// Z-15 / claims

Claims register

Every claim Z-15 makes is classed here. The machine-readable mirror, claims.jsonl, ships with every proof packet. Contradict a row by filing a falsification (see truth engine).

ClaimStatusSourceChannels
Z-15 has never been independently benchmarked. public-safe operator statement site / github / x / hn / email / nda
Z-15 proof engine is under construction. public-safe git log on this repo site / github / x / hn / email / nda
Benchmark and falsification are invited. public-safe /z15/truth-engine site / github / x / hn / email / nda
33.94% token reduction (1.514x) on a 192KB held-out corpus, real run, pre-registered. measured Z15-REAL-RUN-RESULT.md; trace bab2ef11…, 2026-07-08 site / github / x / hn / email / nda
90.48% deterministic anchor-fact survival, 95% CI [86.90, 94.05], n=252 — HELD against a pre-registered bar; CI dips into the registered DEGRADED band. measured Z15-V2-RERUN-RESULT.md; bundle 27f960ae…, 2026-07-09 site / github / x / hn / email / nda
Cross-family judge (llama-3.3-70b) scores 78.4% mean claim preservation where self-grading said 96%. measured z15-v2-judge-verdicts.json, 2026-07-09 site / github / x / hn / email / nda
Compression destroys long-tail specificity first: 4-digit numbers 82.4%, filenames 84.1%, round anchors ~98-100%. measured z15-v2-result.json, numbers_survival_by_digit_count site / github / x / hn / email / nda
Distill-once pays back after ~26 consumptions at the pipeline's own measured cost. measured z15-v2-result.json economics; dollar basis nda-only site / github / x / hn / email / nda (count only)
Training equivalence of the distilled corpus. unproven registered, unbuilt: downstream train-and-eval (PREREG-Z15-V2 §9) never asserted anywhere
96% claim preservation (V1, self-graded). do-not-use superseded by the 78.4% independent figure, 2026-07-09 nda only, with the V2 caveat attached
Latency: -39.9% reduction. do-not-use no source ever existed; removed from / 2026-07-09 forbidden everywhere
59.9% token reduction (micro-benchmark). simulated z15/memo/z15_results.md, March 2026 nda
2.49x RU-ROI (micro-benchmark). simulated z15/memo/z15_results.md, March 2026 nda
40-75% compute reduction at scale. unproven historical projection only nda only when marked unproven
38-42% runtime token reduction at scale. unproven historical projection only nda only when marked unproven
Crossover near 1.9B parameters. needs replication extrapolation only none until reproduced
Every Z-15 proof packet's BUNDLE_SHA256 verifies against the artefact set. public-safe z15_engine/trace.py; tests/test_trace_no_secret_leak.py site / github / x / hn / email / nda
The Z-15 gap-finder requires two consecutive empty searches before declaring a branch closed. public-safe z15_engine/gap_finder.py; tests/test_gap_finder.py site / github / x / hn / email / nda
All Z-15 model calls and GPU rentals default to dry-run; live mode requires explicit env flag plus in-session operator approval. public-safe z15_engine/model_router.py; ops/scripts/z15-gpu-budget-check.sh site / github / x / hn / email / nda
Frontier-scale superiority. do-not-use no measurement exists forbidden everywhere

Status legend

measured a real, pre-registered run produced this number; artefacts are hash-bound and the falsification condition is published.

public-safe may appear on the site, GitHub, X, HN, email, NDA.

simulated exists as a measurement on toy/synthetic data; not asserted as frontier proof.

unproven historical projection or extrapolation; not a measurement.

needs replication requires an independent re-run before publication.

do-not-use never appears in public artefacts.