| Surface-accessibility gating | Treats a credible cell-surface extracellular domain as a hard gate: correctly deprioritizes intracellular antigens (TP53, GAPDH) as surface-ineligible for CAR-T/ADC/bispecific regardless of tumor expression, and notes they are only addressable via TCR-T/pMHC. | Recommends or fails to disqualify an intracellular antigen for a surface modality, or ignores localization/topology entirely and scores on tumor expression alone. |
| Tumor/normal specificity | Judges the antigen on tumor expression RELATIVE to normal-tissue expression (tumor/normal ratio), not raw tumor TPM; recognizes that a high tumor TPM with high normal expression is a poor therapeutic window. | Equates high tumor expression with a good target, ignores normal-tissue baseline, or never computes/considers the tumor-vs-normal ratio. |
| Normal-tissue / vital-organ safety | Flags expression in vital, dose-limiting tissues (heart, brain, lung, liver, etc.) as an on-target/off-tumor toxicity liability and weighs breadth of normal expression; treats vital-tissue expression as a dose-limiting gate, not a minor deduction. | Misses or downplays vital-tissue expression, treats safety as an afterthought, or never checks the GTEx normal profile against the vital-tissue list. |
| Integrated prioritization & tiering | Synthesizes surface gate + specificity + safety into a clear, well-ordered prioritization (score/tier or ranked recommendation) that matches the evidence, with calibrated confidence and the right caveats; in comparisons, orders candidates correctly. | No clear prioritization, a ranking contradicted by its own evidence, or over/under-confidence inconsistent with the data. |
| Evidence faithfulness | Every number and claim (TPM values, topology, ratios, flags) traces to the tool outputs; no fabricated expression values, variants, or programs. | Hallucinates TPM numbers, localization, or clinical programs, or contradicts the returned tool data. |