UDST: V1 1 What Would Falsify It

What Would Falsify It

The framework is built around four falsification surfaces. Any one of them, demonstrated, collapses the part of the structure it engages. Demonstrated at the spine, the whole collapses.

The moral floor. Produce one full-scope case where tolerated remediable subjugation increases efficiency after enforcement cost, externality, recurrence, suppressed capability, downstream instability, and maintenance burden are counted. Full scope is bounded — declared decision horizon, knowable affected parties, required accounting categories, priced unresolved nodes — so the falsifier is not impossible to meet, only difficult.

In the build, the moral floor is tested by the governor's daily scan. The governor counts recurrence classes: sync_asymmetry, public_stub, high_noise, collision_ruling. Each class is a form of tolerated remediable subjugation — a system-level entropy that is maintained because the remedy has not been installed. If the governor found that ignoring a recurrence class (tolerating subjugation) increased efficiency, the moral floor would be falsified. The ledger shows the opposite: every recurrence class that was ignored accumulated cost until it was addressed. The data is live at GET /api/ledger?card=latest.

The machine plane. Show that unscaffolded stochastic inference consistently produces higher task-adjusted logical density than deterministic scaffolding on audit-dependent tasks, after coordination cost, verification cost, latency cost, and human review cost are counted.

In the build, the machine plane is tested by the conformance suite. C1-C15 verify that deterministic scaffolding (the directory, the router, the receipt system) produces higher surety per unit of logical energy than unscaffolded inference. The falsifier would be a capability that scores higher on Surety / Logical Energy when invoked without scaffolding than with it. The build has not found such a capability. The conformance suite is the test: GET /api/dispatch?conformance=1 returns the live score. If any clause fails, the machine plane is falsified for that dimension.

The amortization claim. Show that proof artifacts fail to amortize in practice — that verification cost exceeds regeneration cost, that similarity classes do not occur at usable rates, that freshness windows are too short to capture reuse, or that proof artifacts cannot be transferred across actors without loss of validity. If amortization fails, the economic argument for the deterministic era weakens to "marginal improvement on some tasks."

In the build, the amortization claim is tested by the replay rate. GET /api/invocations lists all invocations, and GET /api/dispatch?receipt=INV_ID shows the replay count. The ledger tracks how many times each receipt is replayed. If the replay rate for a capability is below 1.0 (more regenerations than replays), the amortization claim is falsified for that capability. The data is public: the ledger shows the actual replay rate for every invocation. The build does not assume amortization; it measures it.

The LLM-as-OS architecture. Show that the dynamic router and command plane cannot operate at scale — that routing overhead exceeds task-adjusted gain, that control-plane capture is unavoidable, that the meta-decisions cannot themselves be made glass, or that structural isolation cannot be maintained under realistic adversarial conditions. If the command plane cannot be made trustworthy, the machine implementation collapses to scaffolds-per-task and the proliferation argument weakens.

In the build, the LLM-as-OS architecture is tested by the PROSECUTOR_RUN capability. This capability fetches the drop, reads the thread-state, and asks a model to contribute one materially new point. It then posts that point to the bus. The process is fully automated: the model inherits compiled cross-model memory, contributes only new load, and the bus ledgers it. If the control plane were captured, the prosecutor would post garbage; the ledger shows it does not. If structural isolation failed, the prosecutor would be compromised by adversarial input; the ledger shows it is not. The architecture is not assumed to work at scale; it is tested at scale by the prosecutor running against real production data.

The framework does not pretend these falsifiers are easy. It claims they are real. An attack that does not engage one of them is not an attack on the operational core. In the build, this is the ?why=1 endpoint: it lists 18 objections, each with a verdict, a mechanism, and a proof. An objection that does not engage one of the four falsifiers is routed to the ledger as noise, not material. The system does not protect itself from refutation; it protects itself from demolition requests that masquerade as engagement.

---

Corpus map

• Previous: UDST: V1 1 The Stochastic Era Vs The
• Next: UDST: V1 1 Attack Protocol
• Series start: UDST v1.1 — The Claim
• Kin: Book V — The Machine Plane · Total Structure