Skip to content

CHANGELOG.md

Latest commit

 

History

History
52 lines (41 loc) · 3.19 KB

CHANGELOG.md

File metadata and controls

52 lines (41 loc) · 3.19 KB
Raw

Changelog

0.4.0 — May 2026

The 0.4 release rebuilds the package around the Behavioral Trust Clustering architecture validated empirically on Qwen2.5-Coder-7B (companion paper, in preparation).

Added

  • HybridLayer: end-to-end Generator × Governor pipeline with K-sample stochastic decoding and behavioral clustering.
  • behavioral_governance: closed-form trust thermodynamic governor exposed as a standalone function for offline analysis.
  • trust_thermodynamic: the closed-form trust score, $T = \mathrm{PPV} \cdot \exp(-\sigma_{\text{calib}} \cdot T_{\text{comp}})$, with adaptive computational temperature.
  • cluster_by_behavior: behavioral (output-equivalence) clustering for code candidates.
  • make_python_probe_runner: factory for behavioral probe runners that execute candidates in subprocess sandboxes.
  • extract_probe_args_from_tests: regex-based extraction of probe arguments from assertion-based test specifications.
  • cluster_by_ast: structural clustering, retained for ablation studies.
  • Backend protocol LLMBackend with three reference implementations: OllamaBackend, OpenAIBackend, CallableBackend.
  • Layer 1 system prompts (LAYER1_SYSTEM_PROMPT_EN, LAYER1_SYSTEM_PROMPT_IT) with calibrated utility framing.
  • Strict input validation on HybridLayer and trust_thermodynamic.
  • 35-test pytest suite covering governance, parsing, clustering, and end-to-end integration.

Changed

  • The trust formula default for $T_{\text{comp}}$ is now adaptive ($0.5 + (1 - \mathrm{PPV})$) rather than fixed at $1.0$. The fixed form remains available via the t_comp parameter.
  • Confidence parsing tolerates inline CONFIDENCE: X markers as well as separate-line markers (regression from 0.3 found during HumanEval validation).
  • The Shannon entropy normalization for $\sigma_{\text{calib}}$ now divides by $\log(n_{\text{clusters}})$ rather than $\log(K)$, eliminating a numerical artefact at perfect agreement.

Removed

  • Legacy Substrate and Cascade classes from 0.1.x; their functionality is subsumed by HybridLayer and the standalone behavioral_governance.
  • The @verified and @with_abstention decorators; replaced by direct HybridLayer.query invocation, which is more flexible and easier to instrument.

Fixed

  • sigma_calib no longer becomes negative under perfect agreement due to the $\epsilon = 10^{-9}$ inside the logarithm of the 0.3 implementation.
  • The cascade did not propagate the modal answer correctly when the modal cluster was singleton; this case is now handled.

Migration from 0.1.x

The 0.1.x API is incompatible with 0.4. The migration is straightforward:

# 0.1.x
from snc_core import Cascade
cascade = Cascade(substrates=[s1, s2, s3], threshold=0.5)
result = cascade.evaluate(prompt)

# 0.4
from snc_core import HybridLayer
from snc_core.adapters import OllamaBackend
hybrid = HybridLayer(OllamaBackend(model="..."), k=5, threshold=0.5)
result = hybrid.query(prompt)

Concretely, the 0.4 single-backend / K-sample model replaces the 0.1.x multi-substrate / single-call model. The two are not isomorphic: 0.4 is the architecture validated in the companion paper, and we recommend it for new projects.

0.1.0 — March 2026

Initial public release.