Papers & Open Standards

Preprints, verification packages, and specifications.

Kyle W. Luecke · Driftbound Labs · ORCID 0009-0008-0202-7852

Every study released by the lab fixes its pass, fail, and inconclusive thresholds before outcome data are inspected and reports failed gates, null comparisons, and regime crossovers alongside positive findings. Companion repositories publish frozen preregistrations or pre-specification manifests, SHA-256 digests, OpenTimestamps anchors where established, and reproduction harnesses so every number can be checked from source.

Six manuscripts spanning geophysical clock-integrity bounds, typed runtime boundaries for small language models, oscillatory mixture-of-experts control, network-native computation, physical-layer AI infrastructure bottlenecks, and state-space retraction algebra.

01 · Geophysics · Time Metrology · GWOSC O3

Exterior Clock-Integrity Auditing via Canonical Coherence of the Schumann Resonance Field

Frozen Preregistration · SHA-256 Verified P1, P2 & D1 Supported · D2b Not Supported

Evaluates the global Schumann resonance magnetic field (6–25 Hz) as an exterior physical timing channel across the 3,002 km LIGO Hanford–Livingston (H1–L1) subterranean vault magnetometer baseline during a preregistered six-hour O3 epoch (GPS 1264298000–1264319600). Two-axis canonical coherence recovers the shared wavefield without compass or gain calibration, reaching γ²can(f1) = 0.3291 (z = 24.54 above the off-band empirical null μ0 = 0.0399 ± 0.0118; 49.9× null-corrected vault gain over an unshielded GW170814 pilot) and establishing a six-hour Carter Cramér–Rao timing floor of στ = 76.6 μs (45.7 μs raw).

Scope & failed stability gate — Preregistered gate D2b is not supported: while 11/12 thirty-minute windows were individually significant (p ≤ 0.05) with a physically bounded median lag (−3.85 ms), between-window scatter (MADN = 4.95 ms) exceeded the 0.796 ms stationarity threshold by 6.2× (proposed mechanism: diurnal lightning-source migration; companion WWLLN tests in k-luecke/Geo also not supported). Reported strictly as an exterior audit bound, not a positioning system or GNSS replacement. Acknowledges GWOSC (LIGO/Virgo/KAGRA) and Robert H. Holzworth / WWLLN (http://wwlln.net).

BibTeX
@techreport{luecke2026schumann,
  author      = {Luecke, Kyle W.},
  title       = {Exterior Clock-Integrity Auditing via Canonical Coherence of the Schumann Resonance Field},
  institution = {Driftbound Labs},
  year        = {2026},
  month       = nov,
  number      = {v2.0},
  url         = {https://github.com/k-luecke/time-integrity},
  note        = {ORCID: 0009-0008-0202-7852}
}
02 · Agentic Systems · Runtime Verification · Security

Typed Boundaries for Small-Model Web Agents

5/5 Bitcoin OpenTimestamps (.ots) Confirmed Zero Invalid Actions Executed

Evaluates a non-learned typed runtime boundary (SCP)—combining a compact accessibility-tree projection with a deterministic four-rung guard ladder (D1–D4) and provenance-preserving abstention—on MiniWoB++ under BrowserGym (25 tasks × 10 seeds) and a confirmatory 300-world synthetic cohort. Across Qwen2.5 0.5B, 1.5B, and 3B models, SCP tracks the reference agent on task success (+0.8, +0.4, +0.8 pp) at roughly one-third of exact BPE observation tokens (3.1×–3.8× reduction) with zero invalid actions executed (0/300 committed violations at D4 vs. 152/300 unassisted, and 0 vs. 88 violations against an FP16 teacher-forced confidence baseline).

Mechanism & boundaries — The gate filters and does not repair: synthetic cohort task success was identically 101/300 across every rung; ±10 pp TOST equivalence was supported at 0.5B ([−2.4, +4.0] pp), marginal at 1.5B, and unbounded at 3B; preregistered superiority over raw accessibility trees was not supported (+6.0 pp, p = 0.24 vs. α = 0.01). On Qwen3.5-2B, an attribution ladder shows action-schema typing (+22.4 pp) accounts for completion lift rather than observation compression alone.

BibTeX
@techreport{luecke2026typed,
  author      = {Luecke, Kyle W.},
  title       = {Typed Boundaries for Small-Model Web Agents: Tracking the Reference Agent Across Small Scales at a Third of the Tokens with No Invalid Actions Executed},
  institution = {Driftbound Labs},
  year        = {2026},
  month       = sep,
  url         = {https://github.com/k-luecke/scp-typed-boundaries},
  note        = {5/5 Bitcoin OpenTimestamps preregistration proofs confirmed. ORCID: 0009-0008-0202-7852}
}
03 · Computational Neuroscience · Mixture-of-Experts

Phase Selects; It Does Not Amplify: A Preregistered Test of Oscillatory Control as Expert Selection

12/12 Bitcoin OpenTimestamps (.ots) Confirmed Stages T5–T13 Preregistered

Tests across twelve preregistered experimental stages whether an oscillatory division of labor—slow tonic selection stencils paired with phase-tagged expert superposition—does useful work inside a small Mixture-of-Experts transformer. Fixed-dynamics coupled units hold an expert-selection regime after control input stops (Thold = 26.8 tokens at n = 8; 65.0 tokens at n = 3 with ρeff = 0.989 vs. <2 tokens stateless), obey cosine lock-in demodulation (X(Δφ) ≈ max(0, cos Δφ)), and act as a single-pass expert-attribution medium in looped MoE under matched compute (rφ = 0.869).

Falsifications & process record — Coherent superposition produced no ensemble amplification (Gmedium = 0.044 vs. Gsum = 0.251); under continuous evidence, stateless lookup matched or outperformed persistent coordination; across mid-gap lures (T12b), held state beat one-step memory (0.6525 vs. 0.2987) but tied k = 3 majority smoothing (0.6390); and stage T10 was closed as not informatively tested with its classifier claim withdrawn due to a disclosed tag-placement side-channel.

BibTeX
@techreport{luecke2026phase,
  author      = {Luecke, Kyle W.},
  title       = {Phase Selects; It Does Not Amplify: A Preregistered Test of Oscillatory Control as Expert Selection},
  institution = {Driftbound Labs},
  year        = {2026},
  month       = sep,
  url         = {https://github.com/k-luecke/tonic-gating},
  note        = {12/12 Bitcoin OpenTimestamps preregistration proofs confirmed. ORCID: 0009-0008-0202-7852}
}
04 · Distributed Systems · Query Complexity · Protocols

Network-Native Computation: Foundations, Structural Leverage, External Validity, and the Limits of Prospective Governance

8 Reproducible Micro-Experiments Theorem 1 Lower Bound + Exp08 Negative Result

Investigates when ordinary, RFC-compliant protocol transitions and continuously maintained substrate invariants evaluate computational distinctions at lower physical cost than local execution without deploying custom code into network nodes. Across eight experiments: DNS DNAME/CNAME rewriting evaluates finite automata with 100% agreement but runs 153.6×–268.6× slower than local CPU (Exp01); Merkle frontier reconciliation achieves 25.2× wall speedup and 3,158× wire reduction at N = 218, k = 1 while explicit mutation logs (Δ) beat NNC by 8.9× when available (Exp02–Exp04); and geometry-dependent frontier scaling survives stock Git v2.55.0 (23.8× wire cut; |F| = 19 clustered vs. 34 scattered at k = 16), 6-peer IPFS Bitswap + Kademlia DHT swarms (13.9×–48.5× wire cut), and AO/Arweave checkpoint-witness emulation (Exp05–Exp07).

Prospective governance boundary — In Experiment 08 (200 stratified held-out workloads, probe budget B ≤ 8), zero-hindsight controllers failed pre-specified thresholds (ηt = 1.803 > 1.25; catastrophic misroute rate 11.0% > 5.0%) and lost to unconditional bulk streaming (ηt = 1.361), explained by Theorem 1, the Merkle Ambiguity Lemma, and tail-loss asymmetry (27.4× vs. 3.4×).

BibTeX
@techreport{luecke2026network,
  author      = {Luecke, Kyle W.},
  title       = {Network-Native Computation: Foundations, Structural Leverage, External Validity, and the Limits of Prospective Governance},
  institution = {Driftbound Labs},
  year        = {2026},
  month       = oct,
  url         = {https://github.com/k-luecke/network-native-computation},
  note        = {ORCID: 0009-0008-0202-7852}
}
05 · Infrastructure Research · Power Grids · Survival Analysis

The Physical Layer of the AI Infrastructure Buildout

Hash-Locked Cohort Manifests 31 Builds · 2,184 MW Audited

Reconstructs point-in-time utility interconnection dockets, municipal structural permits, and statutory 811 excavation telemetry across 31 evaluable North American hyperscale data centers and commercial semiconductor fabs (2,184 MW announced load). Physical or regulatory divergence preceded corporate schedule revision in 9 of 9 resolved delay cases with an arithmetic median lead of 211 calendar days (IQR 200–225 days); >90-day slip scaled monotonically across four utility interconnection states (0.0%, 37.5%, 100.0%, 100.0%); only 1,128 MW (51.6%) cleared a three-condition execution gate; right-censored Kaplan–Meier substation energization reached a 112.0-week step median across 18 transitions; and point-in-time physical features reduced Brier calibration error by 84.5% (0.0383 vs. 0.2476) across 8 of 8 held-out test builds.

Evidentiary boundary — Cell sizes in the most constrained queue tiers (N = 5 and N = 4) and the held-out replay partition (N = 8) are modest, and the Brier evaluation is a point-in-time retrospective replay with zero live prospective resolutions claimed in this release.

BibTeX
@techreport{luecke2026physical,
  author      = {Luecke, Kyle W.},
  title       = {The Physical Layer of the AI Infrastructure Buildout},
  institution = {Driftbound Labs},
  type        = {Institutional Research Whitepaper},
  year        = {2026},
  month       = sep,
  url         = {https://driftboundlabs.com/papers/AI_Infrastructure_Buildout_Whitepaper.pdf},
  note        = {ORCID: 0009-0008-0202-7852}
}
06 · Symbolic Algebra · Concurrent Systems · Program Synthesis

Autonomous Discovery of Idempotent State-Space Retractions and Conditional Commutation Manifolds (FNWD v0.4)

6-Gate Suite + 12-Operator Blind Audit Zero-Dependency Python Package (<5s Replay)

Introduces Formal Normalizer–Workload Discovery (FNWD v0.4) to identify enabler–optimizer compositions (E ⊗ O) pairing an observational state-space retraction E (E² = E, &mathcal;O ˆ E = &mathcal;O) with a conditional commutator O valid on Fix(E). Full-VM replay across 127 idempotents over 𝔽40934 × 𝔽40938 × &mathcal;M64 falsifies unconstrained right-zero bands and constructs the 115-class quotient Projection Poset (P, ≤) (18 strict chains; 100.0% zero-probe deductive prediction across 100 pairs), while passing six pre-specified verification gates (trace equivalence, 30/30 black-box linearizability fault recall, −94.3% factorial lock reduction, paired permutation padj = 0.0006, and 12-operator blind triage promoting 3/3 valid and 0/9 unsound operators).

Algorithmic crossover boundary — While Domain 1 (Affine Automata, 43.1×) and Domain 2 (CRDT OR-Set, 1.68×) accelerate wall-clock execution, Domain 3 (DSW Balanced Search Tree) reduces locks by −93.3% at N = 150 but incurs a statistically significant +69.6% latency increase because Θ(N) whole-tree vine rebalancing exceeds lock-avoidance savings below the large-tree crossover.

BibTeX
@techreport{luecke2026fnwd,
  author      = {Luecke, Kyle W.},
  title       = {Autonomous Discovery of Idempotent State-Space Retractions and Conditional Commutation Manifolds (FNWD v0.4)},
  institution = {Driftbound Labs},
  year        = {2026},
  month       = oct,
  number      = {v0.4.0},
  url         = {https://github.com/k-luecke/fnwd-v04-retraction},
  note        = {ORCID: 0009-0008-0202-7852}
}
S1 · Cryptographic Provenance · Open MIT Standard (Release v0.2.0)

Paxiom Receipts (paxiom.receipt.v0) & x402 Delivery-Receipt Extension

JCS / Raw SHA-256 + Ed25519 npm · PyPI · MCP Registry

An open, vendor-neutral cryptographic receipt format (Payment-Receipt) binding HTTP requests, canonical JSON (jcs-sha256) or binary (raw-sha256) outputs, seller key discovery (/.well-known/x402-keys.json), and CAIP-2 / x402 payment references under Ed25519 signatures. Offline checks verify integrity without vendor callbacks, reject unauthorized signer substitution as unauthorized_issuer, and label unchecked settlement claims explicitly as claimed_not_checked.

All manuscripts are typeset from version-controlled Markdown and BibTeX sources with vector figures and reproducible verification scripts. Pre-commitment hashes, OpenTimestamps Bitcoin calendar proofs, and negative-control audits are preserved in each companion repository.