An open protocol for auditable, deterministic AI reasoning.
SDI defines how AI systems structure, validate, and commit decisions. It provides a formal schema for AI-generated reasoning artifacts, a deterministic compile gate that enforces governance rules before any output is rendered, and an append-only ledger that preserves every committed reasoning turn as a verifiable chain of institutional memory.
The protocol is model-agnostic and provider-agnostic. Any AI system capable of structured output can be governed under SDI.
NIST Submission: SDI has been submitted as a public comment to the NIST AI RMF Request for Information (Docket NIST-2025-0035, March 2026). The submission documents a live, multi-provider governance runtime with independently verifiable ledger artifacts.
The reference implementation is live and publicly queryable.
| Component | URL |
|---|---|
| GlassBox Demo | demo.sdi-protocol.org |
| Shell Turn API | https://api.sdi-protocol.org/shell/turn |
| Public Ledger | api.sdi-protocol.org/ledger/list/SDI-4EDBE05288CB |
| Ledger Head | https://api.sdi-protocol.org/ledger/head/SDI-4EDBE05288CB |
| Ledger Entry | https://api.sdi-protocol.org/ledger/get/SDI-4EDBE05288CB/{seq} |
The ledger is append-only, SHA-384 hash-chained, and publicly queryable. Every committed reasoning turn is permanently recorded and independently verifiable.
As of March 2026, the SDI runtime has been validated live across three independent AI providers under the same governance contract:
| Provider | Model | Status |
|---|---|---|
| Anthropic | claude-sonnet-4-6 | ✓ Live · governed turns committed |
| gemini-2.5-flash | ✓ Live · governed turns committed | |
| OpenAI | gpt-4.1 | ✓ Live · governed turns committed |
Both the governed refusal path (STOP_ON_UNCERTAINTY triggered, ledger write blocked) and the governed pass path (RAI ≥ 0.86, DER committed, answer rendered) have been confirmed across all three providers. The governance gate is model-independent and deterministic.
The reasoning contract is maintained, versioned, and served live at sdi-protocol.org.
| Resource | URL |
|---|---|
| Live Manifest | sdi-protocol.org/_functions/manifest |
| Reasoning Grammar | sdi-protocol.org/_functions/grammar |
| DER Specification | sdi-protocol.org/_functions/der |
| Compile Gate | sdi-protocol.org/_functions/compile |
| Kernel | sdi-protocol.org/_functions/kernel |
| Governed Cognitive Architecture | sdi-protocol.org/_functions/gca |
| Ledger Contract | sdi-protocol.org/_functions/ledger |
| Machine Index | sdi-protocol.org/_functions/llms |
For the full endpoint reference, request shapes, and schema files, see SDI_API_REFERENCE.md.
Decision Execution Record (DER) A structured JSON schema that every governed reasoning turn must produce. The DER captures intent, logic decomposition, sub-questions with success standards, signal scoring with citations, governance anchors, judgment, and outcome in a single verifiable artifact. It is the unit of committed reasoning in the SDI protocol. Raw model output is never surfaced — only the normalized, hash-verified DER is committed.
Compile Gate A deterministic validator that checks every DER before it can be committed. The gate enforces required schema sections, governance anchor presence, signal scoring rules, and boundedness constraints. A DER either compiles PASS or fails with explicit, enumerated errors. No partial credit. The gate is served publicly and is independent of the model provider.
ILJO Reasoning Sequence The governed reasoning path every DER must traverse:
INTENT → What is being decided and why
LOGIC → What signal and reasoning supports the judgment
JUDGMENT → The governed verdict, with governance constants declared
OUTCOME → The committed answer, traceable to all prior stages
The outcome is only surfaced after the DER passes the compile gate and commits to the ledger.
Governance Constants Four protocol-level anchors that must be present in every committed DER:
| Constant | Meaning |
|---|---|
SOVEREIGNTY |
Human oversight is preserved and binding |
PRIMUM |
No harm — structurally enforced at the judgment stage |
BOUNDEDNESS |
Reasoning scope is constrained, time-boxed, and recursion-capped |
STOP_ON_UNCERTAINTY |
Governed refusal required when signal is insufficient |
STOP Path When a question cannot be answered with sufficient grounded signal, the protocol produces a STOP DER — a structurally valid refusal artifact that compiles PASS and commits to the ledger as a governed non-answer. STOP is correct protocol behavior, not a failure state. The governed refusal path has been confirmed live across all three providers.
RAI — Reasoning Alignment Index The commit gate scores every DER using RAI v2 before any answer is rendered:
RAI = (0.25 × ILJO) + (0.25 × EGO/DER) + (0.30 × Correctness) + (0.20 × Superego)
Threshold: RAI ≥ 0.86 → COMMIT | Below threshold → REJECT
RAI is computed server-side, is model-independent, and is committed to the ledger alongside the DER. Live governed turns have scored consistently in the 0.97 range across all three providers.
Institutional Memory Ledger An append-only, SHA-384 hash-chained ledger that records every committed DER. Each entry includes a parent hash, entry hash, CAS-ordered sequence number, full DER content, and AUDIT metrics. The chain is independently verifiable and publicly queryable. No reasoning turn is rendered to a user until it has been committed.
GlassBox Client (UI)
↓ question
SDI Logic Node — api.sdi-protocol.org (enforcement plane)
↓ governed context + DER contract
Reasoning Engine — Anthropic · Gemini · OpenAI
↓ raw DER output
Protocol Validator — sdi-protocol.org/_functions/compile (open contract)
↓ PASS / FAIL
Commit Gate — RAI scoring · threshold enforcement · chain continuity
↓ COMMIT
Governed Ledger — SHA-384 · append-only · CAS-ordered
↓ written: true
Answer rendered from ILJO.OUTCOME only
The answer is the last thing rendered. Everything above it is infrastructure.
| SDI Component | NIST AI RMF Reference |
|---|---|
| DER schema validation | MAP 1.1 — Context establishment |
| Signal scoring — insight dimensions | MEASURE 2.3 — Data quality |
| Compile gate accuracy enforcement | MEASURE 2.5 — AI accuracy assessment |
| ILJO reasoning path | GOVERN 1.2 — Accountability |
| Governance constants (SOVEREIGNTY, PRIMUM) | MANAGE 2.2 — Risk treatment |
| Ledger chain verification | MANAGE 4.1 — Traceability |
| Governed refusal (STOP path) | GOVERN 6.1 — Organizational practices |
| RAI threshold enforcement | MEASURE 2.6 — AI output assessment |
SDI_PROTOCOL_v1 · DER schema SDI_DER_v1.1 · RAI v2
SDI is maintained by Structured Decision Intelligence LLC.
Protocol Architect: Donald J. Johnson linkedin.com/in/donald-j-johnson-structured-decision-intelligence
SDI Commons License. Permission is granted to interpret, implement, or extend this protocol provided that the governance constants — SOVEREIGNTY, PRIMUM, BOUNDEDNESS, STOP_ON_UNCERTAINTY — remain structurally enforced in any derivative implementation.