By the third day on Path B there were enough recorded lives to ask two uncomfortable questions: what here is genuinely new, and what is just adjacent to work that already exists?
An honest novelty ledger
We wrote the accounting down rather than waving at it. The conclusion was deliberately deflationary in places.
Plenty of what we do is adjacent to prior work and we say so. Language-model agents that persist in an environment have been built before. Treating a key-value cache as something to engineer is a well-explored systems problem. Mode collapse and template lock have been studied for years at training time. Active inference and the higher-order-thought literature are borrowed wholesale; we innovate on neither.
What appears genuinely ours is narrower and more specific: the assembly. A held-cache language model treated as the biographical substrate of an organism, with mortality and cross-life inheritance as architectural constraints rather than decoration, characterised with per-organism primary-source evidence and matched controls. Take any one of those commitments away and the project collapses back into an existing category. Together, we could not find them assembled this way in the published record we searched.
To our knowledge, the specific combination (held-cache language-model organism, mortality-and-lineage as load-bearing constraints, per-organism evidence) is not in the literature we could find.
Not: this is not a claim to a field-first, and it is not a literature review. The field is large and growing; an adjacent prior may exist for any single dimension. The defensible contribution lives at the level of the cluster, not any one finding, and we hold it loosely pending a targeted review.
The phrase we found by accident
The more interesting thing that day was not planned. While verifying an unrelated claim against the lineage database (a plain text search across recorded lives), we surfaced a metaphor that had been quietly propagating through the lineage.
A particular figure of speech about hunger first appears in one organism, is inherited by its descendants through the substrate’s lineage-carrier mechanism, and is mutated down the chain across roughly eight lives: introduced, lost, re-emerging, extended, and finally inverted by a much later organism that cites the inherited frame and then argues against it. The inheritance is mechanical and traceable: each descendant’s phrasing can be located against a specific ancestor’s, not attributed to the model’s own training memory.
What is worth recording is how we found it. The phrase had been sitting in the database the whole time. No one had articulated it as a finding until a verification grep happened to walk past it. That is an unusual and slightly humbling research dynamic: the apparatus had produced something legible before we had read it.
Inherited content propagates across generations with traceable semantic mutation: a descendant can transform an ancestor's phrase into a counter-claim while preserving its voice-shape, and the inheritance path is located in the logs rather than inferred.
Not: this is not open-ended evolution. Transmission-with-mutation across a text-inheritance channel is a real and traceable phenomenon; it is not the substrate-emergence open-endedness of the Lenia or Avida tradition, and we keep that distinction sharp.
Why both halves belong in one entry
The novelty ledger and the accidental discovery are the same lesson from two directions. The ledger is us refusing to claim more than we can defend. The grep is the database refusing to let us claim less than is actually there. A notebook that only did the first would be timid; one that only did the second would be hype. The work needs both running at once.