Skip to content

v0.7.8

Choose a tag to compare

View all tags
@github-actions github-actions released this 27 Sep 16:18
· 764 commits to master since this release
v0.7.8
bc78785

Cortiq 0.7.8

Cortiq Decision

  • Typed decisions in one CMF file, run by the cortiq CLI on a CPU: Resonance Routing — every label is an isolated affine subspace, the answer is the smallest reconstruction error, with a calibrated confidence, a novelty score and a gate certified on held-out calibration rows, so the model abstains instead of guessing. The router encoder (bge-small-en-v1.5 base, MLP compressed by 75 % with Cortiq NVG) and 4096 hashed lexical features are native CMF tensors — no ONNX Runtime, no new external crates. The model: infosave/cortiq-decision.
  • cortiq decide FILE -p TEXT (single, JSON, batch with per-stage timings) and cortiq decision train | add-skill | learn | keys | verify | rollback … — your own skill from JSONL {text, label}, existing skills kept byte for byte.
  • cortiq serve FILE speaks two protocols: the request/response shape of the Jev decisions API on OpenRouter (POST /api/alpha/decisions, /v1/models, usage and cost in every answer, prices 0 by default) and the cortiq-router API, schema 1.1 (/v1/route, :batch, feedback, taxonomies, usage, escalations, admin keys, metrics). cortiq decision keys import moves existing router keys (sha256, no reissue); serve --shadow-of URL answers from the old router while comparing silently, for a safe switch.
  • Oracle in two steps: export OPENROUTER_API_KEY=… and cortiq serve FILE --oracle deepseek/deepseek-v4.1-flash --oracle-budget 5. Only questions the model does not decide go to the oracle, with a reserved budget, stop rules, PII redaction and a semantic cache; the answers teach the model without changing any other label. cortiq decision oracle check [--test-call] checks the key, the model and its price.
  • Language-model commands refuse a decision file, and cortiq 0.7.7 refuses it too (DECISION feature bit).

Fixes

  • Row-exact q4tp batch multiplication now equals the single-token path on ARM as well (Apple Silicon): the MiMo MTP verification scope is exact there too, and the macOS CI test that failed since 0.7.7 passes.

Verified measurements

Test sets vs the stored answers of Jev 1.13 on the same rows: CMF alone 93.34 % / 96.18 % / 86.15 % (BANKING77 / CLINC150 / MASSIVE; Jev 85.58 / 96.76 / 85.78); certified gate 97.24 % at 90.6 % coverage / 98.70 % at 92.1 % / 97.89 % at 54.3 %; CMF + DeepSeek V4.1 Flash on abstentions 93.96 / 97.47 / 88.00 % at $3.1 / $3.6 / $8.5 per million decisions (Jev $184 / $272 / $111). Text → decision p50 3.9 / 3.8 / 3.0 ms on one Apple M4 core.

Known limits

The benchmarks are public and reused; Jev likely saw them in training. The encoder is English. The gate is certified in-domain only. Speed was measured on macOS arm64. Building from source needs Rust 1.88 or newer (as 0.7.7 already did).

Install

cargo install cortiq-cli --version 0.7.8 --locked
cortiq --version

Prebuilt CLI and FFI archives are attached with SHA-256 files.

Assets 22
Loading