Evidence
How Thornveil proves it works
Every claim Thornveil makes is backed by working code, measured benchmarks, and where appropriate, cryptographically auditable reduction-to-practice. Five papers below. Public companion repositories on github.com/thornveil-ai for the rest.
Research papers
Peer-reviewable papers documenting the algorithms behind Thornveil's systems. HawkStack topology paper preparing arXiv submission; remaining four are technical reports.
Selective-Buffer Streaming Safety for AI Coding Agents
Zeroth-Order Preference Optimization on 100B+ Quantized MoE Models via Live Inference API
Compressed KV Cache Attention as a Plugin Backend
Speculative Decoding Fails on Sparse MoE: A Negative Result and Practical Multi-Model Cascade Alternative
RigRun: Complete Local AI Infrastructure on a Single GPU
Trade-secret-grade IP across the portfolio
Each Thornveil system carries one or more proprietary subsystems where the engineering work materially differentiates it from open-source alternatives. Listed below: the named subsystem, what makes it defensible, and the cryptographic or empirical evidence that backs the claim.
Substitute-on-failure MoE expert dispatch
Day-27 RTP with HMAC-chained audit log, 2026-05-08
Scope-as-code engagement compilation + 13-check Signet gate
Autonomous AD takeover validation with auditable replay package, 2026-05-16
Compile-time classification-gated routing with cross-domain guard
Type-system enforced; 44 NIST 800-53 controls implemented
8-signal confidence pipeline with adversarial self-audit
~4,220 LOC across 8 signals; cross-app parity in Go/Dart/TS
HMAC-chained tamper-evident audit ledger with RFC 3161 anchoring
2,400 LOC across chain.py + verifier.py + compactor.py; 7-state break taxonomy
Compute-aware neural topology recipe (3-parameter)
15 verified zoo checkpoints, R²=0.9895 power-law scaling fit on NUDT-SIRST
24-layer adaptive control loop (PID + Holt + UCB1 + ε-greedy)
Race-clean CI across Linux/macOS/Windows; PAVA calibration validated vs sklearn
Phase 10 trade-secret primitives — attest, secdef, beacon, vehicle-id, log
~5,500 LOC across 5 crates; not shipped in ArduPilot or PX4
14-step agent-production pipeline with Opus calibration + 5-probe adversarial hardening
9 production domain-expert agents generated by the pipeline