Review semantic inference

Harnery validates every semantic reply before Codec can use it. Validation checks shape, citations, confidence, vocabulary, and privacy. It does not decide whether the model chose the best interpretation. The local meaning review adds that missing human judgment.

Open /codec/review while the Harnery dashboard is running. The page prepares a frozen set of up to 12 accepted readings. Each item places two panels side by side on wide screens and stacks them on narrow screens:

• Bounded source facts shows the task, lifecycle, intent, operation, waits, attention, and recent controlled observations available to the model.
• Accepted model reading shows the model, attestation, headline, summary, phase, optional expression, and other accepted semantic fields.

Judge the evidence first, then answer each controlled question. Overall meaning, phase, and expression are separate because they fail for different reasons. When phase or expression is wrong, choose the expected value from the controlled vocabulary. When the current reading changed from the prior captured reading, label that transition as real work, flicker, or unsure.

Submitting the final item writes one private receipt and removes every item in that set from the pending queue. The receipt keeps the model identity, proposed controlled phase and expression, verdicts, corrections, confidence, opaque hashes, and bounded timing. It does not keep the task, source labels, semantic prose, citations, or raw model reply.

Candidate context stays local for seven days and is capped at 100 files. Review receipts are capped at 1,000 files and one year. Neither store enters Event Ledger V3.

CLI workflow

The page uses the same commands available to an embedding host:

harn semantic review prepare
harn semantic review submit --file result.json

prepare backfills current accepted documents only when their evidence digest still matches the active projection. It then prefers changed readings, expression-bearing readings, and newer readings. submit checks the frozen study hash, candidate IDs, controlled values, correction requirements, and duplicate-review state before it writes an atomic receipt.

The review capture is observational. A local evaluation-storage failure cannot turn an otherwise accepted semantic reading into a failed reader pass.