Evidence

What has been measured

Every competitor's site asks you to trust it. This page hands you the means to disbelieve. Nothing below is a claim — each line is a number and what produced it.

44× → 8.7×
memory for ten times the history

Our own engine, before and after. Ten times the history cost forty-four times the memory; it now costs 8.7. Six seeds, three thousand fit points, acceptance bars written before the measurement so the answer could not be tuned. Saved state fell from 894 MB to 125 MB with output byte-identical throughout.

300,000
computation steps

Re-run start to finish across six independent runs on a pinned toolchain. Identical to the last digit.

200,000
output rows, cross-compiler

Four runs of 50,000 steps, compared against references produced on a different compiler two months earlier, on different hardware. Every row identical. Two of the four endpoint values were published before their runs finished. Both landed exactly.

29 / 29
corrupted results rejected

Deliberately corrupted results submitted blind to our verifier. All 29 rejected. A checker that only ever approves is worthless.

400 / 400
adversarial probes rejected

Backdoor insertion, certificate forgery, deliberate weakening, gradual degradation — every probe rejected, and banked as a permanent regression test.

155 / 27
iterations across phases

155 development iterations across 27 phases, every one adversarially reviewed. Findings were filed against every author, including the founder.

6 seeds × 50,000 steps
the post-fix cohort

The measurement run that produced the memory result. Both interventions active, endpoint hashes locked by adjudication, old baselines retained rather than replaced.

What we publish about ourselves

A verifier that only ever agrees with you is worthless. The same applies to a company's own record.

We keep an append-only register. Prior entries are never edited; corrections are filed as new entries with the original left intact.
When a result does not survive scrutiny, we retire it on the record and say why. Several figures that once appeared in our own materials have been withdrawn this way.
A performance projection we made was contradicted by measurement. Rather than adjust it quietly, the contradiction was filed. A cleaner instrument later confirmed the original projection within 7%. Both facts are on the record — including the interval where we were wrong.

What is not yet proven

Flat memory

Memory grows more slowly than the work, but it does not stop growing. We measured why: our engine’s past keeps acquiring new connections, so it genuinely uses all of its history. Two routes to flat were tested and both closed by measurement. We do not claim flat and never have.

Cross-toolchain state stability

Not claimed. Output trajectories are identical across compiler generations — measured, four runs of 50,000 steps. Full internal state is identical within a pinned build, and we say so.

Prefer to check us rather than take our word for it? Send one month of decision logs and your own assessors verify whether they reproduce. No fee, no integration.

The Reproducibility Audit

All figures on this page are internal verification measurements from Sentium X's development record. They describe what the deterministic core has been tested to do; they are not a performance benchmark or a guarantee of behaviour in your environment.