Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Foundations: Invariants and Moves

These principles summarize sections 2.2 and 2.3 of Siegfried Handschuh’s Hybrid AI Patterns (October 2026), also introduced in Studio EX06. They apply to every pattern and composition in this section.

Five invariants

InvariantImplementation consequence
Model text is dataStore proposed rules, queries and actions as candidates. Check and admit them before execution.
Model answers are dated observationsRetain prompt, inputs, model, timestamp and result so observations can be audited and replayed.
Rules decideEncode acceptance, transitions and completion in executable rules.
Evidence decides acceptanceRequire checkable evidence, such as a permitted property and a supporting source quote. Confidence is ordinary data.
Failures stay distinctPreserve failed execution, malformed replies, unclear interpretations and logical unknowns.

The host performs calls and measurements. It should not silently replace a model verdict with its own domain decision before the rules see the receipt. If a correction matters to the outcome, represent it explicitly so the proof can show it.

Four moves

MoveWhat crosses the boundary
RequestA rule derives a request with the exact inputs the tool may see.
ReceiptThe host records the outcome as facts linked to that request and its subject.
AcceptA rule admits a proposed claim, records a reason to refuse it, or leaves it unresolved.
AskA rule identifies a question for a person or the evidence that would settle an open goal; the answer becomes a receipt.

A received field and an accepted domain fact are different stages. Merely recording quoteFound provides no protection unless a rule requires it. See Grounded Extraction for this boundary in executable form.

Distinguish the outcomes

SituationRepresentationMeaning
Provider or tool failuretool#status -> "failed" with errorExecution did not produce usable output
Invalid response formatGeneric host failure, or dialogue formatError and received fieldsA reply arrived but broke the expected contract
Ambiguous interpretationJudge result("verdict") -> "unclear"The model could not settle the criterion
Logical unknownQuery status unknownCurrent facts and completeness assumptions do not settle the goal
Evidence refusedAn application rejection rule and its reasonA received proposal failed an acceptance condition

These cases can call for different next steps: bounded format repair, additional evidence, human clarification, or termination. None should be converted into an affirmative answer merely to continue the workflow.

Stored values such as false or "unknown" remain data. See World Assumption for query truth, and Dialogue and Interviews for a flow that stores a judge’s three-way interpretation as slot state.

Keep authority and evidence separate

A quote existing in a document is a lexical measurement. It does not show that the document is authoritative for that claim. A conversational source may establish a language preference while being insufficient evidence of a degree or qualification. Give sources explicit permissions and keep those permissions visible in the acceptance rules.

Likewise, an old approval does not authorize a new state. Include relevant revisions in evidence joins and request identities, as introduced in Lesson 15.