Architecture thinking

GxP Evidence Architecture

How regulated decisions actually happen — and where evidence has to be captured to remain reviewable later.

GMP is a time, context and decision problem — not a documentation problem.

Cloud
Contextualizes and recommends.
Line
Decides fast. Irreversible.
Site
Decides formally. Approval required.
Audit
Built from snapshots, not logs.

Scope: this page is an architecture model, not a list of features shipping in TrailMQ 3.1.0. Current core capabilities are MQTT transport, two-gate policy enforcement, attributed decision records and a scope-defined system/action chain. Context extraction, historical baselines, deviation calculation and decision traces are planned.

Perspective
Decision layer

Proximity

GMP scales with proximity

The closer to the physical process a decision is made, the faster and more irreversible it becomes — and the harder the evidence is to reconstruct later. Switch perspective above; filter by decision layer to focus on one system level.

Line

Decides fast. Irreversible.

At the line, control decisions happen in milliseconds and physically commit. If the surrounding context is not captured at that instant, it is gone — there is nothing to reconstruct afterwards.

ms · irreversible
Site

Decides formally. Approval required.

Release, deviation and disposition decisions are deliberate, reviewed and signed. They are slower by design — accountability matters more than speed.

reviewed · signed
Cloud

Contextualizes and recommends.

Cloud and analytics layers can use aggregated history to surface signals and support review. Decision authority and required evidence should follow the intended use and risk assessment.

context · recommendation
Line

Real-time window

Context exists for milliseconds to seconds. If it is not recorded at decision time, it cannot be reconstructed — the value loses its meaning the moment it passes.

ms – s
Site

Batch & shift window

Decisions span a batch or a shift. Intervals are not abstractions — they are commitments that define meaning, responsibility and comparability.

min – h
Cloud

Historical window

Baselines and trends are built over days to months — from structured snapshots, not raw logs. This is the reference live values are compared against.

days – months
Attributable

Who & what

Every record carries the identity behind it — which client, user or service triggered the event, and under which permissions.

Contemporaneous

At the moment

Evidence is captured when the decision happens — not assembled after the fact from scattered logs.

Original

First record

The structured record is the source of truth, not a copy or a transformation that can drift from the original.

Accurate

Checked against intended context

Conceptually, a value should be evaluated against the context and reference that apply. TrailMQ context and baseline plugins are planned, not current.

Complete + Consistent

Missing context stays explicit

A complete evidence design records that required context was unavailable. Deferred context-driven calculations are part of the planned plugin direction.

Enduring + Available

Scope-defined integrity

TrailMQ's local hash-linked check covers system and action entries. It is not WORM storage, external notarization or a check over the separate MQTT decision store.

Line

Accept / Block

Current core: role permission and namespace/topic policy are enforced inline, and the attributed MQTT outcome is recorded.

current core
Line

Rate-limit

Architecture direction: protect downstream systems from floods while keeping throttling decisions visible.

planned
Site

Defer / Queue

Architecture direction: make a missing-context calculation explicit and reviewable instead of silently skipping it.

planned context flow
Site

Escalate / Approve

Architecture direction: route a GxP-relevant decision to the organization's approved human workflow.

conceptual
Cloud

Flag / Recommend

Architecture direction: surface anomalies and recommend a review while decision authority remains defined by intended use, risk and approved procedures.

conceptual

Planned plugin scenario: the seven steps below illustrate the proposed live-versus-historical recipe. They are not executable in TrailMQ 3.1.0.

1

Machine publishes a live value

An OEE / temperature value arrives over MQTT from the filler on line 1.

Line
2

Domain context extracted

Line, machine, batch and metric are resolved from the topic and payload.

Line
3

Historical baseline resolved

The expected reference value is fetched from the historical context feed.

Cloud
4

Deviation calculated

Live vs. baseline is compared; the deviation and its severity are computed.

Line
5

Decision gate evaluated

The proposed decision trace explains whether the value was accepted, blocked, rate-limited or escalated.

Line
6

Missing context is deferred

If a baseline is unavailable, the calculation is queued for review rather than skipped.

Site
7

Evidence linked for review

The proposed plugins link value, context, baseline, deviation and decision trace into a reviewable path.

Site

This is a conceptual model of how regulated decisions and evidence can relate across system layers. It informs product direction; it is neither a certification claim nor a statement that every illustrated capability ships today. Read the underlying concepts →