Evolution

Better every day. Proven in public.

The Lumen Index is the public composite behind every tier — reasoning, coding, knowledge, 中文 quality, and speed, tracked from real production duels. This is the trajectory, published as it happens, not a benchmark screenshot.

01 · The trajectory

Every tracked model, since its first measurement

One line per model. Gaps in slope are real — a day a run didn't reach that model, never smoothed away.

02 · Components

Five components, one number

The Index is a composite. Each of these is judged and tracked in its own right before they combine into the single number on the chart above.

Reasoning

Multi-step, hard-to-verify problems — judged on whether the reasoning holds up, not just whether the final line looks right.

Coding

Real coding tasks, judged on correctness and quality — not whether it merely compiles.

Knowledge

Breadth and accuracy of what the model actually knows.

中文

Chinese-language quality, judged as its own component — never assumed from English performance.

Speed

Real responsiveness in production, not an advertised token rate.

03 · Cadence

Two cadences, one trajectory

A lighter run and a deeper run keep the line moving without either one dominating it.

A lighter run happens daily, rotating its sample so coverage broadens over time — no single day tries to be a mega-run. A deeper run happens weekly, over a larger, fixed sample, and anchors the trend the daily runs trace between two deep readings.

Both run exactly like everything else in this system: sampled production traffic, judged off the serving path, at no cost to you. See Methodology for how a single duel actually works.

04 · The anchor

Fifty is a start, not a ceiling

Every model's index starts the same way, and moves only on evidence from there.

The first time a model completes a full five-component measurement, that run sets its baseline at 50. Not a hand-picked constant — whatever that model's own first real evidence works out to, normalized once, at that moment, and never recomputed afterward.

50
Each model's own first measurement — not a hand-picked constant, not a ceiling.

From there, the index moves however the evidence moves it. There is no ceiling: sustained, genuine improvement can carry a model's index past 100 over time.

Fifty is a starting line, not a scale limit.

05 · Confidence

What the shaded band means

Every point on the chart carries a range, not just a headline number.

The shaded band around each point is how sure that measurement is — not decoration. A wide band means the evidence behind that point is thinner, or the result closer to a toss-up; a narrow band means it's well-established.

Production duels don't produce a perfectly even amount of evidence every single day, so the band says that honestly instead of hiding it behind one clean line.

06 · Certification

The ladder is a hypothesis, not doctrine

Tier order isn't asserted once and left alone. It's checked continuously against production evidence.

Lumen FlashLumenLumen ProLumen Ultra

Win-rate monotonicity

Judged head-to-head comparisons on real production traffic must show Flash ≤ Standard ≤ Pro ≤ Ultra in answer quality, region by region of the request space. A tier holds its slot only while it separates.

Merge on non-separation

If two adjacent tiers fail to separate beyond noise for a full evaluation window, the tier count is wrong and the tiers merge. Four is a hypothesis the data must keep earning, not doctrine.

This page is the evidence trail. For the mechanism that produces it — shadow duels, position-swapped judging, promotion gates, and automatic rollback — see Methodology. Or compare all four tiers side by side on Models.