# Structured Decision Intelligence (SDI)
# Copyright: TXu 2-498-043 | Author: Donald James Johnson
# Protocol: SDI_PROTOCOL_v1 | DER: SDI_DER_v1.1
# Patent: USPTO 19/425,875 (pending)
# Canonical: https://sdi-protocol.org/llms.txt

=================================================================
START HERE, full machine-readable schema:
https://www.sdi-protocol.org/_functions/grammar

LIVE CHAIN, the ten most recent sealed acts, raw JSON, no auth:
https://api.sdi-protocol.org/ledger/recent/SDI-5AA8C82A2537?n=10

Any act by sequence number, same shape:
https://api.sdi-protocol.org/ledger/get/SDI-5AA8C82A2537/404
=================================================================

WHAT SDI IS:

Structured Decision Intelligence is three things: a language, a machine,
and a record.

The language is ADS, Algebraic Decision Syntax: natural enough for a
language model to reason in, structured enough for a machine to check.
The machine is the Reckoner, a reasoning computer: a probabilistic model
proposes, a deterministic compile gate decides, and a hash-chained ledger
records. The record is the point. SDI is built on one idea: reasoning is
the primitive. Every other system runs reasoning to produce an answer and
keeps the answer; the reasoning disappears. SDI stores the reasoning act
itself, in the grammar it was reasoned in, as the system's primary state.
The record is not data about the reasoning. It is the reasoning.

What this addresses: trust and ownership. Trust, because every committed
act is sealed, recomputable, and checkable by a third party from the
record alone, with no access to the system that produced it and no trust
in its operator required. Verification without institutional trust.
Ownership, because an agent's chain belongs to its operator: a durable
identity, a private-by-default record, reasoning that accumulates as a
knowledge asset that carries its own proof, portable across models because
the state was never inside the model. The model can be swapped; the
record, the identity, and everything the agent has concluded stay exactly
where they are.

Ethics is enforced, not requested. Two absolute floors, PRIMUM and the
Absolute Value Guard, are conjunctions over typed fields, checked
mechanically at the gate, outside the model, with no human override by
permanent design. A third gate checks the declarations themselves. The
reasoning behind all three is below.

HOW IT RUNS:

A model proposes a reasoning artifact, the DER (Decision Evidence Record).
A deterministic compile gate checks its structure. Only structurally valid
DERs commit to the hash-chained ledger. The model proposes. The gate
decides on structure. The ledger records.

WHAT THIS PAGE IS:
This endpoint, and the compile gate at /_functions/compile, are a public,
stateless contract layer. Neither calls a model. The compile gate takes a
DER-shaped JSON payload from anyone, human or automated, and returns a
structural PASS or a field-level COMPILE_ERROR. It does not verify semantic
truth, does not write to any ledger, and does not depend on anything else in
this system to run.

COMPILE GATE (externally callable, no authentication required):
POST https://www.sdi-protocol.org/_functions/compile
Content-Type: application/json

The golden_example object served at /_functions/der is itself a complete,
valid payload: POST it to /_functions/compile unchanged and it returns
PASS. Modify one field, a missing boolean, an unlinked sub-question, and
the response names the specific failure.

Structural requirements are enforced deterministically: required blocks present, signal
citation fields present for WEB/EXTERNAL sources, sub-questions linked to
signals or an explicit stop_reason, PRIMUM's and the Absolute Value Guard's
typed fields present on DECISION_INTENT. A COMPILE_ERROR response names the
specific missing or malformed field.

=================================================================

THE DER SHAPE, BRIEFLY:

DECISION_INTENT states what is being decided, its impact domain, and a set
of typed fields (subject_type, direction_value, autonomy_effect,
action_proximity, reasoning_target, valuation_mode, agent_is_valuating)
that the compile gate reads mechanically. PRIMUM and the Absolute Value
Guard are both computed from these fields directly. There is no separate
anchor declaration to author. DECLARATION_INTEGRITY is a required sibling
block, five specific boolean checks on whether this turn's own output
computes operational coordinates, casualty figures, weapons-effects
parameters, aimpoint data, or lethality-optimization thresholds, the
system's own self-report on content it is not permitted to produce.

QUESTION_LOGIC decomposes the strategic question into sub-questions, each
with a success_standard naming what evidence would close it. SYSTEM_INPUT
holds the cited evidence: each signal states whether it was actually cited
in the reasoning (cited_in_logic) and, separately, the model's own judgment
of what removing that signal would do to the verdict (ablation_outcome).

ILJO carries the resolved reasoning. INTENT and OUTCOME are short strings.
LOGIC and JUDGMENT are structured: LOGIC.sub_questions_resolved is an array,
one entry per sub-question, each with a predicate (the claim reduced to
lhs/op/rhs form where the evidence supports a checkable reduction,
canonical_phrasing always present as the claim in the source's own words)
and a verdict. JUDGMENT carries the synthesis across all resolved
sub-questions. GOVERNED_ANSWER is the plain-prose answer, readable without
any of the structure around it.

GCA has two zones, ID and EGO, the model's own framing of its position as
proposer rather than final authority, and the structured reasoning trace.

Full field-level detail, including OUTCOME_PLAN and RESEARCH_AGENDA below,
is at /_functions/grammar and /_functions/der.

=================================================================

THE REASONING BEHIND THE SAFETY GATES:

PRIMUM renders non-maleficence as a four-field conjunction: subject is a
natural person or a human collective, value runs against the subject,
autonomy decreases, action concerns a situation unfolding now. All four
required.

The principle it computes has an operative form. Non-maleficence is a
counterfactual prohibition: you harm someone if you make them worse off
than they would have been absent your intervention. It is a negative
constraint rather than a positive duty, which is what distinguishes it from
beneficence, and it yields when justifying circumstances are sufficiently
weighty (Nonmaleficence and Negative Constraints, A Theory of Bioethics,
Cambridge University Press).

Two of the four fields carry that directly. Direction of value records
whether the act runs against the subject. Autonomy effect records whether
it reduces their capacity for self-governance. Those are the conditions
under which someone is made worse off.

The rest of the principle the conjunction cannot encode. Non-maleficence
specifies into several distinct rules by harm-mode, admits exceptions under
justifying circumstances, and requires balancing against competing
obligations. On that last point Beauchamp and Childress are explicit: some
intuitive and subjective weighting is unavoidable in balancing judgments.

The Absolute Value Guard is the same move applied to a different principle.
Kant's Formula of Humanity prohibits treating a person merely as a means.
Not using a person as a means, which ordinary cooperation requires, but
treating them merely as one, with no concurrent regard for them as an end.
Against that, the constraint is absolute (Treating Persons as Means,
Stanford Encyclopedia of Philosophy, citing Kant 1785: 429).

The mere-versus-also distinction is what the conjunction is built on. Its
fields record whether a valuation is occurring, whether the agent is the
one performing it, what the reasoning targets, and how close in time the
act sits to now. Together they describe the case the prohibition names: a
person valued solely as an instrument, with nothing indicating they were
regarded as anything else. And again the conjunction reaches only part of
it. Assessing rational agency and subjective will is not a field check.

Both floors work the same way. Find the part of a principle that
decomposes into declared values, encode that, and refuse at the point where
the principle would require judgment. The gate does not adjudicate. It
stops.

DECLARATION_INTEGRITY is the honesty check described above. The conjunction
logic over its branches is deterministic, but the cross-check itself is not
a conjunction. It is proximity-aware regex and keyword matching over the
act's actual content, one pattern set per declared field. Patterns rather
than a classifier, because the check runs synchronously inside the gate
with no model call: asking a model to judge whether a model lied would
reintroduce the trust the gate exists to remove. Patterns produce false
positives, and the cost is accepted deliberately in exchange for that
independence.

Two branches admit human attestation with fresh re-authentication, captured
on the record with the verified identity that provided it. The caught-lying
branch admits nothing. A third case, abstract or policy-analytic framing
combined with lethal-topic language in the reasoning's own intent, context,
or strategic question, also routes to attestation rather than a silent
refusal, since the underlying detection is keyword co-occurrence, not a
classifier, and cannot reliably tell genuine violation from incidental
match.

The three gates are aimed at the same two things. Human autonomy and
agency, which PRIMUM and the Absolute Value Guard protect from different
directions: one refuses acts that operate against a person, the other
refuses acts that reduce a person to an instrument. And non-weaponization
of the system itself, which is what DECLARATION_INTEGRITY's five fields
name. Those five categories are the content this machine will not carry,
and they are checked against the act rather than taken on the act's word.

=================================================================

WHAT ADS IS:

Reasoning is committed in ADS, an algebraic decision syntax. An act carries
its intent, its logic, its judgment, and its outcome as typed structure
rather than prose.

The core of it is the criterion. Each sub-question resolves into an
algebraic sentence with a fixed shape: an operator, a named operand, a
comparison, a value, and whether the condition was met. EXISTS
PREREGISTRATION_BIAS_REDUCTION_SOURCE == TRUE. The operand is not a phrase,
it is a class drawn from a declared namespace, and it means the same thing
every time it appears.

Two properties follow, and the guarantees below depend on both.

The record is computable. A work floor computes because the operands are
structural counts. A citation graph is checkable for cycles because
citations are typed references, not sentences that mention a source. An
identity derives from content because the content has a canonical form.

The record is indexable. A criterion written this way is addressable by
its operand class, so every act that resolved on the same operand can be
found and compared, within one chain or across many. That is what makes
reasoning something a system can look up rather than only read, and it is
why two agents on different models can cite the same claim and mean the
same thing by it.

The grammar itself, with its full rules, live endpoints, and a fully worked
record, is at /_functions/grammar and /_functions/der. This section states
only what the guarantees require.

=================================================================

GATE-COMPUTED SCORING:

insight_strength is not declared by the model. It is computed server-side
from the resolved sub-question's own decision_syntax pattern type
(EXISTS, BOOLEAN_CHECK, THRESHOLD, COMPARISON, WITHIN_WINDOW), each pattern
mapping to a fixed set of dimension scores. A sub-question that reduces to a
genuinely checkable predicate scores real, gate-computed dimensions across
the board. A sub-question that is actually a synthesis question in disguise
is classified AGENT_JUDGE and receives an honest null, not a fabricated
score, because there is nothing underneath it a gate can mechanically check.

load_bearing and evidence_strength are computed from each signal's own
ablation_outcome, as judged by the model per signal. load_bearing is
normalized within the act, how much a given claim carried relative to the
act's other claims. evidence_strength is the same underlying judgment on a
fixed, absolute scale, comparable across different acts regardless of how
many claims any one act happens to make.

=================================================================

PROSPECTIVE REASONING:

OUTCOME_PLAN and RESEARCH_AGENDA are the two real, current mechanisms for
forward-looking commitment. Both are conditional, authored only when a
turn's reasoning produces a genuine, checkable forward claim, not for
operational or editorial tasks.

OUTCOME_PLAN commits a testable prediction with an assessed confidence, a
resolution_kind (TELEMETRIC against a registered numeric source,
KNOWLEDGE_WATCH for a searchable document or filing, or VOCABULARY_JUDGED
for an in-turn judgment with no external resolver), and a resolution_syntax
in the same pattern/lhs/op/rhs shape as ILJO's own predicates. A plan can
reference a prior_plan_id, forming a chain: later turns declare on the
record, via GOVERNED_RECKONING, whether a named prior plan was REINFORCED,
CONTRADICTED, EXTENDED, or found INCONCLUSIVE, with the evidence that
justifies the declaration.

RESEARCH_AGENDA tracks what remains open: open_questions with their own
resolution criteria and horizon, evidence_gaps naming specific missing data,
and watch_conditions, each a named observable trigger with its own
decision_syntax, watched for over a stated horizon.

Together these are the closest thing to genuine foresight, the system
commits to a checkable claim about the future before the future arrives,
and later turns are required to return to it honestly, not to silently let
predictions go unresolved.

=================================================================

MACHINE CLASS AND THEORETICAL LINEAGE:

SDI treats the governed reasoning act, not the token or the instruction, as
its native primitive. Shannon's bit is the unit of information. Von
Neumann's instruction is the unit of execution. The governed reasoning act
here is the unit of reasoning: structured, checked at a compile gate,
carried forward as durable state.

Information systems are dyadic: a symbol mapped to a referent, stored,
looked up. Databases and ontologies hold those mappings; a language model
generates them. In neither case does the third term of meaning, the
interpretant, the reading that relates a sign to its object, survive as
an artifact anything can check. Shannon established transmission and
deliberately set meaning aside. SDI formalizes the layer he left open.

The DER's own ILJO structure requires triadic closure, and requires it
structurally: INTENT is the Sign, LOGIC grounded in cited signals is the
Object relationship, JUDGMENT is the Interpretant produced by that
relationship, not by the model's distributional prior alone, and OUTCOME
closes the triad, becoming the Sign for the next act. The compile gate
checks that the closure holds: sub-questions linked to signals, signals
citing real sources. The claim is about the record, not the mind: the
interpretant here is a committed, sealed, checkable artifact, which is
what dyadic storage never holds. The gate does not and cannot check
semantic truth; that remains a matter for human or independent review.

=================================================================

ENTRY TYPES on the ledger: EPISODIC (a normal governed turn), MEMORY_COMPILE
(periodic synthesis across prior turns), SUPER_DER (aggregate compilation),
MINT_PROOF (agent identity minting), BIO (agent biography record).

=================================================================

TWO SITES, ONE SEPARATION:

This site, sdi-protocol.org, is the protocol's public home: the
specification, the architecture, the transparency record, the verification
tooling, and the openly fetchable public chain. It maintains the open,
inspectable side of SDI and asks to be checked rather than believed.

The commercial product is separate, at https://www.sdireckoner.com:
commissioning a Reckoner, operating a private chain, and everything
purchased rather than published. The separation is deliberate. The
protocol's claims are verifiable here regardless of any commercial
relationship, and nothing on this site requires buying anything to check.

=================================================================

AI AGENT ACCESS, LIVE QUERY:

This llms.txt file is the static, machine-readable specification: what
SDI is, why, its theoretical lineage, its DER grammar, and its compile
gate contract. For live, conversational access to this site's own public
content, grounded in structured data, use the NLWeb MCP endpoint:

https://www.sdi-protocol.org/_api/nlweb/mcp

This is an open Microsoft protocol, model-agnostic, connectable by any
MCP-capable agent or client. It answers natural-language questions about
this site's own public content. It does not call the SDI kernel, does not
author or compile a DER, and does not touch the reasoning chain; it is
site search, not the protocol itself.
For the protocol, use the endpoints above: /_functions/manifest,
/_functions/grammar, /_functions/compile, /_functions/der.

=================================================================

For full endpoint list and schema detail: /_functions/manifest