Research · External decision authority
The model can propose. It should not define the canonical answer.
The bounded repair experiments showed that output form could sometimes be reconstructed while canonical identity remained unresolved.
That converted invariant preservation from a generation problem into a control problem.
The authority problem
In normative technical translation, admissibility cannot be inferred from fluency or from the model's confidence. A canonical term is admissible because an external authority defines it for the relevant document family.
The architectural consequence is separation: the translation model generates text; an external deterministic layer checks and, within bounded rules, enforces the governed identity.
SIF 1.0: terminology admissibility
SIF 1.0 operationalised this separation through an Anchor Registry for metrological terminology.
On the frozen terminology evaluation, deterministic Registry enforcement raised Anchor Identity Precision from 0.1905 to 0.7143. Detection recall was 1.000 and the false-positive rate was 0.000 on the defined null probes.
A naive exact glossary replacement performed worse than the unmodified baseline in a separate comparison, reinforcing that surface replacement alone does not establish governed identity.
What SIF 1.0 established
- Canonical identity can be represented as an external control object, separate from the translation model.
- Terminological admissibility can be audited independently of general fluency.
- Registry coverage and authority correctness bound the verdict.
Why SIF 1.0 was not enough
Critical technical structure is not limited to terminology: formulas, constants, measurement units, numeric values and context-sensitive concept identity can also be non-negotiable.
SIF 2.0 therefore extended terminology-only control to post-generation structural admissibility.
SIF 2.0: structural admissibility
SIF 2.0 combines governed concept identity, protected technical carriers and context-qualified bounded intervention.
The control architecture remains external to the generator: the domain specification defines what is governed, and deterministic layers decide what can be admitted, repaired, reviewed or escalated.
This architecture is implemented operationally in SIF Inspector.
Why the story continues into Governance
SIF 2.0 still assumes that the applicable governing specification and admitted representation are themselves valid enough to govern the transaction.
That assumption becomes the research question of SIF Governance: does what currently governs still have valid support after sources, evidence, relations, configuration or authority change?