Supporting formula report / 13 October 2026
What does a perfect reconstruction score establish?
A small counterexample makes the reward's boundary concrete. Reading the limitations also corrects an objection: the authors already distinguish consistency from biological correctness.
Decision summary
- Executed
- Transcribed printed Eq.3 evaluated on four illustrative two-gene pairs; same-set permutation checks included in the test suite.
- Read and compared
- Reward definition, p.6; author qualification, p.10. An objection that the qualification is absent is rejected.
- Not executed
- Released-code audit, model evaluation, biological validation or experimental replication.
Claim-by-claim findings
The executed formula check
Reference genes: [A, B]. Their reciprocal-rank weights are 1 and 1/2. For reconstruction [B, A], each shared gene contributes (1 + 1/2) / 2 = 3/4 to the numerator. The reference contributes 1 + 1/2 to the denominator; there are no generated-only genes.
The order differs, but the value is maximal. More generally, for equal-length lists containing the same unique genes, the sum of reciprocal ranks is unchanged by permutation in this equation.
| Recon. | Num. | Den. | Value | Same order? |
|---|
The example uses symbolic genes and does not model biological expression. The code is our transcription of the printed equation, not the authors' implementation. No trained-model failure follows from this check.
What would resolve the remaining question?
First compare the printed equation with the released implementation. Then test whether ordering changes affect the intended biological task, using an independent criterion rather than the optimized reward itself. These are proposed checks, not completed results.
A useful reward can improve prediction without certifying a mechanism. The appropriate response is to scope the claim and investigate the intended use, not discard the paper's reported improvements.
Source, version and attribution
Analysis provenance
This report examines the reconstruction reward in the CellDuality research paper by Yuhang Chen and colleagues. It reproduces our documented formula check and preserves the authors’ stated limitations. The source excerpts are unchanged; the interpretation and illustrative check are ours.
Limitations of this report
No independent paper replication, implementation audit or biological experiment was performed. The source file is identified by its hash; we have not independently established its precise revision date. The public OpenReview page can change and may require browser verification. The retained PDF remains available with attribution.