Shipped product updates that change how you validate, rank, or prove guardrails in production. Dates are when the capability reached marketing readiness — not every internal commit.

2026-10 — Policy simulator

A tighter rule should not meet production traffic for the first time. On an agent's validator editor, Simulate replays a sample or a redacted quarantine snapshot against the unsaved rules and diffs the outcome with the saved policy.

No model call, and the saved policy is not changed
Save stays available; the editor says when the diff is stale
Each run is audited as counts only — fixture text is not stored
Quarantine replay uses the failure shape, not the original text

Start here: validators · test validators in CI

2026-10 — Proof of guardrail execution

Auditors ask whether a rule actually ran. ValGuard seals each request with an HMAC evidence envelope: policy snapshot hash, ordered verdicts, and payload digests — not raw prompts.

Verify on request detail (Proof panel) or offline with valguard-cli verify-proof
Export a date window for auditors from Settings
Metadata-only retention stays the default; payload retention is opt-in

Start here: observability · verify guardrail proof tutorial · pillar article

2026-09 — Agent health score and ranking

Availability charts do not tell you which agents are quietly degrading. Each agent gets a transparent 0–100 health score (agent-health/1.1) from validation fails, blocks, upstream errors, latency, optional cost, and self-heal credit.

League table with rank, Δ7d, and CSV export under Agents → Statistics
Score-drop alerts when health falls vs the prior week
Optional Monday digest for the ops ritual

Start here: observability · monitor flows tutorial · pillar article

2026-09 — Response repair before re-ask

Re-ask burns tokens and latency. ValGuard prefers deterministic repair on structural failures (JSON shape, enums, required fields) before another model attempt.

Configure repair / block / reask / warn per rule (disabled · shadow · enforce)
Estimate savings with the response repair ROI calculator
Same OpenAI-compatible API — no client rewrite

Start here: validation · self-healing without re-ask tutorial