Viability calculator

Purpose

The viability calculator turns a qualitative proposal into a visible benefit-versus-cost argument. It is designed for structural actions such as adding a specialist, replacing a router, merging two models, changing quantization, or retiring a package. The calculation always compares the proposal with no change.

Interpretation

The result is a normalized decision margin, not an objective fact. Positive margins indicate that the stated benefits exceed the stated costs under the selected weights. A negative margin favors no-op. A small positive margin should normally trigger more evidence collection rather than immediate promotion.

Benefit dimensions

• Task utility: expected improvement on the defined workload, including critical slices.
• Robustness: resistance to distribution shift, adversarial input, component failure, or degraded dependencies.
• Coverage: new tasks, languages, devices, jurisdictions, or data types served acceptably.
• Diversity: non-redundant behavior that improves coalition resilience or future adaptability.

Cost dimensions

• Memory: resident, peak, cached, and artifact-storage impact.
• Latency: end-to-end tail latency, not model-only median latency.
• Energy or compute: inference, training, evaluation, transfer, and idle-residency cost.
• Complexity: operational surface area, dependencies, on-call load, and change coordination.
• Risk: security, safety, privacy, licensing, provenance, and governance exposure.

Production rule

Do not permit a single aggregate score to hide a hard failure. The candidate must pass every invariant before weights are considered.

FUNCTION evaluate_structural_action(action, evidence, policy)
    FOR each invariant IN policy.hard_invariants
        IF NOT invariant.PASSES(action, evidence)
            RETURN REJECT("hard invariant failed", invariant.id)
        END IF
    END FOR

    benefit <- WEIGHTED_SUM(
        task_utility = evidence.delta_utility,
        robustness = evidence.delta_robustness,
        coverage = evidence.delta_coverage,
        diversity = evidence.delta_diversity
    )

    cost <- WEIGHTED_SUM(
        memory = evidence.delta_memory,
        latency = evidence.delta_latency,
        energy = evidence.delta_energy,
        complexity = evidence.delta_complexity,
        risk = evidence.delta_risk
    )

    margin <- benefit - cost
    IF margin <= policy.minimum_margin
        RETURN NO_OP(margin)
    END IF

    RETURN ELIGIBLE_FOR_SHADOW(margin)
END FUNCTION

Calibration workflow

1. Score several historical decisions using only information available at the time.
2. Compare the calculated margin with the actual production outcome.
3. Adjust weights by workload and scope tier rather than globally.
4. Document uncertainty ranges for high-impact dimensions.
5. Re-score after shadow evidence; estimates should be replaced by measurements.

Common misuse

A team can make any candidate win by inflating utility or shrinking cost. Counter this by assigning dimension owners, recording evidence links, using ranges, and requiring an independent reviewer for high-risk changes. Where uncertainty dominates, the correct result is experiment, not approve.