Interface-Induced Trajectory Censoring

Reported agent tool-call rates are a property of the model-interface stack, not the model: a mismatched serving adapter can zero out benchmark scores and starve RL training of tool-mediated trajectories even though the model is emitting well-formed tool calls.

On BFCL v4 (200 cases) and τ-bench (115 retail tasks) with weights, decoding, seeds, executor and scorer held fixed, changing only the serving adapter configuration; a 2x2 over chat template and parser; a scale ladder across Qwen2.5-Coder spanning 21x size with n=100 per arm; and RL rollouts in verl's AgentLoop at 1.5B and 7B (GRPO/RLOO, 150 steps, 540 held-out items). Intent is measured via a validated classifier (kappa 0.871 against third-party adjudication).

BFCL v4 same-model scores move across the full metric span: 0.00 vs 0.96 (simple_python) and 0.00 vs 0.19 (multi_turn_base); parsed calls 0/200 to 196/200. On τ-bench, server-parsed calls go 0 to 636 and tasks reaching tool execution 0 to 103. Qwen2.5-Coder emits 0/100 parsed calls at every scale while well-formed emitted calls rise to 80/100 at 32B (about 72 after precision correction), vs a preregistered matched-envelope counterfactual where silent fraction stays at 0-2. Llama-3.1-8B's 23% self-tool-calling failure falls to 0 under strict:true. In RL at 7B, 45/115 generations carry a complete call and 0 execute. Effects are strongest when emission format, chat template and parser mismatch; repairing one side alone buys nothing (both 2x2 main effects exactly zero).

Authors concede: a repaired-FC training arm does not yet exist, so training-side remediation is unproven; multi-turn repair restored mechanism but not a significant pass-rate gain (53 to 62, n.s.); the GRPO pass@1 gains of 2.6-2.8 points were not individually significant (p=.064-.078); results are one code task, 100 items, two benchmarks, small model families (Qwen, Llama), single samples at temperature 0, several errata (wrong max_tokens in seven ReAct arms, corrected initiation rates, retracted p-value), and the classifier undercounts via a 4000-char storage cap. Unstated: the phenomenon may be confined to weaker open-weight models that were never trained on tool-call tokens; frontier API models with matched native serving may show no censoring at all; and the paper's benchmark intervention uses a community adapter rather than a verified-optimal one.

A skeptic would say this is a known configuration error (vLLM issue #32926 was already public), dressed as measurement science: the 'discovery' is that Qwen2.5-Coder was not trained on tool-call tokens and must be served with a matching template and parser, which any careful practitioner checks. n=100 per arm on one code-generation task, a single author, temperature-0 single samples, and non-significant headline training gains all limit generalization. The claimed training-loop consequence is demonstrated only for rollouts with zero accepted actions at two scales, without the counterfactual repaired arm that would show RL actually recovers capability when fixed.

Little or nothing yet for the macro AI-economy debate in 12-24 months. The immediate implication is narrow and practical: agent-benchmark comparisons across open-weight models can be confounded by serving-stack configuration, so any enterprise procurement decision or capability ranking that rests on parsed tool-call rates should verify the serving contract first; the released 98-line preflight check makes that cheap. If unmatched interfaces also silently hollow out RL training data, that modestly slows open-weight agent capability catch-up relative to labs with vertically matched stacks.

Technical Capability: Anyone benchmarking or RL-training open-weight agents must treat the model-interface stack as the unit of measurement, since parsed tool-call rates can be zero by construction rather than by capability.

Market & Valuation: Valuations resting on open-weight model 'good enough' agent parity could be mispriced if reported tool-use scores embed silent interface censoring that varies by serving stack.