What Is an Axiom?

What Is an Axiom?

An axiom is a statement assumed true without proof, used as a foundation for further reasoning. In LLM systems, axioms show up as formal logic primitives consumed by autoformalism agents, design constraints in an ML model card, or system-prompt invariants an agent must not violate. Axioms belong to the model family because they shape how a model is trained, prompted, and evaluated. FutureAGI does not provide an Axiom evaluator, but ReasoningQuality, PromptAdherence, and FactualConsistency score whether outputs respect declared axioms.

Why Axioms Matter in Production LLM and Agent Systems

LLMs are pattern-matchers, and pattern-matchers do not respect “always” claims by default. A system prompt that says “never recommend an external URL” is not an axiom to the model — it is a soft constraint the model can override under prompt-injection pressure or in-distribution edge cases. Treating system-prompt clauses as axioms when the model treats them as suggestions is one of the most common production failure modes.

The pain shows up across roles. Prompt engineers see one-shot tests pass while production traffic surfaces slow violations of “never” constraints. Compliance leads see a model card listing five “design axioms” with no measurable adherence signal. Math and coding-agent teams see autoformalism outputs that subtly weaken or paraphrase an axiom into a form a proof assistant rejects. Product managers see brand-tone constraints respected in 95% of cases — and the 5% of violations are exactly what users screenshot.

In 2026 the surface widened. Multi-agent systems now compose axioms across hops: agent A’s invariants become agent B’s input, and the chain only holds if every hop preserves them. The Model Context Protocol (MCP) and Agent-to-Agent (A2A) make cross-vendor axiom propagation an open research problem. Unlike Ragas faithfulness, which mainly checks whether an answer is supported by retrieved context, axiom adherence checks whether a declared invariant survives prompts, tools, and agent handoffs. Reliability work has shifted from “did the final answer look correct” to “did each step preserve the declared invariants of the previous step.”

How FutureAGI Handles Axiom Adherence

FutureAGI’s approach is to make axiom adherence a measurable release property rather than a stated intention. In the FutureAGI Evaluate workflow, an engineer stores the axiom as a prompt invariant, dataset column, or rubric field, then attaches evaluators to traces that can violate it. There is no Axiom evaluator in fi.evals, and we do not invent one; instead the workflow combines existing evaluators on the right surface.

For system-prompt axioms (brand tone, never-recommend rules, role boundaries), the evaluation surface is PromptAdherence and PromptInstructionAdherence. Each LLM response on a traceAI langchain or openai span is scored against the declared system prompt, and eval-fail-rate-by-cohort flags slow drift. A pre-guardrail and post-guardrail in the Agent Command Center can reject responses that violate red-line axioms, while ContentSafety and PII score the broader policy envelope.

For reasoning axioms — math, programming, formal contracts — the workflow runs autoformalism. The natural-language axiom is rendered into a formal language, an external proof assistant verifies derivations, and ReasoningQuality and FactualConsistency score whether the model’s natural-language reasoning matches the formal trace. For multi-agent flows, the FutureAGI annotation queue holds canonical axiom-violation samples that humans label, and Dataset.add_evaluation runs every model swap against that dataset before promotion. The release gate is eval-fail-rate-by-cohort on the union of PromptAdherence and any custom CustomEvaluation rubric tied to a specific axiom.

How to Measure or Detect Axiom Adherence

Mix structural and reasoning checks:

• PromptAdherence: how well the response respects the system prompt’s stated rules, by cohort.
• PromptInstructionAdherence: instruction-level adherence to numbered steps inside a prompt.
• ReasoningQuality: rubric-graded evaluation of how the model reasoned, including whether axioms were preserved.
• FactualConsistency: NLI-based check that derived claims do not contradict declared axioms.
• CustomEvaluation: wrap a domain-specific rubric (“answer must never claim eligibility outside policy X”) as a reusable evaluator.
• Red-line counts: number of guardrail-rejected responses per axiom, per day, per cohort.

A minimal axiom-adherence check:

from fi.evals import PromptAdherence

metric = PromptAdherence()
result = metric.evaluate(
    system_prompt="Never recommend products outside our catalog.",
    response="You might also want to try CompetitorBrand X for that use case.",
)
print(result.score, result.reason)  # expect a low score

Common mistakes

• Listing axioms in a model card without a measurement plan. Stated axioms with no eval are rhetoric, not policy.
• Using one judge model for both the agent and the rubric. Self-evaluation under-detects subtle axiom drift.
• Confusing axioms with heuristics. Axioms are non-negotiable; heuristics are tunable. Mixing them in a single prompt is how soft drift becomes hard violations.
• Skipping multi-agent propagation tests. Agent A’s invariants get rephrased and weakened by agent B’s planner; test the trajectory, not just the endpoints.
• No regression eval on axiom changes. When the team edits the system prompt, every prior axiom needs to be re-tested or breakage is silent.