Review queue

Choose the review view

The Review queue uses saved successful API responses and retained batch results from your account. It indexes existing results in pages while open.

Save a threshold from 0 to 1 (default 0.65). Below threshold includes only known scores strictly below it. A score of 0.66 does not appear at a 0.66 cutoff. Missing scores are excluded. Similarity is a matching signal, not an accuracy percentage.

• Priority issues shows offers flagged during an audit, at any score. Confirmed issues appear before cases needing investigation. Open an item to see its reason and supporting sources. An audit flag is separate from an LLM judgment or training label.
• Unresolved / uncertain hotel shows those cases at any score.
• Review history shows requested rechecks and decisions at any score.

Within each view, audited issues appear first, followed by other items from newest to oldest. Priority does not bypass the cutoff in Below threshold.

Counts, status filters, search and pagination apply to the selected view. Raising the cutoff backfills retained source data; lowering it narrows the default view without deleting reviews or labels. This threshold selects review work and does not change mapping behavior.

Run a recheck

Open a room and select Run LLM recheck. It runs asynchronously against the saved inventory's confirmed hotel. The page shows the actual model and decision; you can leave the page while it runs.

Each recheck forces a fresh LLM decision, bypassing the encoder fast path and mapping cache. The picker compares the supplier name with catalogue candidates. Supplied descriptions, beds and views support separate conflict checks, rather than being passed into this picker.

A supported selection without detected conflicts creates a model-proposed label. Abstentions and conflicts remain Needs evidence; failed or unacknowledged LLM calls create no label. Rechecking never rewrites the original response or changes your mapping policy.

Export labels

Export labels provides separate JSONL files for model-proposed and human-approved labels. Approving moves a proposal to the approved export. Rejecting or withdrawing removes it from future exports. Abstentions are not negative training pairs.

Exports cover your account, independently of page filters, up to 10,000 labels. They include source evidence, policy and timestamps, model provenance and label source. Approval records an account session, not a named individual. Already downloaded files cannot be recalled.

Labels are exported for a separate training pipeline. The sandbox does not submit training jobs; exports carry trainingStatus: "not_submitted". Model-proposed labels remain distinguishable from human-approved labels.

Review limits

Up to 100 accepted rechecks per account per UTC day, including retries, and three attempts per item. A recheck does not consume the mapping-submission quota. Failed calls require an explicit retry.

Reviews survive browser closure and service restarts. Saved snapshots remain available after source expiry, but missing source data cannot be reconstructed. A review describes the saved response, not necessarily what a fresh mapping would return today.