, and every comparator score on the same axis. New model families were graded against the inheritance of the old, not against a fresh spreadsheet.

184

rubrics under version control

62,140

graded comparisons

rebuilt eval suites

Reading · aura trace

CASE 05 · EVAL LEDGER

A frontier AI lab · reading kept private under reference path.

CASE 06 · TRUST READINESS

↳ TRUST PACKET

An enterprise software program

Procurement got the proof, not the pitch.

Controls, audit logs, retention windows, and export lineage were already attached to the workflow operators used day-to-day. Buyer security review started in parallel with the pilot instead of stalling at the end.

buyer security packets

custom one-off docs

shared trust record

Reading · aura trace

CASE 06 · TRUST PACKET

An enterprise software program · reading kept private under reference path.

On the record · attributed by type, not by name

“Release review went from a three-week scramble to a repeatable gate. When an edge case slips, we catch it, we save it, and it never ships again.”
Head of Model Quality · a regulated decisioning program

BRING THE WORK

Your release, your reviewer, your record.

Bring the release, review queue, or escaped issue that matters most. We will show how it runs — and what it leaves behind.

Talk to AI Labs Back to customers