Wyrd Labs gives unconventional questions a serious engineering trial.
The divisions commercialize only the ideas that survive evidence. Everything on this page is labelled by what it currently is — an active program, a prototype, a hypothesis under test, or a recorded option.
Innocence Translation.
It does not mean turning dangerous technology into a civilian product. It is a discovery method that protects curiosity without protecting weak ideas from evidence.
Innocence asks the question. Expertise tries to kill it. Proof decides.
What is being tested, and at what class.
ICARUS ForgeGenerate, test and compare governed agents or policies.Research / integration
Can agent and policy variants be generated, evaluated and selected under explicit criteria rather than hand-authored one at a time?
DAEDALUS ForgeSimulate plans, failures, counterfactuals and deployment blueprints.Architecture / research
Can rehearsal against a twin become a normal, cheap step between idea and pilot — instead of an expensive special project?
NINA FoundryGenerate and benchmark optimization methods under constraints.Research prototype
Does a portfolio of generated methods, verified by an assurance layer, beat a single brittle winner on real domain constraints?
The earlier framing that a language model simply "beats solvers" is retired. Prior measurement was found unreliable; the instrument is being repaired before any performance claim.
Substrate HypothesisDoes nested constrained evolution produce useful reusable methods?Scientific research
A hypothesis, not a result. It is pre-registered internally and reported as unproven until an instrument we trust says otherwise.
Agentic geneticsVariation, selection, mutation and evaluation of agent strategies.Research prototype
Explored under controlled evaluation. The hypothesis is not proven.
Engine CorpusReusable library of solvers, classifiers, planners and specialist engines.Platform direction
Every engine needs an owner, a benchmark, tests and a consuming product. Otherwise it is not in the corpus.
Sovereign coding stackLocal or air-gapped coding agents with repo truth, proof and governed handoff.Active R&D
Repo Truth Layer, coding proof packs, governed handoff between external agents, and time-boxed execution leases through PEAOS.
Proof systemsReceipts, ledgers, attestations, transparency, evidence packaging.Active software
Signed evidence, hash chains, threshold controls, Merkle anchoring, and proof packs that a third party can check without trusting us.
OSS intakeResearch, license, provenance, tailor, govern and integrate external open source.Operating method
Using open source responsibly is a licensing, provenance and maintenance commitment, not a download.
What we explore.
Grouped by field. These are recorded directions — expansion optionality, not present commercial promises. Several will be killed by the Accuser's Office, and that is the point.
It is not a claim that any of these exist as products. And where a recorded idea has been parked or rejected — a from-scratch OS kernel, a phone manufactured as version one, a weaponized RADOME, an agent currency before authorization — it stays parked. The option portfolio demonstrates reach. It does not replace focus.
Conflict as a quality mechanism.
An idea is not protected because it is ours. It is attacked, repaired, retested — and then it survives or it dies.
The survivors go to the divisions. The rest are recorded and left alone.
If your institution has a hard problem and real constraints, that is exactly what the method needs.