Tests and benchmarksEvery result read again at benchmarks.eigna.ai on 24 October 2026

We publish our results. Including the ones we lose.

At the gate so far: 3 held, 0 released
Proof Ladder: 1 of 54 cells done
Published: Benchmark 001, 14 October 2026

The lead story · Benchmark 001

17.8%

Our first training run cut forecast error 17.8% against its base model. It is not yet ahead of a simple last value baseline, and we publish that too.

The full report benchmarks.eigna.ai

Base model, Kronos-small1.750%

Our first training run1.439%

Last value baselineStill ahead of ours1.392%

Benchmark 001. Forecast error (CRPS), lower is better. 243 held-out days, hourly BTC and ETH, 24 hours ahead. Run on 14 October 2026.

0 of 3Simple baselines beatenLast value, EWMA, GARCH(1,1)
57.3%Band coverageTarget 80%
+4.4%Paper tradingBase model +15.0%

Case 001

Held at the gate.

Three candidates of our first client’s forecasting model met the gate in October 2026. All three were held. Release 1 was not issued, and nothing untested reached the desk.

Read Case 001

Eigna TLM, Case 001

Release record

Candidate 3
23 October, shrinkage head
A tie, p 0.46 to 0.53Held
Candidate 2
14 October, calibrated
Coverage 88.3%Held
Candidate 1
14 October, LoRA, 4 runs
CRPS −17.8% against baseHeld
Release 1
Not issued
No release

On record

Every test and case, in one format.

Methods and corrections

1.750%1.439%1.392%Test14 October 20267 minBenchmark 00117.8% lower forecast error than its base model. Not yet ahead of a last value baseline.

The Figure cast as light on a mineral green sphere

Case25 October 20268 minCase 001: held at the gateA forecasting model of its own. Measured in the open, held at the gate.

Test24 October 20263 minThe Proof LadderSix rungs, nine model families, 54 cells. One done. What runs next.

A hand writing on cream paper on dark slate, the Figure drawn at the top of the sheet

Test24 October 20263 minMethods and correctionsMetrics, test design, digests to reproduce, limits, and the standing corrections rule.

The Proof Ladder

1 of 54 cells done.

Six rungs, nine model families. We publish the empty cells too: they are what we still have to prove.

What runs next

Proof Ladder cells by family and rung
TLM, time series	Done	Partial	To run	To run	To run	To run
LLM, language	To run	To run	To run	To run	To run	To run
Voice	To run	To run	To run	To run	To run	To run
Image	To run	To run	To run	To run	To run	To run
Video	To run	To run	To run	To run	To run	To run
Robotics	To run	To run	To run	To run	To run	To run
Typed router	To run	To run	To run	To run	To run	To run
3D	To run	To run	To run	To run	To run	To run
Guardrails	To run	To run	To run	To run	To run	To run

Done 1 Partial 1 To run 52

Read on 24 October 2026 at benchmarks.eigna.ai.

How we claim

A metric, a baseline and a date. Never adjectives.

01Name the metricIn plain words, with its unit, and whether lower or higher is better.
02Name the baselineThe simplest one first. Beating the base model is not enough.
03Date itThe day of the run and the day we read it again.
04Publish itWhether it wins or loses, next to the win, at the same size.

Runs today: the first expert, the scoring harness and sealed tests, training on local machines and H200 GPUs, benchmarks.eigna.ai.

What runs today benchmarks.eigna.ai

Test Lab

Your proof, on your work.

In about four weeks, on your own data in a sealed setup, we show whether an Eigna expert beats your baseline. If it does not, you keep the report.

Request a Test Lab Book a call