Synthesis is the step where several responses become one answer. Done badly it averages everyone into mush or quietly copies the loudest model. Keplar's design tries to avoid both.
The sequence
- Draft. A stronger "judge-tier" model writes a draft from the positions and claims (with how many responses support each), not from raw transcripts. When little time is left, a structured draft is produced instead of prose.
- Review. The verifier, or the judge when the panel split, reads the draft against the individual responses. See The verification review.
- Final answer. The judge-tier model writes the answer from the positions, the claims and their support, the reviewer's findings and the draft.
The rules the synthesizer follows
- Weigh evidence and reasoning, not headcount. A position held by one model with good reasons can outweigh two with weak ones.
- Show disagreement only when it exists. A "Where the models disagree" section is written only if there was a split, and removed in code otherwise.
- Label counterarguments. If you asked for the best case against something, that section is labeled as a counterargument, not presented as Keplar's own view.
- Return a one-line resolution for each disagreement, used in the disagreement readout.
- Do not invent sources. Only links the models cited are listed, labeled as cited by them.
- Do not claim certainty it lacks. No confidence score or accuracy figure is produced.
If synthesis fails
If the final write-up cannot be produced, Keplar lists the positions the models took. It does not paste one model's response verbatim as if it were the combined answer.
Streaming
The answer is generated as a whole and then streamed to the page in chunks, so it appears progressively. True token-by-token streaming from the model is not in place today.
Quality gate
Before a single-model answer is shown or charged, a quality check rejects obvious failures such as repeated chunks, loops, a high ratio of non-words, or incoherent windows of text (code, JSON and non-English text are skipped). A rejected answer falls back to the next model. If every answer fails, you see "Keplar couldn't produce a reliable answer this time, and nothing was charged. Please retry."
What synthesis cannot do
It cannot add knowledge the models lack, and it cannot make a wrong consensus right. It can only organize, weigh and present what they said, plus the review.
Related
- The verification review: What the verifier or judge actually does: it reads the draft against each model's response and flags reasoning gaps, contradictions and unsupported claims. It does not browse or open sources.
- Disagreement detection: How Keplar finds where models split, how the Disagreements section is built, why it appears only when the split is real, and how to use it.
- The pipeline, end to end: The eight steps between your question and one answer: understand, select, compare, disagree, verify, synthesize, and answer, with where each can stop early.