Research foundation
The public ideas behind bounded instruction control.
Invarra publishes a non-enabling account of the measurement and control principles behind Phalanx. The research explains the questions being tested; it is not, by itself, proof that a specific release implements every theoretical property.
Public research boundary
Five concepts, described without the private mechanism.
The individual head roles, labels, corpus payloads, quad construction, formulas, thresholds, score geometry, transition tables, private reason codes, and signing material remain confidential.
TRP
Phalanx uses bounded specialist neural evidence heads.
LIP
Training and evaluation test stability across meaning-preserving realizations.
CSR
Corpus work separates semantic intent from controlled surface realization.
ISL
The training method combines invariance with structured contrast so specialists track meaning rather than surface form.
TRPCP
Deterministic stateful control composes compatible evidence into pass, clarification hold, or block.
Evidence discipline
A principle becomes a product claim only when it is bound to the release.
Research can motivate a training or arbitration design. It does not establish current runtime behavior. Product claims come from signed artifacts, declared protocols, bounded evaluation results, health checks, and rollback evidence.
Phalanx evidence therefore reports the population, result, environment, artifact identity, and limitation together. No research term is used as a substitute for an observed release result.