The argument, briefly
Classical software spent fifty years at war with ambiguity, and won — at the price of brittleness. Learned systems inverted the bargain: their representations hold many meanings in superposition, and that polysemy is precisely why they generalize. But the infrastructure around them is still the old regime, and most production failures live on the seam where the two meet. That seam is our first product, Digest.
Meanwhile the model's categories were never designed. They are sediment — the compressed record of who wrote the corpus, when, and with which blind spots — presenting itself afterward as neutral knowledge. Making that sediment visible and contestable is our second product, Strata.
And a system that can only accumulate is not learning; it is hoarding. Forgetting — in the weights, in the context window, in the codebase — is a faculty that current stacks simply do not have. Building it is our third product, Lethe.
The bibliography we actually use
We name our sources because clause II of our own manifesto demands it. Three lineages, honestly credited and freely bent to our purposes:
Antifragility is Nassim Taleb's term for systems that gain from disorder rather than merely surviving it. We push it one step: in learned systems, the disorder that strengthens is semantic — contested meanings, hedged answers, disagreeing samples. Hence polysemic antifragility.
Organic ontology leans on the Marxist tradition: meaning as relationally constituted (nothing has intrinsic sense, only position in a field — which is, almost literally, an embedding space) and as historically produced (categories as congealed labor). The provenance label is commodity-fetishism critique turned into a struct.
Active forgetting is Nietzsche's: forgetting not as memory failing but as a positive, plastic force — the precondition of health and of any new beginning. Lethe's mandatory TTLs are that thought with a type signature.
None of these thinkers would fully endorse what we've done with them. That's fine. Sediment is there to be metabolized.
How we work
The open-source cores are dependency-free by policy — you should be able to read every line you put on your seam. Every repair is named, every answer sources its layer, every memory knows its exit. We eat our own philosophy: the proofs of concept are reviewed against the same standards they preach, and every claim on this site is meant to be checkable against the repositories — call out any that is not.
We're a small team and intend to stay legible. If you want the diagnostic version of this page — which of the three deaths your system is dying — book forty-five minutes.
We do not ask our machines to be correct. We ask them to be alive enough to be wrong, rooted enough to be answerable, and free enough to begin again.