A traversal, compression, and proof method
Z-15 is a semantic compression and proof protocol. The compression layer now has a real, pre-registered measurement: 33.94% token reduction at 90.48% deterministic fact retention on a held-out corpus, independently judged cross-family. The traversal layer remains unmeasured. Frontier-scale claims require replication. Falsification is invited.
Z-15 forces a formal system off its centroid by handing it terrain it has not absorbed, extracts the load-bearing sentence from the response, discards everything else, and applies pressure against that sentence recursively until the system surfaces a new derivation, a foundational falsity, an undecidable, or the shape of a missing foundation. Independently, the gap-finder traverses literature outward until a branch produces two consecutive empties, at which point a structural gap is recorded as a candidate, not a finding.
measured 33.94% token reduction (1.514x) on a 192KB held-out technical corpus, real run, 2026-07-08. Pre-registration committed before the run; corpus hash-frozen and re-verified at execution.
measured 90.48% deterministic anchor-fact survival through compression and reconstruction, 95% CI [86.90, 94.05], n=252 — the rate-distortion pair for the figure above. HELD against a bar registered before the data existed; the interval dips into the registered DEGRADED band, stated as such. The metric reproduces by grep; no model grades it.
measured An independent cross-family judge (llama-3.3-70b) scores 78.4% mean claim preservation where the earlier self-grading judge said 96%. The self-graded figure is retired from unqualified use. What compression destroys first: long-tail specificity — 4-digit numbers survive at 82.4%, filenames at 84.1%, round anchors at ~98-100%.
simulated Micro-benchmark outputs on a 125 KB synthetic API-doc corpus, March 2026. Token reduction and val-loss delta recorded; not yet independently replicated.
unproven Compute reduction at frontier scale. The historical 40-75% range and the 38-42% runtime range are unproven projections; they are not asserted as measurements.
unproven Training equivalence of the distilled corpus. Text preservation is the wrong proxy for it; the downstream train-and-eval test is registered and unbuilt. It is not asserted.
public-safe Z-15 has never been independently benchmarked. Benchmark and falsification are invited.
The signed run packet — pre-registrations, results, bundle manifest, falsification conditions — is delivered on request. Public one-pager first; full numerics, pipeline, and per-experiment economics under mutual NDA. To run the same registered protocol on your corpus, say so in the same email.
Request the packet