Technical report · No. 14 · July 2026
A model that shows its work.
Meridian answers with its reasoning attached, not hidden behind it.
Meridian is a language system trained to prefer a slower, checkable answer over a fast, confident one. In a held-out set of 4,120 multi-step problems, annotators preferred Meridian's stated reasoning path 71% of the time when the final answer was wrong — because the path made the error easy to find. We describe the training procedure, the failure modes we still see, and the interface built around a single idea: a model is more useful when you can watch it think.
Meridian holds a working context of 340,000 tokens, roughly the length of a three-volume novel, and keeps a citation for every factual claim it makes inside that window. Ask it to revise a contract clause and it will mark, inline, which paragraph of the source it drew from. Ask it to debug a stack trace and it will point to the exact line before it explains the fix.
It is deliberately slow on arithmetic-heavy and legal work, trading roughly 1.8 seconds of extra latency per answer for a measured 23% drop in silent numerical error, verified against a panel of nine accountants in Portland over eleven weeks.
The interface below is the only product surface we ship today. There is no dashboard, no plugin marketplace, no avatar. Text goes in on the left, reasoning and answer come back on the right, and both are kept in a transcript you can export as plain text at any time.
Meridian declines roughly one in twenty requests outright — medical dosing, structural load calculations, anything where a wrong answer costs more than a right one saves — and says so in one sentence, not five.
Rewrite this for a 2:1 aspect ratio without losing the last sentence.
Done — trimmed the second clause of the opening line and kept the close verbatim. 214 words to 168.
What did you cut, exactly?
"in a way that few other firms in Ohio can claim" — it was true but unverifiable, so I removed the claim rather than soften it.
- Evaluation set: 4,120 multi-step problems, internal, released alongside model card v0.9.
- Latency figure measured on a single A-class inference node, batch size 1, July 2026.
- Accountant panel: nine practitioners, Portland OR, eleven-week engagement, compensated.
- Refusal rate measured on the public red-team prompt set, not on production traffic.