Models

Why it is model-independent by design

1. The genome is data. The log, the replay and the genome contain no model output that the platform relies on. The learning rule credits the genes that were in the prompt, from the platform’s own record. It never uses the model’s account of what it followed. A model that writes a poor report, or none, costs only the review hints.
2. The outline is separate from the text. The compiler first decides what goes in. A dialect then writes it out.
3. There are four dialects. gemini, claude, gpt and open. See Portability.
4. The model is a parameter. In a run, the model is chosen by the run path, not by the genome. In the benchmark, one setting names it.

When a model is replaced, the same genome compiles for the new one. The platform also uses no provider-side tuning, so nothing it controls lives inside the model.

What is tested and what is not

Tested and untested claims about models

Claim	Where it stands
Each dialect renders, escapes special characters and ends with the same report instruction	Tested by conformance and property tests. Built, flag off by default G16
One extractor reads the report from every dialect	Tested by the same tests.
An agent in the IntelliNix platform can use a dialect other than gemini	No. IntelliNix runs use gemini only. The other three are reachable only through the public API.
The genome improves results on Gemini Flash-class models	Measured in a simulation, in three run sets. Measured M01 See Research.
The genome improves results on Claude, GPT, Gemma, Llama or other models	Not tested. The benchmark can call only the Gemini client. Adding other providers is Specified, not built G82 .
The reading rules work on other families	Not tested. They were shortened against Gemini output. The other dialects reuse them in other words, and no benchmark run has measured them.
A small model writes a usable report block	Not tested. The learning rule does not need the report, so a model that ignores it still teaches the genome through customer outcomes.
A study across the four dialects and several families	Research direction R10

The model names in the three run sets are on the Research page, with the caveat on set B’s model.

Longer context windows

The genome’s size limits are constants in the code, not properties of a model. The default budget is 6,000 characters. A per-agent budget is limited to between 500 and 12,000. At most 200 genes and 30 observations are kept, and 8 observations are shown.

A longer window removes the reason to keep these limits low. It does not remove the reason to choose. An irrelevant or conflicting gene can mislead a model however much room there is. A large window also makes the simplest alternative cheap: put every past statement in the prompt. That control (the memory_list arm) exists in the harness and has never been run on a live model. If it matched the genome, the case for the genome’s machinery would be weaker. The test that would settle this is a live run with that arm and the limits raised on each model of interest.

Cheaper tokens

The measured cost is prompt characters, and prompt tokens in one run set. About 1,100 of the genome’s characters are fixed reading rules. Cheaper tokens lower the price of that scaffold. They do not change the latency of a longer prompt, the share of a model’s attention spent on reading rules, or the size of the stable part that a prompt cache can reuse.

An optional layout splits the section into a stable prefix and a changing tail so that a provider’s prefix cache can hit across runs. Built, flag off by default G17 The design target was a section at most 2% longer than the single layout. That is a target only. No committed result file measures the layout’s size overhead or cache hits.

Smaller and open models

The open dialect is compact on purpose, for small models. Compiling is cheap and runs in-process, with no model call. What the code cannot tell is whether a small model follows tagged reading rules, precedence lines and a 6,000-character section the way a hosted model does. The study that would answer it needs other providers in the benchmark first.

Tool-using models

In the IntelliNix run path, the genome is compiled once per run and reused for every model call in the run’s tool loop. External stacks compile per call. The genome holds no facts: tools supply facts and the genome supplies preferences, so the two fit together. Two limits follow. A preference learned in the middle of a run reaches only the next run. And an agent whose tools can act on the world is treated as high stakes, so the leave-one-out experiment never runs for it. The benchmark has no tool-using personas, so there is no measurement here.

Testing your own model

The benchmark can be re-run with another model name, at the cost of about 2,700 paid model calls for one run set. The result is a new, separately labelled run set with its model name, code revision and working-tree state. It is never pooled with sets A, B or C. A new live run will not reproduce an earlier set number for number, because sampling at temperature 0.2 is not seeded and a hosted model can change under the same name. Whatever it shows applies to how that model reads this prompt under this checker. It says nothing about how customers would rate the outputs.