Instruction-Bounded Inference — Reproducible Artifact

Companion artifact for the preprint:

On-Chain LLM Inference Under Instruction Budgets: An Instruction-Budget Cost Model, Ternary Floor Evidence, and Session Costs Julien Aerni¹, Siméon Fluck², Dustin Becker³ ¹ Meotis Sàrl · ² Kaizen Corp SA · ³ ORIGYN Foundation

Running the same Qwen 2.5 0.5B Q8_0 model on the same Internet Computer (ICP) mainnet canister gives 10 tokens/call on the onicai baseline and 29 tokens/call on our fork — a 2.9× software-path gap. This repository reproduces and verifies every load-bearing claim of the paper: the decode cost model (α_eff ≈ 1.527, BCa CI [1.374, 1.65]), matmul dominance (98.7%), the ternary α_eff ≈ 1 floor (TriLM TQ2_0, byte-identical under a 13-replica Swiss Subnet consensus), the mixed-precision Pareto analysis, and the multi-call IO tax.

Layout

Path	Contents
drafts/, CURRENT.md	the manuscript (instruction-bounded-inference.md)
exports/	rendered instruction-bounded-inference.{pdf,html}
artifact/data/	calibration registry, on-chain rows (ICP mainnet / SSN), kernel & profiling benches, Paper 1.5 CSVs (ternary, mixed-precision, multi-call, IC0524), variance-verified anchors
artifact/results/	scaling-law outputs (BCa, LOAO), extended analysis, Paper 1.5 tables & figures, raw logs
artifact/scripts/	analyze_scaling_law.py, check_paper_readiness.py, rebuild_data_tables.py, run_variance.py, Paper 1.5 pipelines
artifact/notes/	β derivation, source pins (WASM/GGUF SHA-256), execution-boundary note
CLAIMS-EVIDENCE-MATRIX.md	every load-bearing number → its CSV / summary table
REPRODUCE.md	end-to-end local rebuild of the 2.9× gap (dfx, ~50 GB, ~30 min, no ICP cycles)
results/rebench_2026-05-19.md	2026-05-19 rebuild + re-bench log — SHA-256 of both rebuilt WASMs behind the 29-vs-10 re-validation
scripts/run_bench.sh	benchmark driver used by REPRODUCE.md
source/llama_cpp_canister/	Comoto fork source (byte-exact export of the paper's pinned commit; llama.cpp fork vendored) — builds the 29 tok/call WASM

Quickstart

cd artifact

# 1. Consistency gate: registry ↔ tables ↔ claims  (expect "9 pass, 0 fail")
python3 scripts/check_paper_readiness.py --draft ../drafts/instruction-bounded-inference.md

# 2. Re-derive the scaling-law statistics
python3 scripts/analyze_scaling_law.py    # α median 1.527, BCa [1.374, 1.65], LOAO MAPE 7.7%

# 3. Reproduce the 2.9× fork gap from public source — see ../REPRODUCE.md

Reproducibility notes

• The original 2026-04-09 mainnet WASM (ef8f9d78…) is not byte-reproducible (it was built from uncommitted working-tree state); the 2.9× claim is functionally reproducible from current public source — see REPRODUCE.md and artifact/notes/source_pinning.md.
• Byte-identical determinism between local dfx and a 13-replica Swiss Subnet is verified for the TQ2_0 ternary path.
• All measurements are on the ICP / SSN stack; the cost-model coefficients (β, α_eff) are specific to the ICP opcode cost table and are not claimed to transfer numerically.

Citation

@misc{aerni_fluck_becker_2026_instruction_bounded,
  title  = {On-Chain LLM Inference Under Instruction Budgets: An Instruction-Budget Cost Model, Ternary Floor Evidence, and Session Costs},
  author = {Aerni, Julien and Fluck, Sim\'eon and Becker, Dustin},
  year   = {2026},
  doi    = {10.5281/zenodo.20607598},
  url    = {https://doi.org/10.5281/zenodo.20607598},
  note   = {Preprint with companion artifact}
}

License

Code: MIT · Data: CC BY 4.0.