Invarra

Invarra research

How do we know a system is responding to meaning?

A system can behave differently when the wording changes, even when the underlying meaning stays the same. Invarra's research examines how to measure that sensitivity and what stable behavior can tell us about the thing being measured.

These papers describe research frameworks. Phalanx V1 uses deterministic execution controls and does not require LIP, CSR, or a Phalanx-trained semantic model to operate.

Paper 01

June 28, 2026

The Latent Invariance Principle

Evidence of stable behavior when the target cannot be observed directly.

Read summary

When we observe a latent target through prompts, documents, or other representations, a correct response to one example may not tell us enough. LIP asks what happens when the representation changes while the relevant meaning is preserved.

The principle concerns measurement validity. It does not claim to supply an algorithm that establishes truth or correctness automatically.

Representation process

r=g(Φ,c,ϵ)r = g(\Phi, c, \epsilon)

An observable representation is modeled as a function of the latent phenomenon, the representation channel, and residual variation.

Observed behavior

B(r)B(r)

The evaluator observes behavior under a representation, while the measurement question concerns behavior with respect to the latent phenomenon.

Public boundary

The publication presents the measurement principle and its argument. Private implementation materials are not included.

Full paper

Open the embedded PDF here, or use the direct links.

Open paper viewer
Read the publicationView PDFDownload PDF

Paper 02

June 28, 2026

Canonical Semantic Realization

A framework for measuring behavior across controlled expressions of meaning.

Read summary

CSR separates the semantic unit being studied, the way it is expressed, and the outcome observed from a system. That separation helps evaluate how changes in representation affect behavior.

It treats controlled realizations as repeated measurements of a shared semantic unit, with explicit conditions for what must remain the same.

Realization relation

p=π(s,c)p = \pi(s,c)

An observable realization is modeled as an expression of a canonical semantic unit under a representation condition.

Semantic preservation

p1≡semp2for valid realizations p1,p2∈E(s)p_1 \equiv_{\text{sem}} p_2 \quad \text{for valid realizations } p_1,p_2 \in E(s)

Valid realizations preserve the meaning-bearing commitments of the same semantic unit.

Public boundary

The publication explains the measurement framework and the conditions for interpreting semantic variation. Private implementation materials are not included.

Full paper

Open the embedded PDF here, or use the direct links.

Open paper viewer
Read the publicationView PDFDownload PDF

Publication license

Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International.

Start with a shorter explanation.

The field notes introduce the research questions in plain language. For the current product, see how Phalanx controls agent actions.

Read the field notesExplore Phalanx