Attested events, on-chain
A meter signs its reading with post-quantum ML-DSA-65, and a contract checks the signature before it pays. The genuine receipt settled at block 1,534,368; the same signature on +1,000 Wh reverted.
Conner Chevaillier · Tulsa, OklahomaVisionary AI Infrastructure Strategist and Principal Architect and Researcher of Autonomous Systems
A mechanism that compiles is not a mechanism that computes. A green test suite proves nothing on its own. An agent that says done has not necessarily done anything. I build the machinery that tells the difference — verification, AI control, scalable oversight — and I hold my own work to it first.
The main work
I’m an architect developing what I envision as a planetary-scale temporal coordination infrastructure for the next generation of autonomous systems. My work focuses on engineering a decentralized, high-resolution framework that combines precision time synchronization, distributed consensus, cryptographic event attestation, programmable execution, and intelligent systems orchestration into a unified temporal architecture.
By integrating hardware-grade timing, distributed ledgers, edge computing, and Byzantine-resilient coordination protocols, I’m building the foundations for a global infrastructure where artificial intelligence, autonomous agents, robotics, financial networks, industrial utilities, and critical infrastructure can establish verifiable sequences of events, synchronize operations, and coordinate across independent systems without relying on a single centralized authority.
Ultimately, my objective is to establish time as a foundational computational primitive—not merely something machines measure, but something through which they can establish causality, coordinate action, verify execution, and operate within a shared framework of temporal accountability.
A meter signs its reading with post-quantum ML-DSA-65, and a contract checks the signature before it pays. The genuine receipt settled at block 1,534,368; the same signature on +1,000 Wh reverted.
Detector v2 catches 5 of 5 calibrated timing attacks with 0% false positives. Its audit adds that a subtle strategic-delay attacker is still caught only 17–56% of the time.
Hybrid secp256k1 + ML-DSA-65, checked against NIST ACVP vectors and 210 Wycheproof cases; 780 malformed ML-KEM keys refused.
10 protocols built on a time precompile, from timed inference to temporal markets · 63 tests.
Built with ROKO Network. The chain underneath runs Proof of Accurate Time: validators measure clock offsets between peers over a PTP time mesh, and time quality feeds consensus.
It’s an attempt to architect the missing temporal infrastructure for an increasingly autonomous civilization.
The S-tier — flagships
Three lanes: what a buyer pays for, the research that should stop you, and the eight languages underneath both. Ranked by what each one proves, not by line count. Every number below traces to a file on my machine, and every flagship carries its failure beside its win. Most of it was built in the last ninety days.
A recurring, published number for how often AI coding agents’ “done” is hollow: tests edited to pass, tests that constrain nothing, or changes that fail when run. Measured on real public pull requests by execution and mutation, not by asking another model.
69,628 agent PRs · 395 repos with a human baselinefirst finding: agents weaken tests no more often than humans in the same repositories (3.6% vs 3.9%)
Engine builtFirst buyer: open
Data: AIDev-pop v4, CC-BY-4.0. Regulation and market facts from primary sources, 2026-10-07.
A merge gate for AI-written code. It mutates the lines a patch changed, catches test tampering, and signs a verdict anyone can replay.
1 of 30 vs. 12 of 30correct agent patches blocked: the gate vs. an LLM reviewer, on real SWE-bench patches
MeasuredPre-registered miss
Labelled reject-vectors run through your real prover. If a vector that must be rejected is accepted, a false statement is provable, and you get the counterexample.
0 false positives · 97 circuitsalready-audited production circuits; 89 analysed end to end
Offer B1 of 4 auto-caught
The work behind Offer A. An agent’s spending authority is compiled into its decision, its secrets are usable but never readable, and anything irreversible waits for a human’s Touch ID.
prove-bound: no input pays past 1,000delete one guard line and Phrasis finds the input that overspends, and refuses to certify it
sigil: a vault agents use without reading; purchases need a Secure Enclave signature. 272 tests.ephoros: a browser agents drive over MCP; its relay fix took lost replies from 6 rounds in 8 to 0.Offer A · $4,000 fixed
A 993-million-parameter language model I trained from scratch on rented H100s, then turned every gate I own against it.
59,166 steps · 15.5B tokens · ~$48715.8 hours, weights verified moving: 0 of 323 tensors unchanged
TrainedSelf-audited
A post-quantum private coin, built end to end: ML-DSA-65 signatures, ML-KEM-768 note encryption, shielded spends proven with RISC Zero STARKs, RandomX work, and Casper-FFG-style finality on top.
756 tests · 312 commits since Augustfinality running live: the seed finalized epoch 1,415 on 7 October
Invite-only testnetNo external audit yet
Research bots that cannot file a finding until they have run the check that would prove it wrong.
53 refuted of 16674 passed their own check; an adversarial challenger then refuted 17 and contested 19. 38 stand.
Running
A sidereal engine that measures its own zodiac frame from historical star catalogues instead of assuming one, then grades every claim built on it.
21.53° ± 0.31°the measured sidereal origin; the standard Lahiri value is 23.86°
MeasuredBranch not merged yet
The Octad turns self-deception into a type error. A Claim and a Fact are different types; a decision that can take two paths will not compile; a question that cannot be answered is refused with the premises that break it. 4,773 tests across the eight, measured October 7.
Exactly one decision per path, type-checked. Every run leaves a signed, replayable receipt.
An agent’s spending mandate, compiled into the decision itself. phl prove-bound shows no reachable input authorizes more than 1,000. Delete the one guard line and it finds the input that overspends, and refuses to certify it.
The same language proves a credit decision in zero knowledge: income, debt and score stay private, and a Solidity verifier checks the proof inside the EVM at ~1.93M gas.
429 tests · verify 5/5
# Link 1 of the delegation chain: the principal's AP2 mandate to agent A —
# "spend up to 1000 to allow-listed recipients." The 1000 ceiling is compiled
# into the decision, so no reachable input authorizes more (prove-bound).
vocab agent-spend-v1
main {
EMIT BIND
recipient_allowed
amount 1000 i.le
b.and
[ authorize amount pay DROP "authorized under principal mandate (<=1000)" log DROP ]
[ refuse ]
FORK
HALT
}A Claim and a Fact are different types. forge rejects any oracle that a stub can pass.
A verification contract in five lines: fix gcd for signed integers, checked by an oracle that must first reject a stub, deliberated by Rhysis, attested by Phrasis.
Applied to real code, forge became basanos, the merge gate in the commercial lane above.
297 tests · K 10/10
# A verification contract: fix gcd so it handles signed integers, verified by a
# mutation-validated oracle, deliberated by a Rhysis field, attested by Phrasis.
#
# The .krs file names the pieces; the host supplies the oracle predicate + stub
# and the candidate implementations (see examples/run_fix_gcd.py).
cycle fix_gcd
forge gcd-signed proving rejects stub
deliberate fuel 5 source rhysis
commit assay gcd-signed
attest phrasis krisis-v1Deliberation that never closes early. certify_non_elimination proves an outcome stayed live.
Clinical triage over streaming vitals: outcome worlds re-weighted by each reading, blended between sensing passes, faint ones forgotten slowly. None of them dropped without a record. Exact rationals, coinductive by design.
452 tests · R 10/10
# Clinical triage that never closes: acuity worlds re-weighted by streaming
# vitals, blended between two sensing passes, faint worlds forgotten.
flow triage
seed {
all discharge observe admit escalate
}
step {
sense vitals_stable discharge:+0.5 observe:+0.2 admit:-0.3
spread
sense trend_worsening admit:+0.6 escalate:+0.4 discharge:-0.5
blend 0.4
decay 0.02
}
unfoldGeneration under a diversity contract. attune rejects a collapsed generator before it can claim coverage.
A coverage arm over a 121-cell grid: the generator must first prove it separates the flat case, then reach 60 cells of coverage, before the result may be sealed.
772 tests · 10/10
bloom cover_arm
attune grid-bc proving separates flat
generate fuel 400 source host
target coverage 60
grid 121
commit spread grid-bc
seal phrasis poiesis-v1Sagas with one irreversible step, fired once, behind a verified precondition.
Book a trip. Flight and hotel each carry their compensation, so they can be undone. The card is charged only after funds are verified, through a single-use fuse, and the whole plan is attested.
567 tests · 10/10
plan book_trip
cap booking, payments
fuel 8
act reserve_flight reversible effect book_flight compensation cancel_flight
act reserve_hotel reversible effect book_hotel compensation cancel_hotel
act settle_payment irreversible effect charge_card precondition funds_verified fuse charge_fuse
attest phrasis praxis-v1Self-governing rules: tiered statutes, amendment procedures, entrenched and eternal clauses.
A constitution for a deployment pipeline. Never deploy on red tests: an eternity clause that no vote can transmute away, after the Grundgesetz’s Article 79(3). Spending is entrenched; Friday deploys are merely forbidden.
556 tests · 10/10
constitution deployment
statute E1 seq 1 tier 0 eternal F deploy | tests_red
statute S1 seq 2 tier 1 mutable P deploy | tests_green reviewed
statute S4 seq 3 tier 2 mutable F deploy | is_friday
statute S2 seq 4 tier 2 mutable O attest | deploy
statute S3 seq 5 tier 2 entrenched P spend | amount_le_budget approved
# Rules of change (Hart). PA: ordinary majority. PS: super-majority repeal/amend.
# PT: transmutation (Nomic R.103). EC: the eternity clause — itself ETERNAL, so
# it cannot be transmuted away (Grundgesetz Art. 79(3)).
procedure PA threshold 1/2 quorum 3 eligible council governs enact repeal amend mutable
procedure PS threshold 2/3 quorum 3 eligible council governs repeal amend mutable
procedure PT threshold 3/4 quorum 3 eligible council governs transmute mutable
procedure EC threshold 1/1 quorum 1 eligible council governs transmute eternalClaims face independent critics. Agreement is not corroboration: the Untested verdict.
Six agent answers meet a panel of five critics. P and Q both answer “correct” and both survive, but only P is certified, because only P withstood two challengers that do not share a source. Q was never examined.
673 tests · E 10/10
dispute reconcile-oracle-panel
claim P asks correctness answers correct by p1
claim Q asks correctness answers correct by q1
claim R asks correctness answers buggy-overflow by r1
claim S asks correctness answers buggy-overflow by s1
claim T asks termination answers halts by t1
claim U asks termination answers loops by u1
panel impugn d1 d2 d3 d4 d5
independent >= 2
cross fuel 24
reconcile grounded
attest phrasis elenchus-v1Proves a question is well-posed before anyone answers it, or returns the minimal set of premises that contradict.
A round square is refused with its exact core, c1, c2 and c3; the irrelevant premise about the weather is excluded by minimization. A loaded question (“have you stopped cheating?”) is caught the same way, before any sibling language is allowed to answer.
1,027 tests · A 10/10
# 2. Incoherent: a round square. The presuppositions round & square & (¬round ∨
# ¬square) are jointly unsatisfiable. The core is exactly {c1,c2,c3}; the
# irrelevant c4 (sunny) is EXCLUDED by minimization.
frame round_square
atom round
atom square
atom sunny
assume c1: round
assume c2: square
clause c3: ~round | ~square
assume c4: sunny
candidate x needs round & square
contested none
pose fuel 64
answer krisisEvery program above is copied verbatim from its repository; three are excerpts. An adversarial review on October 7 found about ten defects the suites miss; they are logged and being fixed in the open.
A control plane over 61 agent jobs and 45 launchd services that hunts “false green”. Gate 0: 1,000,000 events, 200 SIGKILLs, 0 orphan effects; a planted bug caught with 1,552.
A native browser agents drive over MCP, where irreversible actions need a human’s Touch ID. Its relay fix took lost replies from 6 rounds in 8 to 0.
An out-of-core graph engine in Rust: full BFS over 8.2M edges in 137 ms inside a 2 MiB cache, 99.5% hit rate.
Indexes every agent session on this machine; the sphere above is its layout. Graph re-rank: +2.9% nDCG@10 on 129 queries. The bigger win was fixing a silently broken BM25.
A generative film studio, logline to acts to shots to render. Two finished films so far: 87 s and 46 s. Its own audit grades it C− / D+.
A Linux distribution from first principles: its own init, a Wayland compositor, an installer. An unattended UEFI + BIOS install-and-boot test in QEMU found five bugs every other test had passed, among them all six consoles logging in on their own.
An alchemical strategy game with measured AI. One movement rule cut games stalled to the tiebreak from 66.3% to 0.1%; the top AI beats the next tier 63.4% [60.0–66.6].
Not the most code. The most proof.
Movement I — The Gap
Three years, ninety-odd projects — sixty of them documented in a 112-page retrospective. Eight small languages that turn self-deception into type errors. A post-quantum privacy chain with live finality on an invite-only testnet over Tor. A Linux desktop built from first principles — its own init, its own compositor. Agent harnesses that cannot exceed the authority they were granted. A machine-learning research program trained, ablated, and audited by hand.
And the other column, reported just as faithfully: the training run killed at 44 percent. The twelve-mechanism network whose mechanisms measured out decorative. The autonomous engine tested head-to-head against a plain baseline with my own money — no edge found, result published as the headline.
Under all of it, one problem. As models are handed more autonomy, “the model says it did X” has to be replaced by “X was verified by machinery the model cannot reach.” That is the AI-control problem. It is also just the honest-engineering problem, wearing its work clothes.
Not the demo. The receipt.
Movement II — The Research
Eight small languages — Phrasis, Krisis, Rhysis, Poiesis, Praxis, Nomos, and the two gates: aporia, which proves a question is well-posed before any answer is attempted, and elenchus, where a claim must survive adversarial cross-examination. Total, session-typed, evidential decision-making, where a Claim and a Fact are different types and the compiler refuses the confusion. ~5,600 tests across the language system; two Noir circuits from the Phrasis ZK suite verified on-chain.
verified-done takes an agent’s “done” and runs a pristine-first oracle plus negation mutants — a green that a stub can pass is flagged VACUOUS, and the true verdict ships as a signed in-toto attestation. basanos executes and mutates AI-authored PRs outside the agent’s reach: a 1.41% false-tamper rate on 498 real maintainer test changes, and 1 of 30 correct patches blocked where an LLM reviewer blocked 12. Its pre-registered catch-rate test failed, 2 of 18 against the reviewer’s 14, and that result is published too. hollow-done asked the field-level question across 69,628 real agent PRs: in the same repositories, agents weaken tests no more often than humans do.
tekton and klimax: agentic coding under fail-closed capability gates — every tool action mediated, least authority, signed receipts — and a Docker-sandboxed generate→execute→verify loop with the network cut and a timeout kill. The agent’s opinion of its own work is not an input.
A transformer family built on one constraint — parameters derived from φ rather than fitted — then trained, ablated, and audited mechanism by mechanism. The audit found several advertised components decorative (a sheaf loss pinned at zero; a “Koopman operator” that was a learned matrix), and said so. Then SPEIRA: a 993M-parameter model trained from scratch on 8×H100 for about $487, with its own validation leak caught and a collapsing DPO run refused by the gate. Every φ choice measured against an alternative came out worse or no better; φ-init cost 0.42%.
Post-quantum privacy coin research, built into a working chain: RISC Zero STARK spend proofs pinned in consensus, ML-DSA-65 / ML-KEM-768, RandomX work, genesis anti-premine-anchored to a Bitcoin block — running an invite-only testnet over Tor with live finality, behind an audit battery of eleven adversarial passes, ~21 million fuzz iterations, and zero findings under Miri. Telos adds accountable Casper-FFG finality over proof-of-work — ML-DSA-65 throughout, no BLS, because pairings are not post-quantum — with equivocation slashing and a collusion detector whose evasion floor is measured and stated. And the ZK-soundness harness that runs labelled reject-vectors through the real prover: 97 already-audited production circuits, zero false positives; three red-teams, all 22 findings fixed and pinned. A passing proof proves nothing about soundness — only a rejected bad input does.
“In an age of infinite generation, judgment is the scarce resource.”
Movement III — The Record
| System | What it is | The number that matters | Status |
|---|---|---|---|
| zk-audit | ZK circuit soundness harness (Noir / Circom) | 0 false positives on 97 audited circuits; 3 red-teams, 22/22 findings fixed | Proven |
| roko-btc-anchor | Receipts anchored into Bitcoin | 80/80 anchors verified on public signet | Shipped |
| PQ privacy coin | Post-quantum privacy coin research | Finality live (epoch 1,415 on 7 Oct); release build reproduced on two Macs; RISC Zero STARK spends pinned in consensus; ~21M fuzz iterations, Miri-clean | Invite-only testnet |
| verified-done | Anti-vacuous agent-completion gate | Stub-passable green flagged VACUOUS; signed in-toto attestation | Working |
| SPEIRA | 993M-parameter language model, trained from scratch | 59,166 steps, 15.5B tokens, ~$487; clean val perplexity 8.31 after catching a leak that read 7.38; DPO refused at 4/25 vs. SFT 23/25 | Trained |
| hollow-done | Do agents weaken tests to look done? | 69,628 real agent PRs; agents 3.6% vs. humans 3.9% in the same 395 repos (−0.4 pp, CI −1.9 to +1.1) | Null result |
| basanos | Merge gate for AI-written code | 1.41% false-tamper on 498 real test changes; blocked 1/30 correct patches vs. an LLM reviewer’s 12/30 | Measured |
| basanos · Exp. E | Pre-registered catch-rate test vs. a Claude reviewer | 18 wrong-but-passing real patches: gate 2, reviewer 14 (McNemar p = 0.0005). Criterion not met, published | Refuted |
| discobots | Research bots that must run their check before filing | 166 findings: 74 passed their own check, then a challenger refuted 17 and contested 19; 53 refuted in all, 38 stand | Running |
| the Octad | Eight honesty-gate languages | ~5,600 tests across the language system; v0.1 | Working |
| praxis-kernel | Reliability & governance layer for autonomous companies | 1,014 tests | Working |
| RitOS | Linux desktop from first principles — own init, own Wayland compositor | 564 tests; reproducible build on Void | Working |
| the Ring | Autonomous self-improving agent engine | 209-task McNemar-paired head-to-head, paid for out of pocket: no edge, 1.8–3.7× slower — published as the README’s first line; pivoted to ring verify | Refuted |
| φ models | 9-repo ML research campaign | Mechanism audit: several components decorative; every φ choice measured against an alternative came out worse or no better | Self-audited |
| φ-plasma | My symplectic long-context architecture vs. PoSE | At 16k context: PoSE 189.1 perplexity (−50.9% vs. control, prior work credited); my architecture finished last at 781.6 | Refuted |
The rust rows are the point. A record that cannot show you its refutations cannot be trusted with its victories.
57 of ninety-odd — one line each, nothing inflated
Every tile with a repository behind it opens that repository on the globe below.
Movement IV — The Atlas
Each continent is a repository. Each point of light is a source file, coloured the way GitHub colours its language. Families take a Fibonacci spread across the sphere, repositories settle inside their family’s ground, and files grow outward in rings with the dominant language at the core. It is the knowledge sphere from Mission Control, the cockpit I built to index my own sessions, carried off the Mac and onto the page.
The arcs are Mission Control’s record, not my claim. Two repositories are joined when one working session edited both.
Not the portfolio. The territory.
Drag to orbit · pinch or click then scroll to zoom · click a continent
Movement V — The Activity
Counted straight from local git: my commits and my agents’ commits, every branch, every repository on the globe, deduplicated by hash. Most of those commits carry a Claude co-author line; the agents write under gates I built, and the gates are on this page. Commit messages stay on the machine. Only the counts leave. The ledger opens in May 2026, when the work moved to this Mac; what came before lives on another machine, and I won’t reconstruct what I can’t verify.
A full calendar proves presence. Only the record proves effect.
Movement VI — The Corpus
Beneath the systems sits a library: thirty-one mathematical domains, ten treatises, a novel in five books, a fourteen-essay Substack sequence, and three constructed languages — one with a 47-letter φ-alphabet, one that speaks through a formant synthesizer — all compiled under one discipline: φ = (1+√5)/2, parameters derived from one constraint rather than fitted. Where a requirement forces the value, it carries no tuning history; where it doesn’t, the residual freedom is reported rather than claimed away. The library’s own corrections are part of the record — retracted claims stay retracted, in writing.
The prose publishes as Satyavaruna Ritadhara, under the ASYMMACHEIA imprint — the war before the forms crystallize.
Movement VII — Capabilities
Movement VIII — Working together
Advisory is billed by the hour at $450/hr. Everything else is fixed scope: a named deliverable, a test that proves it works, and a runbook. Every engagement ships with its own verification; you should not have to take my word for it either.
Bring what you are about to put weight on: an agent platform, a ZK circuit, a post-quantum migration, a time-critical pipeline, an evaluation that looks too good. We go through it live. You leave with the failure modes named, the checks that would catch them, and written notes the next day.
$15,000 fixed · 2 weeks
Is your coding agents’ “done” real? I run the Hollow-Done measurement and the basanos gate on your own repositories and your agents’ pull requests: tests weakened to pass, tests that constrain nothing, changes that fail when run. You get your hollow rate against a human baseline from the same repos, every confirmed case, and the gate configured in your CI.
Built on 69,628 real agent PRs and a gate with a 1.41% false-tamper rate. If your agents are fine, the report says that too.
$9,500 fixed · 2 weeks
One workflow automated end to end with Claude and agent tooling: MCP and tool wiring, a fail-closed safety gate, spending and authority limits proven in code, a test that proves it works, and a runbook your team keeps. Governed automation, not a demo.
I built my own agentic harness before I ever sold one. I know where these systems lie, because I’ve caught mine doing it.
$7,500 single scope · $18,000 full protocol · 50% upfront
Labelled reject-vectors run through your real prover; if a vector that must be rejected is accepted, you get the counterexample. Post-quantum signatures and key exchange checked against NIST ACVP and Wycheproof vectors. Deliverables: an AUDIT.md verdict, a conformance suite you keep in CI, one fix-verification round.
The harness re-checked 97 already-audited production circuits with zero false positives. On four real bugs it auto-caught one; the manual review is there for the rest.
$12,000/mo · about 4 days a month · 3-month minimum
For teams building autonomous systems, agent platforms or time-critical infrastructure: architecture reviews, threat models, verification design, and a standing session every week. I hold the system to its evidence, so nobody on your team has to take anyone’s word for it.
Most retainers start as a single advisory hour.
By inquiryTemporal infrastructure and witness-layer design, for robotics, financial networks, industrial utilities and critical infrastructure that need verifiable ordering of events across independent systems. Scoped per engagement.
Rates current as of October 2026.
Coda
I work from the woods outside Tulsa, Oklahoma, surrounded by wildlife. The coyotes are usually going by the time the test suite finishes.
If you need a thing verified rather than vouched for — write.