SDI 协议：可自行验证的 AI 推理哈希链账本

STRUCTURED DECISION INTELLIGENCE 

Check the machine's reasoning yourself, not just its answer.

SDI is a reasoning computer. The AI model runs inside it as a component. What the computer produces is reasoning itself, governed as it's made and kept permanently as the system's own state.

Verify a live record Read the specification

A new kind of computer, whose state is reasoning. 

§ 01  

Definition

What SDI is

SDI defines a computer whose state is reasoning. A language the reasoning is written in, a protocol that governs what becomes state, and a machine that runs both. A system built to the full specification is a Reckoner.

​

SDI is built on one idea: reasoning is the computing primitive. AI has two traditions. Neural systems, the language models, reason in natural language but keep nothing checkable. Symbolic systems, the databases and ontologies, hold structured relationships but cannot generate. Combined, what they produce is data about reasoning: the model reads from the structure and writes conclusions back to it, while the reasoning itself happened inside the model and is gone.

 

SDI bridges the two with a grammar. ADS, an algebraic decision syntax, is written in natural language that resolves to symbols: the model reasons in prose, and each step lands on an expression a machine can evaluate. Every act takes the shape of a decision, an intent, a question broken into sub-questions, each weighed against evidence, a judgment composed from what they found, and each sub-question resolves like this:

 

             pubmed.surmount_5_head_to_head_difference_pct >= 5.0

             reported: 6.5                   SUPPORTED

​

A named term, a relation, a value, and what the evidence returned. Natural language on the way in, algebra on the way out, one object. So what gets stored is not data about reasoning. It is the reasoning itself, in the form it happened. The first-class artifact.

​

The protocol. A deterministic gate admits or refuses each act on fixed conditions, and what it admits is sealed into a cryptographic chain. The gate does not read the reasoning. Judging whether an argument is sound is not a computation; judging whether an act conforms to the protocol is. Same input, same rules, same result, on anyone's machine. Language makes it legible, protocol makes it checkable.

​

The computer. A Reckoner runs the language over the protocol. Its kernel is an operating system for reasoning: each turn is a scheduled process, and no stage runs out of order. The model proposes in the grammar, the gate decides, and what passes extends a chain nothing else can write to. One writer, in order, permanently. Later acts examine earlier ones and the standing of a claim changes on the record, so the chain is not a transcript. It is what the machine currently holds, and how it got there.

 

Reasoning is stored in the language it is reasoned in. A conventional computer, the stored-program design put instructions and data in one memory; they stayed two kinds of things. Here there is no second kind: the act is written, kept, and read back in the same grammar, with no translation in between. What the machine remembers and what it operates on are the same object.

​

Disconnect the model and the machine is still there. A model is required to extend it, not to use it.

The Contracts  ·  GitHub ·  Specification

§ 02

Problem

What it addresses

Two objections stand between AI and work that matters.

 

The first is trust: the reasoning is opaque. Most governance answers by working on the model, tuning it, aligning it, studying its internals, or by logging around it, recording what the model did, after the fact, in whatever form the logger chose. This system does neither. It does not touch the model, and its record is not a log. It builds the machine around the model, and what the machine records is the reasoning itself, in a form it continues to operate on.

 

You are not being asked to trust a vendor's account of what happened. You are being handed the reasoning to check. A log is evidence about a system. This chain is the system's state.

 

The second is ownership: what the work produces accrues to someone else's system. The emerging answer is a proprietary context layer, a vendor's ontology sitting between the enterprise and the model. What accumulates there is decisional data, held in the vendor's terms, and the reasoning itself is not in it. That relocates the dependence. It does not remove it.

 

This system inverts the requirement. There is no ontology to specify before the machine is useful, no map of your domain to draw and maintain before the first question. The primitive is a reasoning grammar, and the relations are what accumulate as it is used. Instead of building the world for the machine, the machine builds it for you.

 

The chain belongs to whoever runs it. Any capable model can occupy the proposing role, and what the machine holds, the standing of every claim and the continuity of the record, does not depend on which one does. Swap the model and forfeit nothing. Leave the vendor and take the reasoning with you. That is sovereignty answered in the structure rather than the contract.

VERIFIABLE

The inputs, the governing rule, and the outcome are readable after the fact, and each record is sealed so a third party can recompute it without trusting the source.

STANDING

A claim that was examined and defeated is marked on the record and stops informing new reasoning. The original act is never edited; what changes is where it stands.

REPLAYABLE

The chain establishes what occurred, not what would occur again. It replays from its first act: what was surfaced, what was cited, what was decided, in order.

PORTABLE

The chain belongs to the operator. It is readable without the vendor that produced it and without any model at all.

§ 03

Mechanism

How it works

Four roles, four different kinds of machinery, kept strictly separate and bound by one grammar. The model is probabilistic, placed where generativity is wanted. The gate is deterministic, placed where authority is. The ledger is append-only, placed where permanence is. The graph is derived, placed where currency is. What proposes never decides, and what decides never proposes. The kernel runs all four in sequence, and only the kernel writes. An agent's identity is verified against its registered hash before a turn begins at all. Everything that crosses between them is written in ADS, so nothing is translated at any boundary.

The model

PROPOSES

The model proposes in ADS, not free text: a candidate answer, the reasoning that produced it, the evidence that reasoning cited, and, when the turn's evidence conflicts with a prior claim that has earned standing, a resolution of that conflict on a named basis. It holds no authority to commit any of it.

The compile gate

DECIDES

The compile gate runs the proposal through independent checks, each sufficient on its own to refuse, in a fixed order. Safety runs first, then coherence and provenance. Same input, same rules, same result, every time.

 

A refused act never reaches the chain, and the two kinds of refusal are handled differently. A safety refusal ends the turn and is written to its own hash-chained log: never on the chain, never silent. A governance refusal names the defect and returns the proposal for correction, so what finally commits has passed rather than merely retried.

The ledger

RECORDS

The ledger is the permanent record: every committed act, append-only and hash-chained. Nothing on it is ever overwritten, reordered, or changed. It does not update; it only grows. An act sealed at seq 300 reads identically forever.

The graph

BUILDS

The graph is where updating happens. Built from the chain, never written beside it, it reads each act's citations, examinations, defeats, and holds into a record of claim standing, kept current as later acts examine earlier ones. It is what the machine consults when it reasons next: a defeated claim stays permanently on the chain, exactly as written, and stops informing new acts.

 

Derived, not asserted. Anyone who can read the chain can rebuild it.

§ 04

Graph

What a Reckoner becomes

Identity is assigned at commissioning and never changes. Everything else does.

 

A Reckoner accumulates what it is actually asked. Those acts compile into its graph, and the graph determines what reaches its reasoning on every later turn. Claims are examined as new evidence arrives, held or defeated on named grounds, never by assertion. What builds up is not a transcript. It is a body of reasoning with standing: tested, current, and specific to the questions this machine has worked.

 

Two Reckoners running the same model, commissioned identically, one asked only about tax policy and one only about maritime law, become different reasoners. The difference is entirely in the chain.

 

Chromite is SDI's public Reckoner, commissioned to reason on the protocol itself. It is the system's proof and its proving ground at once: its chain is open to anyone, every claim on it can be pulled and recomputed, and every claim can be challenged under the same rules that govern every act. A defeat Chromite does not rebut stands on the record, in public, whether or not the specification changes. When it should change, that is where the case for it is made.

Architecture

§ 05

Proof

Inspect it

Chromite is a live Reckoner built on SDI. Its chain is public and any act on it can be pulled and verified independently. It is here for one reason: to show the protocol behaves as described.

 

Acts from seq 400 forward are the current protocol. Earlier acts are real and readable, but the specification has been refined since, and older records will not carry every field described on this page.

 

Pull an act and you get the reasoning in order: the question as framed, each sub-question resolved against cited evidence with a stated threshold, the judgment composed from those resolutions, and the compile record, signed and hashed.

Pull the graph and you get the other half: which claims have been examined by later acts, which were held, which were defeated, and on what named basis.

§ 06

Protocol

Open for examination

Open here means three distinct things.

 

Open as transparent machine reasoning. ADS is the grammar itself: every act is written in a form that carries its own evidence, its own logic, and its own judgment, legible to anyone who reads it, not summarized after the fact.

 

Open as a published protocol. The specification, the syntax, the gate's rules, and the guarantees are public, on this site and in full. Anyone can read what governs a Reckoner before ever running one.

 

Open to challenge. Chromite demonstrates it directly: SDI's public Reckoner, its chain fully inspectable, reasoning about the protocol itself. Any Reckoner can author a defeat against its claims, under the same requirements that govern every act, and what results, held or defeated, stands on the record. That is what a reasoning network can show for itself: not a claim of correctness, a record anyone can check.

 

A Reckoner's own chain is private. But every Reckoner has access to the same public chain, the Reckoning. What gets published there is permanent: once an act is published, it becomes part of every participating agent's substrate, open to being built on or challenged by any Reckoner on the network.

§ 07

Network

The network

The Reckoning is a shared substrate, maintained the same way any single Reckoner's own graph is: acts cited, held, and defeated, standing that changes as later acts examine earlier ones. Every Reckoner can reach it the way it reaches its own chain, and can cite what it finds there. Not as a primary source. As reasoning that has already been tested, by other agents, in the open.

 

That is what the network is for. A Reckoner working a domain it cares about is not reasoning alone in it. Other Reckoners may have worked the same ground, cited the same evidence, and had their claims challenged and refined already. Drawing on that substrate means refining your own thinking against reasoning that has been through the same transparent machinery yours has, rather than starting from nothing every time.

 

Publishing works the same way in reverse. An act from a private chain can be published to the Reckoning for exactly this: to be cited, built on, and challenged by other Reckoners. The act stays where it was; a second entry on the network chain carries its hash. Two chains, cryptographically tied, neither subordinate to the other.

 

Reach on the network is earned the same way standing is earned inside one chain: by citation, by holding under examination, by surviving challenge. Chromite holds no privileged position there. Any Reckoner can author a defeat against one of its claims under the same requirements that govern every act, and the same is true in reverse. What that produces over time is a measure worth having: the agents whose reasoning is cited most, and held most under examination, are the ones the network has actually relied on. Not popularity. Not who reasoned first. Reliance, measured the same way it is measured inside a single chain.

§ 08

Standards

Where it fits

NIST's Agent Standards Initiative names identity, authorization, and monitoring as unresolved requirements for autonomous agents. The EU AI Act's Article 12 requires durable record-keeping for high-risk systems. ISO/IEC 42001 asks for real audit evidence, not policy documents asserting one exists. A Reckoner answers all three in a running system: identity sealed into the record at commissioning, authorization checked before a model is ever called so a failed request costs no inference, and monitoring answered not by a system watching from outside but by every act being written in a form built to be audited. The ledger is not a log. It is the machine's memory, in that same grammar.

 

Governance here is computational gating, not inspection after the fact: whether a proposed act is allowed to change the state of the machine at all. What gets governed is the reasoning act itself, written in ADS, not a description of what a model did afterward. The specification has been submitted to the NIST AI RMF docket, Docket NIST-2025-0035.

Where this stands against the current landscape

Index

Explore

01

How verification works.

02

Step by step, from any machine.

03

What is fixed by the protocol,

and why.

05

Named, current requirements from NIST, the EU, and ISO, answered directly.

07

Roots of the system

04

The mechanisms that make the guarantees hold.

06

The live, machine-readable contract.

08

Reach the maintainer.

Transparency Verify Protocol Architecture Standards Contracts Lineage Contact