Work

TRIADE: Enabling LLM Agents to Write Native Code by Hand Under Signed-Intent Governance (Academic Edition v0.22.0) — manuscript, data tables, complete sources and verification evidence

Year: 2026
Type: article
Abstract: The academic edition v0.22.0 of the TRIADE single-case study: a development protocol under which an LLM agent writes native x86-64, RISC-V and TLS 1.3 code directly as assembly notation, governed by signed layer-I specifications. This version succeeds v0.21.0 under the same concept DOI and carries the eleventh instruction wave. (1) The SIMD corridors: the memchr+memcmp hybrid search at 704 cycles per absent 4096/16 (25.9x over the sealed GPR path, inside the class of the library SIMD primitives), the SHA-NI compression corps at 128.5 cycles per block (6.5x), runtime dispatch with baseline fallback and per-session hoisting, on a 559-vector hardened differential (buffer-tail, overlapping, lengths 1-33, page crossing, multi-occurrence, 512 property-based draws) at zero divergence. (2) The AES-NI carrier loop restructured to 7.09 cycles per block envelope (the 12-cycle target met, 7.1x over the v2 carrier), a round-trip cache keyed by notation+engine+profile (680x at opening, fail-closed), incremental crystallisation over the dependency graph (-54 percent at a real diff, the 1/10 audit maintained as a named probabilistic net), and the 37-second governed session decomposed by control (oracles 81 percent). (3) The fourth arm C+G+ (gcc-O2 with governance and the full verification regime at equal oracle volumes): T 39.39 s, C+G+ 18.68 s, C+G 2.18 s, C 1.97 s — the proper notation effect confirmed at 20.72 s (55.7 percent), the verification-regime effect at 16.50 s (44.3 percent), governance at 0.213 s. (4) Lifter soundness validated against observed Unicorn accesses per instruction (942 distinct encodings at zero divergence, the 155 mutant divergences all outside the validated vocabulary — fail-closed domain guard), GHASH H covered as a secret (no tainted branch or address), non-interference re-discharged on 16 blocks across seven families with named holes. (5) The template-versioned edit operation (1106 to 674 native tokens, the natural-language prose down 57 percent), hierarchical localisation (1053 to 627), skeleton-based addition (3074 to 2200, 3/18 feasible — the gap named, never re-drawn). (6) The vault hardened on laboratory copies: ciphertext-fingerprint verification before AEAD (corruption refused without destruction, falsification at valid naive fingerprint destroyed with PV), epoch-bound trap-value derivation, the Argon2 re-parameterisation ceremony with a 60-bit passphrase entropy floor, phrase change under sequestration, the RAM-guard block specification, epoch-witness anchoring, the FIDO2 hmac-secret prototype — with a z3 state-machine model proving four invariants. (7) The six pre-registered kits for the awaiting lines (fence.i on RV64 silicon, the third machine, dudect on two machines with a pre-registered peak investigation, the stratified H7 campaign at 14 strates, the second annotator with a self-tested kappa calculator, the CF3 Tamarin model) — hashed before any measurement, no result produced without the developer's gesture. (8) The D81 delivery instruction purged from the operational pieces (the sequestered 5-of-3 guard active, the deposit public): the Book v1.6 and the specification v0.15.7, with the runbook simulation proving no path commits a phrase. The inter-piece linter is extended to every numeric token of the prose (5,026 tokens, four coverage classes, fail-closed, a lived double negative control).
Language: English
Open Access status: green
OpenAlex ID: W7236717666
FWCI:
Cites:
Cited by:
Related to: 10