The Deep Fold Experiment

oxexp-deep-fold · the unlearning experiment

The discipline. The fold is the Ducci map |diff(v)| — the operation that discards the sign field at every step, “as if there is no coherence tax.” Here the tax is counted exactly: the unlearning number λ(u) = log₂ of the exact number of signed pre-images the fold has collapsed into an operated state (cycle dynamic programming; no sampling). The calibration facts, recomputed by the engine on request: lap 4’s first fold carries 53.6 bits of discarded signed history; the grown (prime-indicator) fold on its own 0/1 ambient carries exactly one bit — the seed and its global flip: the only unlearnable content of a grown fold is the orientation, the decision itself. Three families walk the same ring under the same fold: P (the π laps, the Deep Pi seeds), G (prime-indicator windows), L (the deep line’s own learner bias — the register whose −1/0 is that experiment’s surplus-walk law). Five hypotheses are registered before each rung, THE DECISION BIT among them (every grown fold exactly one bit — a single counterexample is the headline), and the tax–fate association is landed either way. THE FIREWALL binds every panel: no claim on any Millennium Prize problem, ever; no claim about π’s normality; no claim over machine learning’s own results, quantum computing, physical cosmology, quantum mechanics, consciousness, or phenomenal experience, ever. The commissioning frame is the operator’s, fenced in the design of record (oxexp-dfx-design.md). Acts, never eyes. The kin pages: /dlx · /dpx; the readable companion: The Feynman Nines on the SX surface.

The gate — honest acceptance

The observable, the families, the protocol, the engine’s calibration witness (the spike counts recomputed exactly), and the ladder’s standing — derived from the engine’s own acceptance, never asserted here.

The ladder — registrations, runs, and verdicts

One rung per seeds-per-family count (64 · 128 · 256 · 512 · 1024). Each rung shows its registration (the hypotheses and their frozen bounds) and its results (the verdicts as landed, the family medians, the named cases). The acts are the operator’s: register writes the hypotheses; run builds the three families and counts every seed’s unlearning numbers exactly; the engine refuses out-of-order acts. The deep rungs take minutes — the page waits.

the ladder loads on page open; the button re-reads

The external anchor

A rung’s records (registration first, then results) to the immutable EXTERNAL record (mldata CKAN when configured; the dummy-node fallback otherwise) — the archive record is the citable object, never this host.

oxexp-deep-fold · engine daemon/dfx.py (exact big-integer pre-image counting; the walks; the families; the order enforced fail-closed; the registration embeds the frozen parameters and the engine’s own sha) · the design of record ooi_network/docs/oxexp-dfx-design.md · the CLI wrapper remains beside the plugin for shell use · hypothesis before run, always; refutations displayed, never hidden.