Release v0.3.0

CHANGELOG.md

@@ -9,6 +9,18 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [0.3.0] — 2026-07-26
+
+Dataset-quality and training-signal release. It closes the validator-hack / architecture-mismatch
+pattern found in the merged data (portable-code prompts + retargeting prompts to the *actual*
+validating GPU + a non-blocking guard-removal flag), adds denser preference signal (DPO correctness
+pairs, now sha256-pinned and attested in the manifest, plus a token-efficient diff-style repair
+export), and broadens the corpus with license-gated external task seeds (KernelBook / opus / gpt-oss —
+PyTorch tasks only, re-proved from scratch on attested hardware).
+
+`generator_version` remains **0.3.0**; the bundle manifest gained an **optional, attested**
+`preferences_sha256` (additive — bundles without it verify exactly as before). Package release `0.3.0`.
+
 ### Added
 - **Attested preference file in the bundle** (`sparkproof-triton-generate --export-preferences`):
   correctness preference pairs can now be written into the bundle as `preferences.jsonl` and

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "sparkproof"
-version = "0.2.0"
+version = "0.3.0"
 description = "Blackwell and Hopper GPU-attested Triton dataset generation for SparkDistill (Fable 5 + GPT 5.6 via OpenRouter)"
 readme = "README.md"
 license = "MIT"