Sermo Arbiter Inferentiae Determinata (SAID)

What other inference apps do

You pick a model manually. You run it. You hope the output is right. No selection rationale. No audit trail. No governance. If it fails, you try a different model.

What SAID does

Ranks models across 6 dimensions using multi-objective optimization. Routes requests deterministically. Validates output against formal constraints. Retries on violation at zero cost on local GPU. Produces HMAC-signed decision receipts for every inference. Fails closed on contract violations. Same inputs, same model selection, every time.

Others: you pick a model manually

Pareto-optimal model selection

No other inference tool mathematically selects models. SAID uses NSGA-II evolutionary optimization across cost, speed, quality, privacy, capacity, and regulatory compliance — computing a Pareto frontier and selecting via weighted Chebyshev scalarization. The selection is deterministic, reproducible, and auditable.

Others: inference is single-shot

Generate-validate-retry

No other inference tool validates output before returning it. SAID passes every inference through a gate engine with formal constraint checking. Contract violations trigger structured retry with feedback enrichment. On local GPU, retries are zero-cost — enabling aggressive constraint stacking that would be prohibitively expensive through cloud APIs.

Others: no constraint enforcement

Formal constraint system

No other inference tool enforces output contracts. SAID supports set membership, regex, range, cross-reference, and composite constraints — defined in YAML, enforced deterministically. Every attempt, verdict, violation, and remediation path is logged. The gate function is pure: same output always produces the same verdict.

Others: no record of decisions

HMAC-signed decision receipts

No other inference tool cryptographically signs its decisions. Every SAID inference produces an HMAC-signed receipt: the selected model, the full score breakdown across all dimensions, the applied rules, and the hyperparameter snapshot. Replay any decision after an incident. Prove why that model was chosen.

Others: local and remote are separate

Unified selection across local and remote

No other inference tool evaluates local and remote models in the same mathematical framework. SAID runs one selection calculus across llama.cpp, MLX, PyTorch, Metal, and CUDA locally — and OpenAI, Anthropic, Groq, Mistral, and Cohere remotely. A local MLX model competes against a cloud API on the same Pareto frontier.

Others: errors are undefined

Deterministic failure doctrine

No other inference tool distinguishes contract failure from operational failure. SAID defines two explicit failure states: contract violations fail closed (TERMINATED), operational faults degrade gracefully (DEGRADED). Deterministic state transitions. No partial states. No silent failures. Security boundaries are never silently crossed.

Others: every request runs unconditionally

Runtime governance predicates

No other inference tool gates requests before execution. SAID enforces rule-based filtering at inference time: task type, token count, time-of-day, provider scope. Weight overrides, forced model assignment, and block-with-reason on constraint violation. Frozen request context prevents mid-flight mutation.

Others: no reproducibility guarantee

Replay equivalence

No other inference tool guarantees reproducible model selection. SAID logs every decision with full reasoning and score components. Seeded execution with determinism fingerprinting. Same inputs, same model selection, every time. This is what makes incident forensics possible.

Others: governance bolted on top

OpenAI-compatible API

Drop-in replacement interface — your existing code works unchanged. But unlike proxy layers that add governance after the fact, SAID builds selection, validation, constraint enforcement, and audit into the inference path itself. Governance is the engine, not a wrapper around it.