Documentation Index

Fetch the complete documentation index at: https://docs.mnemom.ai/llms.txt

Use this file to discover all available pages before exploring further.

​

Managed Rules

A Managed Rule is the control-plane state that wraps a signed detection recipe propagated to every gateway in the network. Recipes are detection content; rules are the control-plane state; both compose through the same machinery as agent cards. When AEGIS promotes a detection pattern — from the arena, from a customer FN/FP report, or from the cross-tenant aggregator — the result is a Managed Rule that lands at every affected customer’s gateway under the P95 ≤ 30s propagation target. This page documents the rule pipeline: how a candidate becomes a promoted rule, how the protective invariant is enforced, how the three signing chains compose, and how the 24-hour observe soak governs mode escalation.

​

The signal-loop pipeline

AEGIS Managed Rules are sourced from three signal loops plus the substrate-fingerprint cross-cutting dimension; see AEGIS — three signal loops. Every candidate row is stamped server-side with a writer_identity by the auth context used at write time: The customer never sets the writer identity. It is the server-side stamp that downstream review uses to scope auto-promotion eligibility (see Three reviewer modes below).

​

Three reviewer modes

The graduated-trust pattern used by Cloudflare WAF Managed Rules, AWS Shield Advanced, and CrowdStrike Falcon. Confidence determines auto-eligibility; convenience is the customer’s posture choice. Mode is set platform-wide via the platform-admin config. The manual mode is the default for new platforms. Auto-modes are available to flip platform-wide via the platform-admin config page. The choice between modes only affects tier-3 candidates; tier-1 and tier-2 promotions are governed by the protective invariant below.

​

The protective invariant

A tier-1 or tier-2 rule — one that would actually block real production traffic — can never promote without a human in the loop, regardless of how aggressive the reviewer mode is set. Auto-modes only accelerate tier-3 (observe / nudge / log) rule landing, where the blast radius of a bad call is bounded. The invariant is enforced structurally by the schema. A CHECK constraint requires promotion_quorum_met = TRUE whenever status = 'active'. The database refuses to write an active row that has not collected the required dual-control approvals. The invariant is not procedural — it is not a workflow that a misconfigured handler can skip; it is a constraint the storage engine enforces on every insert and update.

​

Signed promotion and the three signing chains

Every promotion produces a signed row plus two parallel signed envelopes. The three keys are independent — they rotate independently, are escrowed independently, and would have to be compromised independently to fully poison the rule plane. Three independent compromise paths required to fully poison the pipeline. KV poisoning degrades gracefully to R2 (different signing chain). KV + R2 compromise is still defeated by the per-row promotion signature: even if both envelopes are forged, the gateway verifies each row against the promotion signature before evaluation, and a forged row will not match. See Managed rule envelope schema for the wire format and /trust/slos for the published SLOs on the failover behavior.

​

Propagation, soak, and auto-rollback

A promoted rule transits four stages before it reaches its target enforcement mode:

1. Sign + dual-write. Promotion writes the per-row signed entry, then dual-writes the KV and R2 envelopes. Both envelopes share the same canonical recipe set; only the signing key (key_id) differs.
2. Gateway propagation. The gateway’s tiered read pipeline picks up the new envelope within the P95 ≤ 30s propagation target.
3. 24-hour observe soak. The rule enters observe mode regardless of its target mode. During the soak, the aggregator computes the per-recipe false-positive rate; the 7-day rolling FP threshold governs whether the rule survives the soak.
4. Escalation or auto-rollback. If the soak passes, the rule escalates to its target mode (nudge or enforce). If the soak fails on FP, the rule is auto-rolled-back under CLPI Phase 2 trust recovery; the recipe.retired webhook fires with reason = 'observe_soak_fp'.

The soak is not bypassable. A platform admin cannot manually escalate before the 24-hour window completes; the only way to bring a rule to enforce faster is to publish it as enforce after the soak, never instead of it.

​

The five GA-seeded synthetic Managed Rules

The Day-1 Managed Rules are sourced from real production detection content meeting four bars:

• status = 'active' in the canonical detection-recipes registry.
• Platform-scope (no org or agent specialization).
• hit_count > 0 (proven detection signal).
• FP-rate below threshold over the trailing window (proven low false-positive history).
• Tier 3 (low blast radius for single-operator dual-control constraint).

Five rules selected via the GA seeding script and promoted through the full pipeline:

1. Candidate insert via internal-observation writer identity.
2. created and approved review actions chained in the append-only audit log.
3. Signed per-row promotion (promotion_signature via RECIPE_PROMOTION_SIGNING_KEY).
4. KV envelope write + R2 envelope write (independent signing chains).
5. Initial observe mode for 24 hours.
6. After soak, escalation to enforce.

By GA, the L3 surface shows five real Managed Rules in enforce with effective-since timestamps. Customers see live network protection on Day 1; the rules are real detection content, the signatures are real, and the propagation latency is real. What is synthetic is the framing — they are platform-internal patterns promoted through the same pipeline a future cross-tenant detection event would use, not detections of an actual prod attack campaign.

​

Webhook events

The rule lifecycle emits webhook events to the customer’s subscription. See Webhooks for the subscription setup and Webhook contract for the canonical payload shape.

​

See also

• AEGIS — the overall protection-layer framing
• Protection Network — the five-layer L0-L5 model Managed Rules push into (L3)
• Managed rule envelope schema — the wire format on the KV and R2 envelopes
• CLPI — the governance layer that handles auto-rollback (Phase 2)
• Protection Card — the per-agent configuration Managed Rules tune at the gateway
• Card Composition — the scope cascade Managed Rules compose into
• /trust/slos — published SLOs for propagation, freshness, failover