Proof methodology · buyer-readable, source-bound

Proof that gets stronger under pressure.

The model did not change. The prompt set did not change. The scoring contract did not change. Aria's measured advantage did—and every boundary stays visible.

Best same-contract lift: +2.85
Larger than first dated lift: 3.8×
Prompts improved in best run: 10/10
Real-browser repair checks: 20/20

Same-contract quality series

Same model. Same ten prompts. 3.8× more lift.

Select either dated run to inspect the result. The line connects only those two observations—there is no smoothing, interpolation, or forecast hiding between them.

+2.10lift gained

103 minbetween run starts

+1.23/hrobserved two-point slope

+3.00

Same provider · model · 10 prompt hashes · 14-point rubric · concurrency 4

Run 2 · 10:36:24 PM UTC

+2.85 mean lift

Prompts won: 10 / 10
Aria mean: 12.45 / 14
Runtime-complete rows: 10 / 10
First-class gap rows: 0

Source-bound run · 2026-05-18T22-36-24-995Z

Inspect the recovery observation

One exact prompt met a hard measurement boundary. The next runtime-complete observation returned positive in 3m 46s; the best reached +3 in 6m 31s.

Selected complete observation+1 lift

226 seconds after the boundary · same prompt · runtime complete

01 · Quality acceleration

The test stayed fixed. The advantage grew.

Two complete runs used the same provider, model, ten prompt hashes, 14-dimension rubric, and concurrency. That makes the change in measured lift inspectable instead of decorative.

Mean lift moved from +0.75 to +2.85 in 103 minutes—a 3.8× larger advantage under the unchanged contract.

Prompt wins moved from 6 of 10 to 10 of 10
Runtime-complete rows moved from 0 to 10
Two observed points are connected directly; no smoothing or forecast is used

02 · Pre-execution challenge

Every counted challenge names what would prove it wrong.

The challenge record counts checks, not vague claims that work was reviewed. Each counted check is attached to an action ID and includes a named falsifier before execution.

123,777 structured checks covered 8,275 unique actions.

Every counted check includes a named falsifier
Observed from June 5, 2026 through July 13, 2026
Multiple checks can examine one action from different failure angles

03 · Expanded comparison

The advantage held when the prompt set doubled.

A later 20-prompt run kept the same-model comparison visible as a separate contract. It is not appended to the ten-prompt line, so expansion cannot masquerade as a trend.

Aria produced stronger operating judgment on 16 of 20 prompts, with a mean lift of +1.675 points out of 14.

Baseline mean · 9.2 / 14
Aria mean · 10.875 / 14
Warnings · 14 baseline vs 6 with Aria

04 · Repair lineage

A visible failure became protection the next client can inherit.

The incident proof begins with one observed interface regression, follows the repair into source, then exercises the corrected behavior in the browser. Reuse begins only after the real interaction passes.

One incident became 5 regression protections; 20 of 20 browser checks passed with zero browser or console errors.

Repair lineage · cdd7a72d1
Pointer, scrolling, overlay, and continuation behavior were exercised
The public record omits private runtime and session identifiers

05 · Learning promotion

Memory is cheap. Earned reuse is the advantage.

Thousands of observations can enter the learning system without becoming policy. The record separates learning rows, candidates, canary results, and accepted promotions so only supported improvements shape future work.

4,843 learning records were examined; 11 improvements earned durable reuse.

4,812 evolution candidates
10 supported canary results
Promotions reached tests, skills, owner guidance, routing, and a control boundary

Claim discipline

Powerful enough to inspect. Precise enough to trust.

The connected quality series contains two complete, comparable observations—not a statistical forecast.
A runtime configuration block is shown as a measurement boundary, never plotted as a decline in judgment quality.
The later 20-prompt comparison is a separate expanded contract and is not connected to the ten-prompt series.
Live provider output is stochastic; these dated observations do not guarantee the same score on future prompts.

Apply this advantage to one live opportunity

Bring the deal. Leave with the result your client has to feel.

Request the private fit decision

Machine-readable appendix

The buyer-facing explanation above is bound to sanitized, dated records. Analysts can inspect the quality series and the broader public operating record directly.

Open quality series Open public record