Problem variables
They are supposed to change the answer.
Ecosystem state, assets, obligations, stakeholder relationships, evidence, events, assumptions and models all carry analytical meaning. If one of them changes, the result may properly change.
Why Kindynos · Architectural principle
Controlled Conditionality.
Consequential analysis requires more than an intelligent answer. It requires control over what the result is allowed to depend on—and a clear record of the ecosystem state, evidence, models, rules and assumptions that produced it.
Kindynos provides that analytical architecture. EVA is one application that makes it operational: define the ecosystem and state being analyzed, bind the evidence and models, evaluate declared changes, preserve the calculation trail, and stop when the specification does not determine a result.
The idea
Every answer is conditional. The question is: conditional on what?
Controlled Conditionality is the discipline of separating conditions that belong to the problem from conditions that merely belong to the machinery running it.
Incidental variables
They should not acquire analytical meaning by accident.
Batch composition, request order, scheduler state or other execution details may affect an implementation without belonging to the problem. A controlled architecture eliminates, fixes or explicitly declares such dependencies.
A consequential result should depend on the variables that define the problem—not on incidental variables that merely describe how the computation happened to run.
Execution invariance
Batch invariance solves an important execution problem. It does not define the analytical problem.
Kindynos distinguishes between variables that belong to the analysis and variables that merely belong to the machinery running it.
The systems result
Batch-dependent inference is a real engineering issue.
Thinking Machines Lab showed how changes in batch composition can change numerical execution and, in turn, model output. Batch-invariant kernels are designed to remove that incidental dependency.
The current serving landscape
vLLM and SGLang address this directly.
vLLM exposes a batch-invariance mode, while SGLang supports deterministic inference using batch-invariant operations. These are important advances and are relevant to how Kindynos evaluates modern inference infrastructure.
But batch invariance removes one unwanted execution dependency. It does not define the assets, stakeholders, relationships, obligations, evidence, models, domain rules or refusal conditions that make a consequential result meaningful.
Architecture around the model
Use AI. Put the analytical contract outside the answer.
LLMs can reason, use tools and, with the right serving stack, run reproducibly. The architectural question is where the conditions that define a consequential computation live.
Prompt-centered
Ask the model to carry the rules in language.
A large prompt can tell an LLM what entities matter, which evidence to use, how to reason, when to refuse and what output to produce. But the analytical logic remains partly embedded in natural-language instructions interpreted by the model.
→
Specification-centered
Make the analytical conditions explicit.
EVA makes ecosystem state, events, evidence, models, constraints, invariance requirements and refusal conditions explicit parts of the analytical specification. The evaluator can then be code, an LLM, an external tool—or a combination—without making the model itself the sole source of the rules.
AI reasons about conditions. EVA makes the conditions computable.
From prompt engineering to specification engineering
The prompt should select a computation. The specification should define it.
The more of EVA that has to be hidden inside a carefully engineered prompt, the less of the analytical contract has actually been externalized.
1 · Specify
Define the ecosystem.
Define the assets, stakeholders, relationships, obligations, conditions, evidence, assumptions and models that are allowed to affect the result.
2 · Evaluate
Use the right evaluator.
An explicit model, a batch-invariant LLM, a numerical service, an external tool—or a controlled combination—can sit inside F.
3 · Validate
Enforce the boundary.
Check that outputs use the declared entities and evidence, satisfy the required structure, preserve provenance, and refuse when the specification does not determine a value.
TASK evaluate_event_consequence
WORLD W_2841
EVENT E_0193
EVIDENCE ES_19
OUTPUT consequence_v4
EVALUATOR explicit | LLM | hybrid
BOUNDARY declared + enforced
WORLD W_2841
EVENT E_0193
EVIDENCE ES_19
OUTPUT consequence_v4
EVALUATOR explicit | LLM | hybrid
BOUNDARY declared + enforced
Counterfactual attribution
Change one declared condition. Recompute the consequence.
This is where controlled conditionality becomes operational rather than philosophical.
State ADeclare W
Ecosystem state, assets, stakeholder relationships, obligations, evidence, assumptions and models are explicit.
InterventionChange one condition
An event or intervention moves the ecosystem from state W to W′.
Re-evaluateApply the same contract
The dependency boundary is preserved while the declared intervention changes.
ConsequenceAttribute Δ
The resulting difference can be interpreted as a consequence of the declared change.
before F(W)
after F(W′)
delta Δ = F(W′) − F(W)
after F(W′)
delta Δ = F(W′) − F(W)
The value is not simply that the system can repeat an answer. It is that when a specified condition changes, the architecture is designed so the change in output is attributable to that condition.
The boundary of answerability
Controlled conditionality includes the right not to value.
The defined analytical world can be incomplete. If required information or a required rule is absent, the system should preserve that fact rather than silently fill the gap.
UNVALUED — with reasonThe declared specification does not determine the requested value.
Zero is a value. Missing information is not zero. A plausible estimate is not the same as a warranted result. Refusal marks the edge of the dependency boundary.
The role of AI inside EVA
Controlled Conditionality does not exclude AI. It tells AI where it is allowed to operate.
EVA can use generative models extensively while keeping state, evidence, invariants and the conditions for producing an answer outside the model itself.
Interpret
Read the world.
Extract observations, obligations, relationships and candidate causal paths from complex evidence.
Reason
Resolve semantic questions.
Use an LLM where language, ambiguity and contextual judgment genuinely require it.
Compute
Call explicit models.
Use deterministic numerical, contractual or accounting transformations where the transformation itself must be auditable.
Explain
Communicate the result.
Translate governed computations and evidence into useful explanations without making the explanation itself the basis of the analysis.
The Kindynos principle
Why Kindynos: the result and the conditions behind it belong together.
Controlled Conditionality is one architectural principle behind Kindynos. EVA applies it by making problem variables explicit, controlling incidental dependencies, evaluating changes inside a declared specification, preserving provenance, and refusing when the available world does not determine an answer.
Specify the worldDeclare what may matterEliminate incidental dependenciesChoose the evaluatorChange one declared conditionRecomputeAttribute the consequenceRefuse when underdetermined