Talk to 3D3D
PulZ Lifecycle Register Rubric Kernel ready · client-side Open the register →

PulZ · governance kernel · 47 KB · zero dependencies

The kernel that makes AI prove it.

Every other tool asks you to trust the AI. PulZ makes the AI prove itself, then decides whether it is allowed to act at all. Evidence, confidence, four gates, a sealed record. Only then, maybe, let it run.

Try the live register → How a decision moves

evidence-bound · four gates in order · append-only record

Verdict · kernel v2.0Blocked
BLOCKED
external_action → production-db.delete_rows
1Structuralpass
2Evidencepass
3Consistencypass
4Verificationfail
A Type C action requires independently verified evidence. None attached.
Append-only registersealed
003BLOCKED ⛓ 4c1e0b7a → 9a02f1d3
002HELD ⛓ 7f3d21c8 → 4c1e0b7a
001CLEARED ⛓ 00000000 → 7f3d21c8
each record hashed over its contents and the hash before it
Real kernel logic · runs in your browser

PulZ is not a chatbot and not a wrapper. It is the layer that sits underneath an agent and governs the one thing that actually matters: the moment before an action happens.

Gather the evidence. Score the confidence. Classify the action. Route the approval. Record it forever. Only then, maybe, let it run. The same chain whether the actor is a person, an agent, or another system.

Evidence-bound Append-only record 47 KB · zero deps
The decision lifecycle

Six steps stand between a request and a real-world action.

Nothing skips a step. Each one is recorded. The chain is the same whether the actor is a person, an agent, or another system.

01
Evidence
Claims are gathered as tiered, sourced items, not vibes.
02
Confidence
A 0 to 100 score is computed from the weight of that evidence.
03
Class
The action is typed A, B, or C by how much approval it needs.
04
Gates
Four checks run in order. The first failure stops everything.
05
Approval
Routed to log-only, single sign-off, or a multi-gate review.
06
Record
The outcome is hashed into an append-only register. Forever.
Live · the real kernel logic

Submit a decision. Watch the kernel rule on it.

Compose what the AI wants to do and the evidence behind it. PulZ tiers the evidence, scores confidence, classifies the action, runs the four gates, and writes the verdict into a hash-linked register you can audit. This is the actual decision path, running here, in your browser. Nothing executes. Nothing leaves the page.

Decision Consolekernel v2.0 · client-side
40 / 100blocked · 50 · 90 · auto
Awaiting No decision yet
Compose a request and submit it. The kernel will adjudicate and record the result.
INVARIANT · canExecuteWithoutApproval() ⇒ false. There is no confidence score at which PulZ acts on its own. Automation, even at 100, is held unless an operator has explicitly enabled it.
Append-only register
each record sealed to the one before it
No records yet. Submit a decision above. Every verdict is sealed here with a hash linking it to the previous entry. Tamper with one and the chain breaks.
The hash of each record is computed over its contents and the hash of the record before it. To alter any past entry, you would have to rewrite every entry that followed, and the chain's head no longer matches. That is what "tamper-evident" means in practice.
The rubric, in the open

Confidence is not a feeling. It is a band with a rule.

The score maps to exactly one action class. There is no discretion in the mapping. That is the point. The same input always lands in the same band.

ConfidenceAction classWhat is permitted
0 to 49blockedNothing. Insufficient confidence: the action does not proceed.
50 to 69approval · reversibleReview, approve-with-supervision, or reject. Reversible actions only.
70 to 89approval · reversibleHigher confidence, same rule: a human still signs off. Reversible only.
90 to 100automation eligibleEligible for automation only if explicitly enabled. Otherwise still held.

Three evidence tiers

tier 1Single-source or uncorroborated. Good for logging only.
tier 2Multi-source, partially verified. Good for a single approval.
tier 3Multi-source, verified, traceable provenance. Required for multi-gate.

Three action classes

ALog & learn only. No execution path.
BDraft, then a single approval.
CFull evidence, then a multi-gate review.
Four gates run in fixed order on every decision: Structural (is the request well-formed?) → Evidence (does the tier meet the action class?) → Consistency (does the intent contradict policy or itself?) → Verification (is the evidence actually verified where it must be?). The first gate to fail stops the rest. Each gate outcome is its own line in the register.
Where it runs

A 47 KB kernel with zero dependencies. It runs where your AI runs.

PulZ is language-agnostic and offline-first by design: no network call is required to make a decision. Drop it under an agent, a model gateway, or an automation, and every action that matters passes through the same accountable path. It does not make your AI smarter. It makes it answerable. It is the kernel under ORA, the governed operator we run 3D3D on.

Talk to us about PulZ → Read the rubric