# Recall development data · October 30–31, 2026

These CSV files are anonymized exports from retained local experiment results, checked on November 3, 2026. These are article attachments, not an independent benchmark dataset.

## Population and metric

- One copied personal library: 145 valid images.
- 80 fixed queries: 76 with a designated target, four negative cases. The positive queries cover 25 different images; repeated target IDs are deliberate.
- 24 images received automatic analysis. User tags were not added for these experiments.
- `target_analyzed` identifies a target in that analyzed subset; it does not mean that the whole library was analyzed.
- Rank is one-based. A target rank of 1 is first; target-in-top-K counts a positive query when its rank is between 1 and K.
- Negative-query ranks are blank. No top-K denominator includes those four cases.
- These are pre-review candidate rankings. Do not read them as precision, a model confidence score, final visual-model rank, or an independent user's success rate.

## Stages

| CSV stage | Meaning |
| --- | --- |
| baseline | Candidate ranking before the 24-image automatic-analysis addition. |
| text_first | Automatic text matches placed ahead of image candidates; rejected after ranking regressions. |
| guarded_fusion | Stricter text eligibility with rank fusion; restored baseline ranking in this set. |
| short_prompt | More concise analysis prompt; same target ranks as guarded fusion. |
| text_semantic | Experimental semantic vectors of generated descriptions/tags/OCR; text candidates can move forward while leading results are preserved. |
| protected_text_semantic | Experimental text admission plus an unchanged original top-48 pool; text can reorder within that pool. |

The last two stages remain isolated experiments, not the regular Linger 0.5.2 search path. The 48-candidate visual-review experiment is different from the app's Load more control.

## Checks and limits

All six stages use the same 80 query IDs and target mapping. Aggregating query-ranks.csv reproduces experiment-summary.csv. Three additional top-24 hits in the protected-text stage belong to one target image. The leading results and top-48 candidate set were protected by construction; that does not guarantee the first 24 are unchanged for a new query.

Most queries were authored while viewing target images; four dragon-related queries came from the user. This reused single-user regression set is not a fresh holdout or a study of independent users. Visual-model review used six fixed target queries and two visual negatives, with gpt-6-astra / low; matching historical batches were reused, and default xhigh had timed out in an earlier attempt. Those model-review results are not represented by the candidate-rank columns.

The separate October 19 model comparison used 48 images and a different query set. Do not combine its percentages with these rows.

Only anonymized ranks and aggregate counts are exported. Private images, raw queries, source URLs, original asset IDs and raw model inputs/outputs remain private. The exports reproduce the arithmetic, not the full image experiment. End-to-end app latency, complete token/billing costs, calibrated absent-target detection and cross-user performance have not been established here.
