Nine Gates

AFA evaluates every enhancement candidate through nine quantitative gates. All nine must pass for an enhancement to be committed. If any gate fails, the enhancement enters a redesign loop (up to 3 retries) or is rejected.

Gate summary

Evaluation order

Gates are evaluated in a specific order to fail fast on cheap checks:

1. Complexity (hard floor first -- instant reject if < 0.5)
2. Risk / Profit (disjunctive -- either suffices)
3. Novelty
4. Quality
5. Utility
6. Entropy (includes anti-gaming checks)
7. Supply Chain
8. KPI

Gate details

1. Risk gate

Measures how much an enhancement reduces risk relative to the baseline.

Delta_Risk = baseline_risk / max(proposed_risk, epsilon)

Gate passes when Delta_Risk >= 2.0 OR Delta_Profit >= 2.0 (disjunctive -- either suffices).

2. Profit gate

Measures improvement in profit proxy (deployment velocity, error reduction, etc.).

Delta_Profit = proposed_profit / max(baseline_profit, epsilon)

Both risk and profit deltas are validated through Bayesian posterior probability: P(Delta >= 2 | data) > 0.95.

3. Novelty gate

Prevents trivial changes from consuming pipeline resources. Uses a logistic function:

G(N) = 1 / (1 + e^(-k * (N - N0)))

Where N is the number of new concepts introduced, k is the logistic slope (default 5.0), and N0 is the logistic center (default 3.5). Gate passes when G(N) >= 0.8.

4. Complexity gate

The only non-overridable gate. See Complexity deep-dive.

5. Quality gate

Aggregate quality score across security, performance, maintainability, and documentation. Gate passes when Q >= 0.7 AND no individual sub-score equals zero.

The zero-sub-score rule prevents gaming: an enhancement cannot have perfect security but zero maintainability.

6. Utility gate (economic optimization)

Answers "is this enhancement worth the cost?" using a utility function with a lower confidence bound:

U = (Delta_P_H + gamma * Delta_V_L) + kappa * |Delta_R| - (phi_S * Delta_C_S + phi_D * Delta_C_D) - Delta_OPEX
LCB(U) = U - z_alpha * sqrt(Var(U))

Gate passes when LCB(U) > 1.5 (configurable). Uses PERT three-point estimation for variance.

7. Entropy gate (anti-gaming)

Prevents trivially small or artificially inflated changes by measuring Shannon entropy of the diff token distribution:

H(diff) = -sum(p_i * log2(p_i))

Gate passes when H(diff) >= 0.25 (configurable). Also runs three anti-gaming detectors:

8. Supply chain gate

Validates build integrity: SLSA level check, signature validity, and SBOM change allowlisting. Gate passes when all three conditions hold.

9. KPI alignment gate

Ensures enhancements align with business objectives. Gate passes when R-squared >= 0.6.

What happens when a gate fails

1. Near miss (all scores within 1.1x of their threshold): eligible for redesign retry
2. Hard failure: enhancement is rejected
3. Hard floor violation (complexity < 0.5): immediate rejection, no retry

Use afa explain <enhancement-id> to see exactly which gates failed and why.

Operating modes

In standalone mode, gates use simple threshold comparisons. With the optional AEGIS connector, gates use full Bayesian posterior evaluation with confidence intervals, SHA-256 audit hash chains, and governance workflows.