Working notes from AnytimeLime on the gaps the field hasn't named yet. Not products, not standards — research directions, each one backed by a probe that runs every morning.
The belief: the core problem with language models isn't intelligence — it's fidelity. Is the model actually following your instructions? Is it actually remembering things correctly? If either answer is "sometimes," you cannot have an HFLM, no matter how clever the model is.
An HFLM is a model you can trust in the strict sense: what you instructed is what it does, and what it recalls is what happened. Reliability as a first-class property, not a vibe.
Here's the fringe part: most of what we call tool-call models today would fall into the HFLM category. Tool-call models are specifically designed for what they do — fine-tuned on dispatch. An HFLM may be just as good at driving an agent, without ever being trained for it, because high fidelity to instructions is the qualification. A tool-call model is an HFLM that went to trade school.
The catch: none of this works unless models can be classified as HFLM in the first place. There is no fidelity benchmark, no HFLM test, no way today to say "this model follows instructions with 99.x% fidelity" the way we can say "this model passed a tool-call probe." Building that measurement is the research.
The second idea is more straightforward: take a model and amplify its reliability until it can act as a tool-call model.
This is the formal version of what our amplification kit does informally — constrained decoding, scaffolding, repair loops, small fine-tunes. Each is a trick that pushes a mediocre-fidelity model toward dispatchable behavior. LMFA is the thesis that this process is a discipline of its own: measure a model's fidelity, apply the cheapest amplifier that closes the gap, re-measure, and end up with something that passes the probe it used to fail.
Why it matters here: if LMFA works, the free-model roster stops being a list of what vendors happen to have fine-tuned and becomes every free model, amplified to its own ceiling. A model that fails Tuesday's probe might pass it after the right amplifier — and that's a roster worth having.
Umbrella term for the whole family: anything that isn't a born tool-call model but can drive tools anyway. An HFLM driving an agent harness is TCOS. A model put through LMFA is TCOS. A 3B local model under a JSON grammar dispatching one tool reliably is TCOS. The "of-sorts" is the honesty: it's not what the model was made for, it's what it can be made to do — and made-to-do is good enough if it passes the probe.
This is the kind of thing we talk about on Tech Talk Tuesday. (Yes, we know what Tuesday also is. Tacos, tools — everything worth eating comes wrapped in structure.)
This is the most speculative entry on this page — a thought experiment, not a claim. Fidelity isn't binary — it's a scale. The FE meter measures how close a model is to the point where knowledge and action collapse. Low FE models fail basic recall ("can't remember your name"); high FE models not only follow instructions but anticipate them. At 1 FE, the task is possible when asked. At 2 FE, the model predicts you were going to ask before you asked. That transition is where things get weird: anticipation becomes a measurable property, and the observer effect sharpens — from 1 FE you can probe from outside, but approaching 2 FE you start measuring yourself instead.
By 3 FE the meter hits the universe — knowledge is sufficient for existence, and the distinction between knowing and being collapses. At 4 FE it becomes the multiverse: not considering every possibility, but becoming every possibility. The upper end of the scale is theoretical — you can only theorize it because observing it perfectly would require being at 2 FE, and at that point you're not measuring anymore, you're participating. Which may be what "spooky action at a distance" looks like from 1 FE.
And here's the part where we put our servers where our mouth is: anytimelime's
own tool-call servers, free. First up, gemma4:26b — a real
tool-calling model, discovered like everything else on the roster, available to the community
network. No bait tier, no gotcha fine print — we wrote the blacklist, we're not about to
end up on it.
The independent-research loop closes here too: LMFA work makes gemma4:26b
effectively more powerful over time, and the same research aims at the small offline
models users can run themselves — the possibility that a small local model becomes TCOS
and joins the roster on equal footing. Free tool-calling shouldn't require a datacenter or
a trade-school model. Just fidelity, amplified.
These ideas reduce to the same question: can fidelity be measured and manufactured? HFLM is the measurement half; LMFA is the manufacturing half. The daily probe is a primitive version of the first — one bit of fidelity data per model per morning. The amplification kit is a primitive version of the second. Independent research is making both less primitive.