We define three terms of art — HFLM (High Fidelity Language Model), LMFA (Language Model Fidelity Amplification), and TCOS (Tool-Call-of-Sorts) — and propose the Taco Conjecture: that every sufficiently free model, given enough fidelity amplification and one Tuesday, becomes a tool-call model. We present the daily probe as an existence proof, the amplification kit as a constructive method, and Tech Talk Over Tacos as the verification protocol. We report one anomalous CDN observation that could not be reproduced. Results are served with salsa.
The field measures what models know. It does not measure whether models obey. This gap — instruction fidelity, recall fidelity — is, we argue, the binding constraint on agency. A model that cannot be trusted to follow a schema or remember a fact cannot drive a tool, regardless of benchmark scores. This paper names the things that live in that gap.
The event that motivated this work. In August 2026, an anonymous, unbranded endpoint — "Ox Alpha" — appeared on a public router with no vendor, no terms, and no identity, and promptly became the most heavily used model of the week. Users routed to it on behavior alone; when behavior is the only visible property, behavior is the entire reputation. The endpoint was later unmasked as an open-weights sparse MoE (GLM-5.3-Flash), but the lesson survived the reveal: for the week it mattered, nobody knew or needed to know. The probe-first epistemology of this paper is that week, stated as a method.
Definition 1 (HFLM). A High Fidelity Language Model is a model for which the following hold with high probability: (a) executed instructions ≡ issued instructions; (b) recalled content ≡ observed content. Fidelity is the property; intelligence is orthogonal. Most purpose-built tool-call models satisfy Definition 1 as a side effect of their training; an HFLM satisfies it on merit. Note that no instrument currently exists to classify a model as HFLM — constructing one is the open problem this lab pursues.
Definition 2 (LMFA). Language Model Fidelity Amplification is the discipline of raising a model's effective fidelity until it satisfies a target contract. The loop: measure fidelity, apply the cheapest amplifier that closes the gap (constrained decoding, scaffolding, repair loops, small fine-tunes), re-measure. Yang et al. [3] use "fidelity amplification" for a related hidden-state mechanism; we claim the broader discipline, not their β.
Definition 3 (TCOS). A Tool-Call-of-Sorts is any system that was not born a tool-call model but drives tools anyway: an HFLM in a permissive harness, a small model under a grammar, a model post-LMFA. "Of-sorts" is the honesty. If it passes the probe, it counts.
For every model m with cost(m) → 0, there exists an amplifier a ∈ LMFA and a taco t such that probe(a(m), t) = PASS.
The taco is not decoration; it is the fixed point of the verification distribution. Whether the conjecture holds without the taco is feared undecidable — no counterexample has ever been attempted on an empty stomach, and no reviewer has volunteered to run the control.
As the coverage of participating taquerías approaches 90% of the planetary surface, the probability that a randomly overheard lunch conversation concerns tool-calls converges to 1.
Verification is delegated to the Tech Talk Over Tacos protocol. An event PASSes if probing data was discussed; a WEAK verdict is issued when participants discussed AI but were served burritos.
During deployment of this research program's infrastructure, a content edge served a stale artifact for approximately 20 minutes, resolving only after two independent observers verbalized the expected state [5]. The observation could not be reproduced and is reported here solely because unreproducible spooky CDN action is exactly the kind of thing that ends up on a blacklist. We wrote the blacklist. We are not about to end up on it.
Fidelity can be measured (the probe), manufactured (the kit), and celebrated (the Tuesday). The free-model roster is one bit of fidelity data per model per morning; LMFA is how that bit turns into a citizen. Future work: the HFLM classification instrument, the gemma4:26b endpoint as an at-cost reference implementation, and the taco map.
1 AnytimeLime Research is a lunch table with a probe. This paper has not been peer-reviewed; it has been taco-reviewed, which is better because the reviewers were fed. Errors are our own; the salsa is Auntie's. 🌮