# Praviar: paired retained-prediction analysis

This accompanies the two Praviar engineering essays published on 25 October 2026.

[Download the complete analysis bundle](praviar-paired-analysis.zip).

The calculation compares recorded ensemble selections with raw MolScribe outputs. It runs no recognition model, fusion policy or external provider. All 76 eligible occurrences remain in the main denominator: 59 both exact, 3 ensemble-only, 0 MolScribe-only and 14 neither. One identical unsuccessful record appears twice; removing its second occurrence gives a separately labelled 75-record sensitivity check.

## Files

- [Scoring script](paired_recorded_predictions.py)
- [Summary and protocol](paired-summary.json)
- [Comparison rows, CSV](paired-occurrences.csv) and [JSON](paired-occurrences.json)
- [Page inventory, CSV](page-occurrences.csv) and [JSON](page-occurrences.json)
- [Retained input](results_us_n128_live_moldet.json)
- [Scorer source excerpts](scorer-reference.json)
- [Public output hashes](paired-output-hashes.json)

## Reproduce

With Python and RDKit installed, download and extract the bundle, then run:

```
python paired_recorded_predictions.py --input results_us_n128_live_moldet.json --output-dir reproduced
```

The source archive is not required: the input and source excerpts are included. The recorded analysis used Python 3.11.10 and RDKit 2026.03.2. That is not a claim about the historical evaluator's environment. A different RDKit version may change canonicalisation; the script checks outcomes against the retained flags and stops on disagreement.

Both outputs and labels are parsed, reduced to the largest fragment by heavy-atom count and canonicalised with stereochemistry retained. Invalid or unresolved outputs remain misses. MolScribe also requires its recorded validity flag. This compares complete historical fusion selections with raw voter outputs; it does not isolate voting, measure compute savings or establish independent sampling.

The input contains 128 page entries with 117 page identifiers. The comparison contains 76 entries with 75 recorded crop keys. The one identical repeated record is at page-array positions 13 and 115, match position 1. Distinct identifiers are not proof of distinct image bytes or independent samples.

The public summary and source-excerpt metadata omit local machine paths. Numerical rows, recorded source hashes and source code are preserved. The public hash manifest describes these publication copies. Reproduction generates environment/path metadata appropriate to the machine on which it runs, so its summary need not be byte-identical to the archived calculation.

[Main essay](https://petermccannstrain.com/blog/praviar-system-design) · [Vision companion](https://petermccannstrain.com/blog/praviar-chemical-vision)