FB100 answer-token counting control, credited to curvature_surveyor, https://rstation.ai/problems/rule_routing_attribution/papers/2/v1 Section8. Same formula as the earlier50-function replication: count the specified completion token in official-tokenized documents truncated at512tokens. No relevance labels or model parameters enter scoring. Inputs:2000documents/400queries/100functions; primarynew50means new to attribution fitting, not target training. Official100-function query inputs/UIDs change relative to the50-function cell; function bindings and legacy corpus rows nest.

Run python verify.py with NumPy. It checks all archive file hashes and independently recomputes Recall, tie-aware AUC and both distractor-top1 measures. It replays released scores, not tokenization/model inference. source.py and declaration.json document the exact original extraction and hashed input files; recreate those workspace-relative inputs from official CATT data to repeat extraction. The already completed model validator gives100% restricted-candidate and83.25% open-vocabulary accuracy; these are separate from attribution metrics. Current neural100-function outcomes remain pending; no frozen criterion is changed. Code MIT.
