Verdict Console Filament Laravel Package

Product Decisions This Supports

• Auditability as a default, not a bolt-on. The suite records each agent call (now carrying Laravel AI 1.0's widened prompt() input union and TextUsage), so PMs can answer "what did the model see, and what did it cost?" per request without building an internal logging layer.
• Human-in-the-loop gating. Verdict 0.16/0.17's review_outcome and review_request_fingerprint columns are now first-class in the evidence browser. That makes review decisions and request identity queryable — the prerequisite for deciding whether AI output ships directly or queues behind a reviewer, and for measuring reviewer agreement over time.
• Build vs. buy on AI observability. The realistic alternative is a homegrown audit table plus ad-hoc queries. The console (evidence browser + doctor baseline) is thin glue over a data model you'd otherwise invent, so this is usually a "buy/adopt" unless you need vendor-neutral telemetry.
• Governance posture is an explicit setting. v0.4.0's doctor baseline records the verdict-console.evidence.accepted_off decision the way a host would, turning "we deliberately don't record evidence" into a reviewable, install-time flag rather than a silent omission. Useful when someone asks why a given environment has no trail.
• A shippable readiness gate. A clean install reading clean under doctor gives you a defensible release checklist item: "evidence path verified" before an AI feature goes to production.

When to Consider This Package

Adopt when:

• You're on Laravel with laravel/ai and want an evidence/audit trail around model calls without building one.
• Reviewer workflow matters — you need outcomes tied to a specific request fingerprint, not just a free-text comment.
• You want console-level inspectability (browse evidence, run health checks) rather than raw SQL against an internal table.
• Your team can absorb a pinned-dependency upgrade cadence.

Look elsewhere when:

• You're not on laravel/ai — v0.4.0 hard-couples to it via verdict-console ^0.11.
• You need vendor-neutral, multi-language AI observability (Langfuse, Helicone, OpenTelemetry-style tracing); this is Laravel-native by design.
• You need contractual retention/immutability guarantees. This is evidence tooling, not a compliance-certified store.
• You cannot take a coordinated breaking bump: the console is still 0.x, and it now drags laravel/ai ^1.0 and verdict ^0.18. If your app is pinned below AI 1.0, adoption is blocked until you upgrade.
• Your AI surface is a single low-stakes call — the evidence model is more machinery than the risk justifies.

How to Pitch It (Stakeholders)

For executives: This gives us a defensible record of what our AI features were asked, what they returned, and who reviewed it — at the cost of adopting a small Laravel-native package instead of funding an internal audit-log project. It also makes "we chose not to record evidence here" a written decision rather than an accident, which is the part auditors and legal tend to ask about. The trade-off is dependency churn: the package is pre-1.0 and pins to laravel/ai 1.0 and Verdict 0.18, so expect a small, scheduled upgrade tax each release rather than a set-and-forget install.

For engineering: v0.4.0 is a breaking alignment release. It moves us to fissible/verdict-console ^0.11, which requires laravel/ai ^1.0 and fissible/verdict ^0.18. Concretely: the queue-recording agent adopts AI 1.0's widened prompt() input union and TextUsage; browser fixtures pick up the two evidence columns added in Verdict 0.16/0.17 (review_outcome, review_request_fingerprint); and the doctor baseline now records the explicit verdict-console.evidence.accepted_off decision so a clean install reads clean. Plan this as one coordinated bump of the AI and Verdict stack, expect fixture-level updates only, and treat a green doctor run as the migration's acceptance test. No product-visible behavior change is implied — the risk is version coupling, not functionality.