Scope · umbrella SIF frameworkCanonical language · English
Structural Invariant Framework (SIF) · Research programme
Plausible output is not governed output.
SIF checks whether generated or translated technical content satisfies explicitly defined structural requirements, using an external governing specification and evidence-based decisions.
Only activated checks within the applicable Registry and technical-carrier scope are evaluated. This is not a claim of complete semantic correctness.
Recorded demonstration / Frozen replay · CASE 04 / 08 · STRICT
A missing technical carrier
English source
The measured mass was 5.25 kg.
Existing Ukrainian translation
Маса невідома.
Governed technical carrier:5.25 kg · source present / translation missing
STRUCTURAL REVIEW REQUIRED — activated check not preserved
Recorded decision: TECHNICAL_STRUCTURE_MISMATCH. Technical status: MISMATCH; Registry coverage: NONE.
This was a deliberately corrupted negative control. The check detected the missing value/unit; it did not independently assess the accompanying change of meaning. Manual correction and re-inspection are required.
Plausible output is not the same as governed output.
A generated or transformed result may remain fluent, readable or technically plausible while changing information that the governing specification treats as non-negotiable.
A valid decision can also lose its support.
When a governing source, evidence item, configuration, relation or authority state changes, a previously admissible decision may need to be reconsidered.
Two assurance branches
Two questions, one bounded framework.
Output Structural Verification
Does the result satisfy what is currently governed?
SIF 1.0 and SIF 2.0 inspect declared structural invariants after generation or transformation. SIF Inspector is the operational reference implementation of this branch.
Governing-State Assurance
Does what currently governs still have valid support?
SIF Governance evaluates changes in governing sources, evidence, relations, configurations and authority, and derives bounded downstream consequences while preserving unresolved states.
Origin
Built from metrology problems.
SIF originated in metrology, where controlled terminology, traceability, versioned authority and exact preservation of values, units and formulas make structural errors measurable.
Metrology remains the primary evidence domain for SIF 1.0, SIF 2.0, SIF Inspector and the core development of SIF Governance. Later studies test which parts of the architecture transfer beyond that setting and where specification or representation boundaries appear.
Scope
What SIF governs
Entities
Relations
States
Constraints
Evidence
SIF is not limited to checking individual terms or fields. Depending on the application, the governing specification may also control relations between entities, logical operators, dependencies, source support, authority, states and decision limits.
Anchor: an explicitly governed concept identity, such as “acceptance limit”. Registry: the versioned list of accepted concepts, canonical forms and applicable rules. Anchor Identity Precision (AIP): a test metric for correct governed anchor identity in the declared Registry and evaluated instances, not a score for whole-translation quality. Technical carrier: a protected piece of technical content such as a value/unit or formula token that must survive the tested transport workflow.
Decision model
A bounded SIF decision
Bounded decision chainEach transaction applies only its declared scope and applicable checks.
Governing source
Applicable governing specification
Admitted input / representation
Applicable checks
Deterministic decision
Decision evidence
Evidence snapshot
Results with their scope and limits.
Evidence is shown with its configuration and boundary. The results below do not form an aggregate SIF score.
Frozen unseen SIF 2.0 evaluation set, under the declared anchor-identity precision metric.
Boundary
This result describes the declared frozen denominator only. It is not a population accuracy estimate or a claim of general translation correctness.
Frozen evaluation
Protected technical carriers
Observed result
73 / 73Tested carrier instances preserved with protected transport; 48 / 73 without protected transport.
Scope / configuration
Frozen Formula Stress configuration.
Boundary
Literal carrier preservation under the tested workflow; not mathematical-equivalence checking or arbitrary recovery of formulas already lost before inspection.
Boundary finding
Paired domain transfer
Observed result
Nuclear / radiation safetyTransfer supported with specification extension.
Trauma- and stress-related psychiatryTransfer not supported under frozen criteria.
Scope / configuration
Declared SIF control architecture unchanged; curated nuclear and radiation-safety specification and trauma- and stress-related psychiatry specification.
Boundary
Architecture reuse does not guarantee domain-specification sufficiency.
Reported R597/3 study
Real-metrology Governance evidence
Observed result
24 / 24 cases356 / 356 assertionsThe R597/3 study summary reports that the predefined end-to-end contract passed; false lifecycle activation was 0 and fresh-unpack reproduction was exact in 24 / 24 cases.
Scope / configuration
One bounded R597/3 real-calibration metrology evaluation context containing 24 held-out test cases / scenarios. These counts are reported in the R597/3 metrology study summary (9 September 2026); its original 356-assertion execution ledger was not independently recovered for this website review.
Boundary
The 24 test cases are not 24 independent real-world metrology contexts. Separate 19/19 development probes, 24/98 impact-derivation results, and the public demo's 26/26 and 12/12 gate counts are different studies or count units. This bounded result does not establish universal Governance transfer or production certification.
Scientific boundary
What a positive result does not establish
SIF does not establish overall translation quality, universal correctness, completeness of an undeclared specification, hidden dependencies, certification or operational fitness.
A positive result applies only to the checks that were applicable under the declared specification.
Research progression
The architecture was not designed in one step.
Research lineageHow the research progressed from invariant preservation to two assurance branches
Invariant-preserving foundations
Limits of pointwise loss
Structural ordering
Behavioral regimes
Representation diagnostics
Representation / output gap
Output failure taxonomy
Bounded repair
External decision authority
Structural verification
Governing-state assurance
Each stage was motivated by what the previous stage could not establish. Negative and boundary results are retained as part of the evidence chain rather than removed from the narrative.
An operational reference implementation of SIF 2.0 for bounded second-pass structural inspection of English-to-Ukrainian metrology translation.
The Inspector exposes the applicable checks, decision and evidence rather than returning only a revised text.
Live locator · pending confirmation
Public research demonstrator
SIF Governance Demo
A public research evidence demonstrator for governing-state assurance: source-derived obligations, bounded impact and non-impact, explicit OPEN states, source-supported consequences and deterministic replay.
The SIF programme includes published and archived work on invariant-preserving technical translation, corpus ordering, tunable representation regimes, structural admissibility, transfer studies and specification assurance.