Ir al contenido

← Blog

Releasesv0.63.0

v0.63.0

Esta es la nota de release tal como fue publicada, no una reescritura de ella. Las notas de release se publican en el idioma en que fueron escritas.

Added

  • Answers grounded in your documents are checked against them, on by default. When a message asks a question about attached documents, and the answer is written from them with no tool call in the turn, a System One verifier reads question, sources and answer.

    • What happens next:
      • If the answer is supported with confidence ≥ 0.8, it ships.
      • Otherwise it escalates to a stronger model (gpt-6-sol), which is verified again.
      • If that also fails, you get "the sources provided don't cover this" in your language, instead of an unsupported answer.
    • Measured in bench/verified_cascade on 400 grounded questions, with gpt-6-luna drafting: wrong answers shipped went from 33 to 21 (Holm p = 0.002, 11 fixed and 0 broken), at 1.87× the cost with the local verifier, and with no answerable question handed off.
    • Not applied to tasks on documents (summarize, critique, translate, rewrite), to tool-using turns or to solve. A deterministic classifier routes tasks past the check.
    • Where to see it: a badge on the answer (verified, sources don't cover it, lexical only, verifier unavailable, not checked), with the withheld text one click away, and the receipt's grounded block records every step.
    • Controls: CHIMERA_VERIFIED_ANSWERS, CHIMERA_VERIFIED_ANSWERS_THRESHOLD and CHIMERA_VERIFIED_ANSWERS_ESCALATE_MODEL.
    • In the terminal, chat and assist gain /attach <file>. (#656)
  • Choose the System One decision model in Settings and in the terminal. A new "System One" card offers three backends: Local (qwen3:4b, free), Hosted judge, or OpenRouter System One.

    • Model list: it comes live from OpenRouter's output_modalities=decisions index, with price, context and a calibrated or uncalibrated badge.
    • Selectable today: typesafe/jev-1.13 and jaredpalmer/kev-4b.
    • Refused, with the reason: the moving alias ~typesafe/jev-latest, and respan/span-01*, whose behavior-scoring contract has not been measured.
    • Terminal: chimera decisions models and chimera decisions use <backend> [model].
    • Fallback: the verified-answers check uses this choice. The local verifier falls back to Jev when an OpenRouter key exists, and to the lexical check otherwise. (#653)

Benchmarks

  • bench/verified_cascade — pre-registered, US$ 4.89 spent.

    arm wrong answers shipped (of 400) result cost
    Decision-gated cascade, Jev verifier 21 significant against luna alone 3.74×
    Decision-gated cascade, local verifier 21 significant; tied with Jev, so the local one is preferred 1.87×
    Today's lexical gate 22 null 10.8×
    gpt-6-sol alone 20 null 19.4×

    The run needed three amendments: a grader swapped after failing its gate, 42 disagreements adjudicated, and a report bug that read zero hand-offs as missing. (#654, #655)

Leer la release en GitHub →