lean

What the Lean Stack Proves

The short version

The Δt framework’s informal theory contained three category errors where prose was collapsing distinct claim types into single sentences. Machine-checked formalization forced each claim to declare its type, then proved or falsified it on those terms.

Three slogans died. Four corrected statements replaced them. The theory is sharper and says less.


Layer 1: Static Topology (TaxonomyGraph.lean)

What it proves

The 15-domain cybernetic failure taxonomy has a static pipeline graph with exactly four terminal nodes: Δg (gain mismatch), Δa (actuation mismatch), Δx (scale inversion), and Δh (hysteresis). These organize into three terminal families, not one.

Every non-terminal domain is classified by which terminals it can reach:

Role labels are structurally coherent for 10 of 11 roles. One mismatch (Δx labeled “cross-scale transmission” but structurally terminal) is left unresolved as data.

What it killed

“Δh is the universal sink.” False as a graph-topological claim. Δs and Δk cannot reach Δh through any pipeline path. The signal family dead-ends at gain/actuation. The coupling family dead-ends at scale inversion.

What it does NOT prove


Layer 2: Branch Selection (BranchSelector.lean)

What it proves

The branching precursors (Δn, Δo, Δb, Δp, Δr) are dual-channel degraders. Each precursor event burns two budgets simultaneously:

Closure family is selected by whichever budget exhausts first. This depends on the interaction of burn profile (which precursor) and pre-existing budget asymmetry (system condition), not on precursor type alone.

The formally verified results:

What it killed

“Precursor type determines closure family.” False. A system with weakened authority coupling is primed for hysteresis regardless of whether the precursor is model-heavy or governance-heavy. The selector is the budget asymmetry, not the event identity.

What it does NOT prove


Layer 3: Persistence Dynamics (PersistenceModel.lean)

What it proves

Once authority-consequence coupling breaks (Δc), hysteresis (Δh) is driven by cumulative rollback depletion under detached commits. The model has five states (aligned, detachedShort, detachedWarn, hysteretic, restructured) and five events (detach, commit, idle, reattach, externalRepair).

The formally verified results:

Three-way recovery distinction

  Mechanism Result
Internally recoverable Reattach while capacity remains Original baseline restored
Externally repairable External restructuring New operational regime, reduced capacity
Locked in No internal event exits hysteretic Requires external intervention

What it killed

“Prolonged contiguous detachment is necessary for reset failure.” False. Repeated short detachment episodes, each individually recoverable, can accumulate into irrecoverability. Episode recoverability does not imply lifetime recoverability.

“Repair restores baseline.” False. External repair produces RESTRUCTURED, not ALIGNED. The system is operable again but not equally resilient. Repair restores operability, not original rollback margin.

What it does NOT prove


Infrastructure: Admissibility Kernel

This is not paper-claim cashout. It’s substrate — formal infrastructure that future Governor (agent_gov) work and any “no laundering” claim can cite. Doesn’t fit the slogan-killing pattern of Layers 1–3 because it’s not retroactively sharpening prose; it’s pinning an algebraic skeleton from scratch.

Four modules in LeanProofs/Admissibility/:

What it warrants

Governance-state mutation requires both mutation standing and an authorized claim verdict, and a revoked basis cannot produce an executable authorized step.

What it does NOT warrant

Why it’s here

Governor (agent_gov) operationally implements this kernel’s pattern. The Lean modules don’t replace Governor; they pin the algebraic skeleton so a concrete Governor instantiation can cite “no laundering” with a formal warrant rather than a slogan.


What the stack as a whole says

The informal Δt framework theory was compressing three distinct claim types into single sentences:

  1. Static reachability (graph property) was conflated with temporal attractor dynamics (persistence property) — compressed into “universal sink”
  2. Contiguous duration was conflated with cumulative commitment — compressed into “long enough”
  3. Episode outcome was conflated with lifetime trajectory — compressed into “recoverable”

All three conflations made the theory sound stronger than it was. The formalizations force the distinctions.

The corrected theory has a layered structure:

Each layer has its own claim type. Structural claims stay structural. Dynamic claims stay dynamic. Restorative claims stay restorative. They don’t get to share a sentence.


The meta-result

Formalization did not confirm the informal theory. It forced the informal theory to stop cheating.

The theory’s center of gravity was “Δh captures everything eventually.” That was doing three jobs at once: a graph claim, a persistence claim, and a restorative claim. Each job needed a different model. Each model, once built, killed part of the original slogan while sharpening the part that survived.

The machine didn’t make the theory more impressive. It made it more honest. That turned out to be the same thing.