Basis for the agentic-coding dataset: version-control survival as a time-validated
quality label; shapes-that-last; the 3 co-registered layers (transcripts/diffs/
outcomes) + human-vs-AI axis; construction pipeline (extract/normalize/link/redact/
format) with fleet aggregation + HF backup; SFT/preference/RL/survival-holdout.
Co-authored-by: Hanzo Dev <[email protected]>
Measured head-to-head vs llama.cpp on the same silicon: the flash-attn dense-prefill
result on GB10 (+42%, still 0.77-0.84x, stated honestly as the one trailing axis),
the CUDA decode-graph / MoE-prefill / GDN-kernel leads, AMD gfx1151 prefill+decode,
and cross-backend decode parity. Compiles clean (13 pp).
AICoinMiner -> AIMiner across prose, status table, and bibliography.
Contract paths updated to contracts/ai/{mining,compute,token}/ (no
ai/thinking/). 'Thinking' no longer appears as any contract/protocol/
type name; 'Thinking Chains' survives only as the loose category phrase.
No change to the Hamiltonian Market Maker math, theorems, test counts,
or technical content. PoT = Proof of Training (a PoAI proof) untouched.
Recompiled clean with latexmk (23 pages).
Part I (proof) tightened and grounded line-for-line in shipped code:
- correct stale test counts to measured values (poi.rs 7, poi_graph 9,
poi_transcript 8, Go crypto/poi 22 funcs; forge: MinerStakeRegistry 8,
AICoinMiner 12, AIComputeRegistry 6, AICoin 11)
- add typed graph-composition soundness (Thm: well-formed + locally-correct
=> committed output equals deterministic forward pass; poi_graph.rs)
- add information-theoretic / post-quantum soundness proposition + PQ envelope
(ML-DSA/SLH-DSA/X-Wing/P3Q)
Part II (price) — the Hamiltonian Market Maker, fully derived and proven:
- phase space (q,p), cost potential C, Hamiltonian H = pq - C(q) = conjugate
C*(p) on the quote manifold (bonded value; no unforced 'physics' terms)
- Thm path-independence; Thm zero-cost round-trip / no in-curve arbitrage;
Thm bounded subsidy with LMSR b·ln N as the entropic special case;
Thm monotone price-supply law (the PoAI coupling); Prop convex slippage;
Prop Uniswap divergence-loss; multi-asset generalization
- LMSR (Hanson) and Uniswap x·y=k shown as two potentials of one frame
- bind the curve to verified supply: MinerStakeRegistry (bonded/slashable),
AICoinMiner no-proof-no-mint, AIComputeRegistry network-wide no-double-mint,
AICoin capped halving issuance — all read from luxfi/standard contracts/ai/thinking
Integrity fixes during review: removed a FALSE angular-momentum conservation
claim (the {L_ij,H}=0 identity does not hold for a cost-function maker; replaced
with the genuine discrete permutation symmetry + an explicit remark on why the
tempting claim is false); made the trade-vs-relaxation dynamics honest.
No fabricated benchmarks. Every empirical number is a measured test count; every
HMM claim is a theorem with a proof. Compiles clean: 23 pages, 0 errors.
Standalone paper: PoAI builds ON the Hanzo engine — the proof is emitted as a
side-effect of native inference/training/embedding execution. Freivalds over the
exact int8 accumulator in F_p (p=2^61-1), the four proofs (PoE/PoI/PoT/PoC) over
one workloadType primitive, determinism prerequisites, commit-now/verify-later
ActivationTraceMMR, the measured-weights reportData binding, floating-point by
tier, architecture-family coverage, and ownership. Honest status incl. the
node-binary support rows (hanzo-engine v1.2.5 / hanzod / luxd / zood). 10pp.
The DOKS-direct deploy job was wrong (CI must not touch the cluster directly)
and also lacked the token. Keep the working self-hosted build→GHCR half; the
native operator rollout is handled by platform.hanzo.ai via the Hanzo Platform
GitHub App + .platform.yml (wiring tracked separately).
Replace the GitHub Pages publish (broken since 2026-04-10 on the org-wide
Actions artifact-storage quota) with a standard GHA workflow that runs entirely
on Hanzo infra: build the Next.js static export into a hanzoai/static image on
the self-hosted hanzo-build-linux-amd64 pool, push to ghcr.io/hanzoai/papers,
then roll it onto the operator-managed `papers` Service CR (hanzoai/ingress +
static plugin) from hanzo-deploy-linux-amd64. No GitHub-hosted builders, no
Pages artifact quota, no kaniko/PAT. DNS cutover (Cloudflare → hanzo-k8s LB) is
the remaining one-time step.
Nav link + PDF bookmark still said 'Sovereign vs. Closed' (pre-rename, b314046).
Aligned both to the paper's current title and rebuilt the 196-page bundle so it
carries the corrected ~90% vs-claude paper.
The 'Cost at scale' section claimed cloud spend 'fell approximately 97%',
contradicting this paper's own ~90% ($35K->$3.5K/mo, ~$380K/yr) and the
canonical cloud-economics staging (achieved ~90%; ~97%/$1,200/mo is the
not-yet-reached target). Also matches the sentence's own 'order-of-magnitude'
framing. Rebuilt PDF + synced to site/public/pdfs (deploys to papers.hanzo.ai).
- hanzo-vs-claude: prominent 'one engineer on Hanzo stack vs team on Claude' section
(anonymized audit: ~250-900x lower $/LoC) + theoretical mechanism + switch CTA.
- one-pager: cloud-cost aligned to ~$3.5K/mo (~90%, ~$380K/yr) matching the paper;
suite count 16->22.
- site: real Hanzo mark (@hanzo/logo) in header; PDF embed hides TOC/thumbnail sidebar
(navpanes/toolbar=0); vs-claude card abstract leads with the productivity hook.
- README: added the 6-paper engine cluster.
Design paper describing the Hanzo unified Go binary architecture:
one statically linked process (hanzoai/cloud) mounting every platform
subsystem, built on the zip framework (Fiber v3) with embedded goja
for in-process JS, esbuild for TS->ES5 transpilation, and ZAP RPC
serving both the wire and the in-process dispatch seam.
Covers motivation (microservices overhead for co-located control
planes), architecture, the TS->goja->native-Go migration path
(esign tRPC->ZAP case study, 14/14 routers server-side),
multi-tenancy by configuration, the in-process-goja vs sandboxed
hanzoai/runtime code-execution boundary, a cost model anchored in
published service-mesh measurements, and performance (function call
vs network RPC, promise pipelining).
8 pages, two TikZ figures (layered architecture, cost comparison),
20 references. Compiles clean with tectonic; no missing citations.
One content-addressed, post-quantum-encrypted block substrate over S3 with two
surfaces: in-process SQLite VFS (hanzovfs) for structured tenant state, and a
read/write FUSE mount for Dropbox-class file sync. Per-object DEKs sealed with
hybrid ML-KEM-768 + X25519 (luxfi/age), ChaCha20-Poly1305 bulk. Realizes
HIP-0302 + HIP-0107.
Co-authored-by: Hanzo Dev <[email protected]>
Fix shared/ include paths to ../shared/ for 55 subdirectory papers so
that hanzocover.sty and lstlang.tex resolve when building from each
paper's own directory (the build pattern used by latexmk per paper).
Build fixes:
- hanzo-operative/sections/03-screen-understanding-pipeline.tex: move
\COMMENT from after \FOR{...} onto the \STATE line so the algorithmic
env doesn't mis-parse.
- hanzo-network-architecture (abstract + sections 04, 12, 13): escape
unicode mu (U+03BC) as $\mu$ for pdflatex.
- ntt-transform/sections/08-benchmark-results.tex: same mu escape.
All 55 Tier-2/3 papers now build clean with latexmk -pdf.
Add section 23 with literature-backed comparison scaffolds for:
- AI agent frameworks: vs LangGraph, AutoGen, CrewAI, OpenAI Agents SDK
- LLM inference engines: vs vLLM, TGI, TensorRT-LLM
Hanzo column entries marked [TBD-measure] until a controlled re-run
on the competitor's published harness is completed. Competitor cells
sourced to repos, papers, and third-party benchmark articles.
Reproducibility spec (hardware, models, software pins, decision rule)
documents what is needed to fill the [TBD] cells honestly.
No invented numbers; existing internal claims in sec 17/16 retained but
flagged in section 23 as not-yet-apples-to-apples vs literature.
GPU-native AI compute network with DEX+EVM+FHE triumvirate. Documents
the four-version chronology (1.0 initial AI compute -> 2.0 Hamiltonian
LLMs + AI Coin -> 3.0 full PQ -> 4.0 GPU-native sovereign L1), the
HMM (Hanzo Market Maker) native AI-compute DEX cross-listed via
Liquidity Protocol with Lux D-Chain and Zoo DEX, the AIVM opcode
surface (8 precompiles 0x0a01-0x0a08), confidential AI inference
on F-Chain (FHE), Proof-of-AI mining, and on-mainnet receipts at
h_0=36,963,000.
- hanzo-4-0-launch/hanzo-4-0-launch.tex (8 pages)
- INDEX.md
- hanzo-ico-infrastructure: redate to 2017 (Techstars year), expand
from 254 to ~600 lines of detailed token sale platform description
- hanzo-payments-platform: new 2018 paper on the crypto-commerce
platform with Bitcoin/Ethereum payment rails, fiat settlement,
and merchant API
Machine-checked proofs and automated verification for post-quantum
distributed systems. Covers Coq/Lean4 formalization of consensus
protocols, threshold crypto correctness, and gas determinism.
zeekay
