Backlog optimizer
The backlog optimizer decides what to fix first. A Mixed Integer Programming model with ML warm-start schedules the accessibility, performance, and sustainability backlog under sprint capacity, severity, dependency, and budget constraints. The output is a prioritized PR queue that the LLM remediation cascade consumes and the team actually ships. Not a flat severity list — a Pareto-front of severity, effort, and dependency.
What it does
-
MIP planner
A Mixed Integer Programming solver picks the subset of findings that maximizes regulatory and severity coverage inside the available sprint capacity. Hard constraints (dependency, budget) are honored exactly; soft constraints (team preference, theme) are penalised.
-
ML warm-start
A learned model proposes a high-quality starting solution so the solver converges quickly. Training signal comes from historical merge / reject / rework outcomes per repo — the planner gets sharper for each customer over time.
-
Dependency-aware ordering
Findings reference each other — fix WCAG 1.4.3 contrast first, then 1.1.1 alt-text, defer 4.1.2 because the underlying widget is being rewritten. The optimizer respects the graph; ungated remediation does not.
-
Cross-axis scheduling
The same MIP+ML core schedules accessibility (L3) backlog, performance (L7) backlog, and sustainability (L8) backlog. Different rule packs, one optimizer — teams stop fighting cross-axis priority wars.
-
Sprint-capacity budget
Capacity is set per-sprint by the team lead. The optimizer fills it — and only it. No backlog explosion, no alert fatigue, no "fix all 4,217 findings this quarter" fantasy.
-
Pareto frontier output
The optimizer can emit several non-dominated plans — "AA in two sprints, AAA in six" — so leadership picks the regulatory posture, not the engineer. The cascade then executes the chosen plan via the LLM remediation cascade.
Layer mapping
The backlog optimizer orchestrates remediation across three scanner-axes simultaneously, then hands the chosen plan to the LLM remediation cascade for execution.
| Axis | Direction | What flows |
|---|---|---|
| L3 Testing (WCAG/EAA) | Inbound — backlog | WCAG / EN 301 549 findings with severity, effort estimate, and dependency edges. |
| L7 Performance | Inbound — backlog | Core Web Vitals / Lighthouse opportunities scheduled by the same MIP+ML core. |
| L8 Sustainability | Inbound — backlog | WSG / SWD-model findings (page weight, energy proxy) sequenced into the same plan. |
| LLM remediation cascade | Outbound — queue | Ordered finding queue with dependency annotations — the cascade executes in order. |
| Canonical scoring | Inbound — severity | The canonical score feeds the optimizer's severity weights. |
| CI/CD gate | Inbound — constraints | Budget thresholds from the CI/CD gate become hard constraints in the MIP. |
Filed IP
ARIADA holds filed-IP positions covering the MIP+ML scheduler, dependency-graph-aware ordering, cross-axis (accessibility / performance / sustainability) joint optimization, and Pareto-frontier plan emission at the core of this module. Provisional application only; conversion to non-provisional and PCT national-phase decisions are pending within the 12-month window.
Application numbers, claim counts, and PCT deadlines are available for accredited-investor due diligence on the Legal & IP page.
Why most teams fail without it
The default remediation workflow is "list findings, sort by severity, hand the top of the list to engineering." That collapses the moment two findings share a dependency or a single sprint cannot absorb the chosen subset. Severity is a scalar; a backlog is a graph. Engineering then re-prioritises by gut feel and the regulatory posture drifts.
The optimizer treats the backlog as constrained optimization. The same MIP+ML core that schedules accessibility issues also schedules Core Web Vitals work and sustainability fixes; teams get one ranked plan instead of three competing lists. Sales pitch: "we get to AA in two sprints, not six; the optimizer figured out the ordering."
The optimizer orders the queue; the cascade writes the patches; the CI/CD gate gates the merge; the AI attribution module attributes the AI author. Each module owns one decision; planning is the optimizer's.
Cross-references
- LLM remediation cascade: consumes the ordered queue and emits source-level pull requests. See the cascade →
- CI/CD gate: supplies budget thresholds to the optimizer and gates the resulting PRs at merge. See the CI/CD gate →
- Canonical scoring: feeds normalised severity weights into the MIP objective. See canonical scoring →
- Cross-portfolio: the same MIP+ML core also schedules the L7 perf and L8 sustainability backlogs, sharing the engine across rule packs.
- Triad context: the backlog optimizer is the planner inside the Remediate tier; see the homepage triad for Architect · Detect · Remediate framing.
Source-level remediation only — agentic suggestions are not autonomous deployments; pull requests require client merge. Not a legal certification body.