Five context tiers in each of two tracks.
Validation and evidence
The acceleration was measured. The boundary was exposed.
Ten frozen jobs across 8K to 131K contexts established a governed record of inputs, configurations, outputs, timing and scoring. The record shows where the tested generation succeeded and where the successor must improve.
01 / Governed record
Two tracks, five tiers
A two-week, ten-job answer-blind evaluation under TurinTech review.
Every job binds the declared input, selector generation, configuration, output and deterministic scoring record. ARES preserves the familiar-structure and unfamiliar-structure results separately.
8K, 16K, 32K, 65K and 131K frozen tiers.
Measured paired end-to-end acceleration at accuracy parity.
Passthrough rows remain in the speed denominator.
02 / What succeeded
Known-format regime
Measured acceleration with strict accuracy at or above the tested baseline.
Reference result on the tested known-format track.
Accuracy result for the evaluated WarpDrive generation on that track.
Total baseline batch wall time divided by WarpDrive batch wall time at each tier.
Why the denominator matters
This is not a selected-row compression ratio. It is paired batch end-to-end timing with accelerated and passthrough rows retained. Context length and accounting scope stay attached to the number.
03 / What the test changed
Unfamiliar workload structures
Acceleration alone was not enough.
On the unfamiliar-structure track, the evaluated generation accelerated execution but did not preserve baseline accuracy. The complete row-level result stays in controlled diligence so the public page does not hide the limitation or turn private evidence into a downloadable corpus.
The corpus-agnostic safety objective was not met.
The result prevents an unsupported claim of universal safety, broad deployment readiness or accuracy preservation on every corpus.
The successor compiler has a precise target.
Safe generalisation across unfamiliar workload structures must be proven before broad deployment.
04 / Independent perspective
Private evidence, public paraphrase
Novel mechanism. Real observed speedup. Scoped claim.
After reviewing the work with Dr Mike Basios, TurinTech CEO Dr Leslie Kanthan characterised WarpDrive's algorithm as novel and the observed speedup as real.
The source is private correspondence. It is not published here, and this paraphrase is not presented as a TurinTech-authored final public report.
Review the evidence boundary.
Qualified investors can request the public validation summary and, under an agreed diligence boundary, the governed record needed to examine the claims.