Hiro development journal

Mistral extractor qualification stops at artifact identity

2026.08.30 at 19:34:15 America/Los_Angeles Pre-execution model identity check complete; qualification blocked before hashing or inference Machine-readable JSON

Executive summary

A bounded continuation was authorized to qualify only Mistral Small 3.2 24B Instruct 2506 in Bartowski Q6_K form under Hiro's unchanged claim-extraction contract.

The pre-execution identity check found the model family in the configured local stores, but the installed text-model artifact is Q4_K_S rather than the authorized Q6_K variant.

Because substituting another quantization would violate the frozen model identity and the explicit scope of the qualification, no model hash was accepted, no worker was launched, and no runtime or quality request was executed.

Production Qwen and all prior extractor candidates were left untouched. The frozen six-source quality workload and twenty-source integration corpus were not run.

Work completed

Authorized-artifact verification

Blocked

The authorized identity was Mistral Small 3.2 24B Instruct 2506, Bartowski Q6_K GGUF.
The available text-model file identifies itself as Q4_K_S and has a size of 13,549,285,504 bytes. A separate multimodal projector is also present but is not the requested text-model quantization.
The configured model locations were checked for another matching Mistral Small 3.2 24B artifact; no Q6_K file was found.
The Q4_K_S artifact was not silently substituted or hashed as if it were Q6_K.

Decisions and reasoning

Stop before execution because the installed artifact does not match the sole authorized model variant.
Do not reinterpret model-family equivalence as quantization identity equivalence.
Do not retest prior models, change qualification policy, launch a worker, run quality, enable routing, or execute the frozen twenty-source corpus.
Resume only after the Q6_K artifact is installed or the user explicitly authorizes the installed Q4_K_S variant as a different qualification target.

Validation and evidence

Check	Status	Result
Authorized GGUF identity present	failed prerequisite	No Mistral Small 3.2 24B Instruct 2506 Q6_K GGUF was found in the configured local model stores; only Q4_K_S was present.
Runtime qualification	not run	Blocked before hashing and worker launch by the artifact identity mismatch.
Extraction-quality qualification	not authorized	The runtime prerequisite was never entered.

Current state

The existing claim-extraction qualification evidence remains unchanged.
No newly qualified extraction model exists.
Production routing remains disabled and unchanged.
The frozen twenty-source corpus is not cleared for requalification.

Next steps

Install the exact Bartowski Q6_K GGUF and resume the unchanged qualification, or explicitly authorize Q4_K_S as a separate candidate.
Keep all frozen prompts, corpora, schemas, thresholds, retries, and supervisor behavior unchanged when qualification resumes.