Paper 1 briefing / devudaaaa Research Lab x Zeolite Studios / Thursday, July 16, 2026

The Governed Runtime.
Programs that cannot lie, actions that cannot stray.

HXF runs near-English programs deterministically. CLAW, the governance piece teased in chat last week, checks every action against policy before it executes. Every run produces a byte-identical audit log. Scattershot is the host that makes it multiplayer and observable. Target: arXiv by July 31. This deck is the full pitch, plus five days of groundwork already done.

01 / 12
The whole paper in one picture / governance you can replay

Two runs. Same seed, same policy. One hash.

RUN A  seed=7741  policy=v3
tick 0041  trade gold=40   ALLOW
tick 0042  trade gold=250  DENY  escrow_required
tick 0043  move hex=E4     ALLOW
log sha256 = 9f2c11a8…
RUN B  seed=7741  policy=v3
tick 0041  trade gold=40   ALLOW
tick 0042  trade gold=250  DENY  escrow_required
tick 0043  move hex=E4     ALLOW
log sha256 = 9f2c11a8…
two runs, one hash: governance you can replay

Prior systems make the gate deterministic and leave the agent stochastic. We make the runtime itself deterministic, with the gate inside it, so the whole governed run replays byte for byte. To our knowledge, nothing published does that today. Proving this picture with real hashes is the paper.

02 / 12
Why now / three facts

The field just validated the problem. We answer it with a different architecture.

A hot field, a different angle

Since January 2026 at least six systems gate agent actions deterministically: Faramesh, Gated Behavior Trees, Reason Less Verify More, ILION and more. We verified the closest three ourselves on arXiv, abstract by abstract. All of them wrap a stochastic agent. To our knowledge, none makes the executed program itself deterministic and byte-replayable. The newest rival landed July 8, so speed matters.

AI agents need it now

Agents hallucinate steps. An agent that calls an HXF program cannot hallucinate a step, because the program, not the model, does the walking. A CLAW policy blocks what prompts cannot. MCP is the industry standard, so the door is open.

Proof beats promise

Today, governance is a promise in a document. Here it is a hash. Same seed, same policy, same log, byte for byte. Regulated industries can finally verify instead of trust.

03 / 12
Already done / July 11 to 16, before this call

We did not wait for the meeting. Five days of groundwork are in.

04 / 12
The correction / what the process catches before your name is on anything

We claimed first. We were wrong. The claim died in a day, inside the lab.

What died

The line that nobody has done deterministic plus policy-gated execution. At least six 2026 systems do it. We verified the three closest ones ourselves, on arXiv, abstract by abstract.

What survives

All of them make the gate deterministic around a stochastic agent. To our knowledge, none makes the executed program itself deterministic and byte-replayable. Our delta is the substrate, not the gate.

What changed

Two permanent rules: every novelty claim gets its own adversarial search, and the council reviews everything before it leaves the lab. First-ever language is banned in all artifacts.

verified on arXiv, July 12
2601.17744  Faramesh          deterministic gate, stochastic agent
2603.05517  Gated Behavior Trees  deterministic gate, stochastic agent
2607.07405  Reason Less Verify More deterministic gate, stochastic agent
ours      HXF + CLAW          gate inside a deterministic runtime

Why this is in a pitch: co-authorship means our mistakes could become your mistakes. These rules now stand between any claim and the outside world, and they protect your name as much as ours.

05 / 12
How it works / the governance piece teased on July 12

One loop. One choke point. One log.

HXF program  →  interpreter step  →  CLAW policy_check(action)  →  allow / deny  →  execute or skip  →  audit log (hashed)  →  next tick

Scattershot hosts the loop: its typed event bus feeds inputs in a fixed order, and its live introspection (VarRegistry, BreakManager, WatchManager) lets us watch a running server without stopping it. CLAW is disclosed at concept level today. Its code moves into the shared private fork once the one-pager is signed, and the generic policy hook gets proposed to your upstream either way.

06 / 12
One example, four readers

The trade above 100 gold gets denied. Here is what that means to each of us.

The game designer

You write the rule in near-English: trades over 100 gold need escrow. It runs the same every time. No mystery bugs, no dice you did not roll.

The server engineer

Every action passes one choke point. Replays are hash-verified, so a desync is findable in minutes, not weeks. Debugging happens on the live server.

The AI agent builder

Your agent calls a program, not a guess. One MCP tool call runs a whole deterministic flow. Policy stops the refund your prompt forgot to forbid.

The studio and compliance lead

Every action is logged. Every denial carries a reason. Every run replays for an audit. You can hand a regulator the hash, not a paragraph.

07 / 12
What the paper claims / running code on a pinned commit

Four claims. Each carries a test, not an adjective.

Evidence plan: twin-run hash test, policy overhead numbers, tick jitter and bus throughput on Scattershot, a denial demo, and a public repo that reproduces everything in 15 minutes. HXF's four known gaps get fixed or stated plainly in the limitations, seeded randomness first because C1 requires it. A red-team novelty sweep runs before the introduction is drafted, and every claim above carries "to our knowledge" until the sweep passes. We claim only what the code does.

08 / 12
Paper 2, in brief / your engine as a scientific instrument

The game is the instrument.

Paper 2 builds on Paper 1. DEADFALL, our multiplayer behavioral game, runs on Scattershot with Parallax bridged in for natural-language decisions. The experiment protocol itself is an HXF program, so every session is recorded, replayable, and audited by the same CLAW layer. Result: behavioral experiments other researchers can re-run bit for bit. First live results ship with the paper. Authorship there flips: Leed first, Michael second (Or like the first one).

09 / 12
Infrastructure / settled in chat this week, locked here, zero dollars

Domain, site, email, board. All workable this week for zero dollars.

10 / 12
The road / fifteen days to arXiv, then two weeks to compound

Tonight we decide. July 31 we submit. August we launch.

NOW · JUL 16

This call. Six decisions, one-pager signed, build starts tonight.

JUL 17-23

Build. Seeded RNG fix, CLAW walking skeleton plus the architecture pick, red-team novelty sweep, benchmarks B1 and B2, plain draft of Paper 1, endorsement secured, site v1.

JUL 24-31

Verify and ship. Twin-run test green, council passes, final line-by-line approval with you, LaTeX, submit to arXiv by July 31.

AUG 1-7

Launch. Repo public with the paper, Show HN, Reddit, LessWrong, Trello live, credit lines everywhere.

AUG 8-14

Compound. Peer venue submission for Paper 1, DEADFALL session locks Paper 2 data, professor outreach with the preprint attached.

Missing a gate never stops the mission. It changes what the paper claims, and the change gets reported, not hidden.

11 / 12
This call / thirty minutes, six calls to make

Say yes or no to these and the build starts tonight.

Standing cadence: thirty minutes every Thursday, one deck like this each time, and the deck arrives before the call so a slipped call never blocks a decision again.

12 / 12
scroll or arrow keys