Meta and Confluence¶
Overview¶
M rules take other rules' findings as their input. Every other series reads the recipe; this one reads the finding set.
An M finding never lowers a score and never duplicates a line: its input is the finding multiset after the ownership resolver, so two rules that matched the same line have already collapsed to one owner by the time M runs. Its value is the relationship - a practice the recipe claims and a finding that contradicts it, or several weak signals whose combination is the story.
This is a reference page. For how weight and scope work, see Rule Mechanics.
Reading an M finding¶
- It reads the profile, not the diff: the evidence names the findings that composed, and the recipe text they came from.
- M001 scores (HIGH); M002 and M003 are weight-0 but escalate the band one
step (never add points), gated on the benign-corpus cap
M escalation stays below the benign cap. - It fires at the layer of its inputs, so it has no static layer.
Threat context¶
An immune system that attacks the body it lives in is an autoimmune disease: the danger is not the pathogen but the response. The meta family is TrustSight's self-tolerance. Every other rule is a claim about the package; meta is a claim about the run: too many alerts for one window, a matcher that fired suspiciously often, a campaign class re-arming again and again. These are tool-health observations, and they exist so the analysis can admit when it is not itself trustworthy. They never score against the package, because a detector's own over-excitement is not evidence about the diff.
The name "confluence" is the other half, and it is the immune system at work again. A single immune cell binding an antigen is not enough to launch a full response; it needs corroboration from other cells before the alarm is real. M003 reports exactly that: independent families agreeing on the same package, a statement about agreement rather than about any one rule. Both halves answer the same question - "how much should this run's own output be trusted?".
Detection mechanisms¶
The rules that make up this category are listed below, each with the series that implements it and the severity it carries. Scoring, matching and identifier allocation are shared across every category: see Rule Mechanics.
Individual rules¶
| Rule | Name | Severity |
|---|---|---|
| M001 | Claim Contradicted By Structure | HIGH |
| M002 | Weak Signal Across Layers | MEDIUM |
| M003 | Composition Not Owned Elsewhere | MEDIUM |
| M004 | Expected Co-fire Absent | INFO |
| M005 | Licensor Absent On Licensed Shape | INFO |
| M006 | Campaign Re-armed Too Often | INFO |
M001: Claim Contradicted By Structure¶
- Target: programmatic (finding set)
- Severity: HIGH (weight 25)
-
Category:
meta -
Condition: The recipe's declared practice contradicts a structural finding in the same run.
It joins the real claims ledger - the P-series ids the run's declared facts imply (trustsight.scoring.declared_claims) - against the structural findings: checksums declared (P001) while H091 reports the arrays do not cover every source, GPG verification declared (P002) while H024 reports it removed, a source pinned (P005) while H101 reports the pinning lost, a trusted forge declared (P007) while H099 reports the host swapped, a prebuilt deliverable declared (P009) while H066 reports an ELF committed to the repository outside source=().
Only M can see this. The declared-practice ledger is weight 0 and the structural finding is its own claim; "the recipe says one thing and structurally does another" is a relationship between two findings that no single rule reads.
A fourth M family is a tool-health check, never a package finding:
cofire_absence reports when a construct that historically co-fires with a
rule appears without that rule - a sign the rule broke or was evaded in a
known shape. It is consumed by the harness/calibration, not the report,
because a hit is about the tool, not the package.
M002: Weak Signal Across Layers¶
- Target: programmatic (finding set)
- Severity: MEDIUM, weight 0 (annotation)
-
Category:
meta -
Condition: Three or more distinct assurance layers fired, none above the alert bar (
[thresholds] m002.min_layers, default 3).
An attacker who spreads weak signals across layers stays under any single threshold; the span is only visible from the whole profile. It adds no points but raises the reported band one step: a spread across layers is exactly the case a single-signal threshold cannot see. Measured on the locked benign corpus: 112 on the v1 corpus (3.0%).
M003: Composition Not Owned Elsewhere¶
- Target: programmatic (finding set)
- Severity: MEDIUM, weight 0 (annotation)
-
Category:
meta -
Condition: Two or more distinct capability families co-occur while none of the cluster owners (H027, H098, X007) recorded the composition - typically because the owner is disabled.
The composition is never lost. It adds no points but raises the reported band one step. Measured on the locked benign corpus: 316 on the v1 corpus (8.5%).
Standing gate: the escalation passes the benign-corpus cap (M escalation
stays below the benign cap, MAX_BENIGN_FIRE_RATE) before it ships, the
cluster-rate gate the M constitution requires.
M004: Expected Co-fire Absent¶
- Target: programmatic (tool health)
- Severity: INFO, weight 0
- Category:
meta
A construct that historically co-fires with another rule appeared without
it. A hit means the tool, not the package, needs attention, so this is
never added to a package's finding list: it rides the report's
cofire_health channel for the harness and calibration.
M005: Licensor Absent On Licensed Shape¶
- Target: programmatic (tool health)
- Severity: INFO, weight 0
- Category:
meta
A rule declaring escalates_when fired while none of its declared licensors
did - the detector travelled without its usual company. Like M004 this is a
sensor-health observation, not evidence about the package, so it goes to the
cofire_health channel and never scores. See
Escalation Licensing.
M006: Campaign Re-armed Too Often¶
- Target: programmatic (tool health)
- Severity: INFO, weight 0
- Category:
meta
A campaign re-armed faster than
[alerts] campaign_min_reactivation_observations allows. The window is still
refreshed - re-activation is never blocked - but a campaign that re-primes over
and over is the instrument churning, not a package misbehaving, so it is
surfaced as a tool-health note rather than passed over in silence. Unlike M004
and M005 it is a cycle-level observation: it is raised in the correlation
cycle's log, not attached to any package's cofire_health, and it never
scores. See Campaign Alert Mode.
Interpretation¶
- An M finding is about TrustSight, not the package under review. It never scores against the package and is excluded from the package's risk band.
M001(volume),M002(a noisy matcher), andM004toM006(cycle/campaign health) are alerts to the operator to retune or investigate filters.M003(confluence) is the exception that reads a package: several distinct families agreeing is a real signal, and it is scored as such.M006specifically flags a campaign re-armed too often: the state is still refreshed, but the churn is reported so a chronic campaign does not inflate the alert stream.
Related documentation¶
- Kinds of Alerts; operator-facing versus package-facing findings
- Alert Budget