Skip to main content

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.

Sideband detection is Mnemom’s fleet-level oversight. Three detectors — coherence, fault-line, and fleet — run on a configurable cadence and emit advisories when their thresholds fire. The thresholds are not hardcoded: they live in a Trust Posture. This guide shows you how to tune each detector for your fleet.
New here? Start with the conceptual overview at Sideband detection for the umbrella picture: what the four detector sources do, how findings cross to runtime via the two-turn dance, and how compliance evidence is preserved. The normative schema is the field-by-field reference for the posture body. The pending advisories schema documents the wire format consumers see.
If you’re new to Trust Posture, start with the overview and the Mnemom-shipped defaults. The normative schema is the field-by-field reference.

​
The three detectors

AxisFires when…Reads from
CoherencePairwise governance falls below floor, conflict edges climb past threshold, or outlier agents emergesideband.coherence
Fault-lineSeverity meets or exceeds the configured floorsideband.fault_line
FleetOutlier agents detected, min pair-score drops below threshold, or fleet partitions into incompatible clusterssideband.fleet
Each axis has an enabled flag, a cadence_seconds sweep interval, axis-specific thresholds, and a severity_on_fire value stamped onto the resulting advisory.

​
The strictest-wins cascade

Every team’s effective posture is composed Platform → Org → Team. Per-axis composition is strictest-wins:
  • enabled (bool): OR-true wins (any layer enabling fires the detector)
  • cadence_seconds (number): min wins (shorter = more frequent = stricter)
  • fire_on.* thresholds: min-among-defined wins (lower = fires earlier = stricter)
  • severity_floor: min wins (lower floor = more advisories = stricter)
  • severity_on_fire: max wins (higher = louder advisory = stricter)
You cannot loosen a detector below the org’s or platform’s floor by tightening at a lower scope. If your org sets coherence.cadence_seconds: 300, every team in that org will run coherence at least every 5 minutes regardless of the team’s own setting.

​
Workflow: clone Standard, tighten for a banking team

The expected workflow is clone the Mnemom default that’s closest to what you want, then tighten. The Mnemom defaults are immutable; you don’t edit them directly.

​
Step 1 — Clone the Mnemom default

Use the dashboard library at /settings/organization/postures (the “Clone Standard” CTA on the empty state) or the CLI: 
mnemom posture clone tp-platform-standard \
  --org org-acme \
  --slug acme-banking-high-compliance \
  --name "Acme Banking — High-Compliance"
The clone starts as a byte-for-byte copy of Standard’s body. From here, every change you make creates a forward-only revision.

​
Step 2 — Tighten coherence cadence

A banking team needs faster anomaly surfacing than the 10-minute Standard cadence. Open the new clone in the editor (/settings/organization/postures/{id}) and drop the cadence slider on the Coherence axis from every 10 min to every 5 min. The “Effective composed” column on the right updates immediately — strictest-wins folds the cadence to 300s. In YAML, the change is: 
sideband:
  coherence:
    cadence_seconds: 300   # was 600

​
Step 3 — Tighten the conflict-edge threshold

For a fleet where any disagreement matters, drop the conflict-edge threshold from 3 to 1: 
sideband:
  coherence:
    fire_on:
      conflict_edge_count_exceeds: 1  # was 3
The “Pinned by org floor” badge in the editor will appear on this field if your org default is already tighter than 1 — strictest-wins keeps the org’s value.

​
Step 4 — Raise severity-on-fire

When this detector fires, you want it loud: 
sideband:
  coherence:
    severity_on_fire: high  # was medium
severity_on_fire is unusual — it’s the one field where max wins instead of min. The reasoning: a louder advisory at a downstream layer should not be quieted by an upstream layer’s lower setting. When two layers disagree on loudness, the louder one wins.

​
Step 5 — Save as a new revision

Click “Save as new revision” in the editor. The dialog reminds you that posture revisions are forward-only — the current revision becomes a queryable historical entry; the new revision becomes live. Add a change_summary so the revision history reads as a narrative:
“Tightened coherence cadence + conflict-edge threshold for SOC 2 review. Severity raised to high.”

​
Step 6 — Assign to the banking team

On the Banking team’s detail page (/dashboard/teams/{teamId} → Trust Posture tab), the Posture Assignment Control surfaces the org library + Mnemom defaults. Pick the new clone and click Apply. The cascade tree updates immediately to show: 
Platform [tp-platform-standard rev 1]
  └─ Org [Acme default — rev 4 — Inheriting Mnemom default]
       └─ Team [acme-banking-high-compliance rev 2 — assigned]
The “Per-field winner” toggle expands to show which layer set each tunable field’s effective value.

​
Tuning fault-line detection

The same workflow applies to fault-line. The two knobs that matter:
  • severity_floor — findings below this severity are dropped. low lets every fault line surface; critical is “only show me the worst.” Most banking fleets settle on medium for the org default and high for noise-sensitive teams.
  • use_reputation_scores — when true, the detector weights impact_score by each agent’s reputation. This is on by default in Standard and should stay on for compliance-grade fleets — it gives more weight to fault lines involving low-reputation agents.

​
Tuning fleet detection

Fleet patterns are independent — each can be enabled or disabled separately:
  • outliers — fires when any agent is more than 1σ below the fleet mean on any value dimension. Always-on for healthy fleets.
  • min_pair_score_below — fires when the worst pair-score drops below this. A higher value fires more often (stricter). Banking-grade clones typically tighten this to 0.7; null disables the pattern entirely.
  • cluster_partition — fires when the fleet partitions into incompatible clusters (compatibility ≥ 0.7 in the SDK). Important for fleets where coordination across the whole population matters.

​
Disabling a detector

Set enabled: false to disable an axis entirely at any layer. Strictest-wins still folds upward — if your org enables coherence and a team disables it locally, coherence still fires for that team (OR-true wins on enabled). The only way to fully disable a detector is to set enabled: false at every layer (including the platform default, which Mnemom owns and won’t change for you).

​
See also