Companion to: What It Takes to Trust a Probabilistic AI in Deep Enterprise

Governor–Engine Scaffolding Model

An attempt to make GenAI outputs defensible in high-stakes environments. The Engine generates. The Governor judges. The loop runs until the output has earned its stability, or the Governor decides it won't.

ACTIVE
REV 2.1 · 2026
CONTEXT: CHAT
MODEL-AGNOSTIC: YES
Core Thesis
Trustworthiness is not a feature of the Engine. It never was. It's a result of who's holding the reins, and how deliberately they hold them. The Engine generates. The Governor judges. That hierarchy isn't a limitation of the technology. It's what I kept coming back to when trying to design for environments where someone has to answer for the output.
§ 01 The Framework Four objectives. Three stages. One feedback loop, as far as I could take it.
Obj 01
Verification & Hallucination Control
GenAI hallucinates at 3–15% depending on domain. In a bank, that's not a quirk. It's a liability. The first thing this scaffold tries to do is catch the problem inside the loop, before anyone acts on it.
E-02 Internal Loop
Obj 02
Auditable Transparency
A vibe-check isn't a governance strategy, and I kept seeing that pattern in the field. The aim here is to get the Engine to show its reasoning, not just its conclusion, so the Governor has something real to interrogate rather than a confident-sounding output to accept on instinct.
E-03 Sensor Report
Obj 03
Reproducible Framework
A process that only works once isn't a process. The three-stage scaffold is an attempt to behave consistently across similar inputs, so governance doesn't have to be reinvented every time a new use case lands on the table.
3-Stage Engine Loop
Obj 04
Portable Governance Artefacts
The scaffold lives in a single, model-agnostic prompt. The idea was to make something portable enough that a colleague with basic AI exposure could pick it up, adjust the thresholds to fit their risk tolerance, and run with it. No ML engineer required. That was the aspiration, at least.
G-01 Set-Point Config
§ 02 The Model Click any node to expand its definition. Click the three stages to see prompt patterns.
⬡ Engine Layer · AI Processing
G-01
System Goal
The Governor inputs objectives. Not just the formal brief, but the context the Engine can't see. Office politics. Stakeholder sensitivities. What "good enough" actually means here.
Obj 04
▶ Goal →
E-01
Stage 1
Internal Verification
Generates a draft and checks its own output. Audits for factual inconsistencies, hallucinations, and logical gaps. Reduces false entropy before anything else runs.
Obj 01 Obj 03
▶ Verified →
E-02
Stage 2
Adversarial Red-Hat
Stress-tests its own Stage 1 output. Adopts an adversarial stance to surface what it suppressed or over-weighted. Introduces useful entropy that the first pass wouldn't have caught.
Obj 01 Obj 03
▶ Scored →
E-03
Engine Scores
Output
Scores the Stage 2 output against SS and DR. If either fails, the Engine loops back to Stage 1 autonomously. GC only enters the scoring once Stage 3 grounding has been injected. The Governor sees nothing until all active metrics pass.
Obj 01 Obj 02
▶ Present →
E-04
Post-Verified
Output
Presents the scoring and reasoning audit to the Governor. This is what the Governor actually reads: not just the answer, but the working behind it.
Obj 02 Obj 04
Stage 1 · Internal Verification
Stage 2 · Adversarial Red-Hat
Stage 3 · Last-Mile Grounding ◈ Governor
↓ scored output + reasoning audit presented to Governor ↓
◈ Governor Layer · Sovereign Human
G-02
Governor Reviews
Output
Reads the audit trail and scores against the original system goal. Has it drifted? Is it telling me what I want to hear? This comparison is deliberately human. Automating it would outsource the judgment that makes the output defensible.
Obj 01 Obj 02
▶ Decides →
G-03
Governor
Decides
Accepts the output, or judges it unsatisfactory. If unsatisfied, Stage 3 is injected. In the field, this is where accountability either lands somewhere legible or quietly dissolves into the workflow.
Obj 01 Obj 03
▶ If unsatisfied →
G-04
Stage 3
Last-Mile Grounding
Governor adds context to fine-tune the output. Injects the organisational specifics, constraints, and real-world framing the Engine could not access. Output re-enters Stage 1 with this grounding baked in. Performing Stage 3 is also what activates GC scoring on the next loop.
Obj 02 Obj 03
[ close ✕ ]
Power Dynamic
Governor: the human with context, authority, and accountability for the output
Engine: fast, generative, pattern-matching at scale. Useful precisely because it's bounded.
The hierarchy is intentional. The Engine doesn't get to validate its own outputs.
Interop Modes
Chat NL prompt
CLI stdin/args
Agentic Tool schema
Engine Core Shared
Gov Config Per-context
Control Params
SS Stress Score
DR Drift Rate
GC Grounding Check
SS and DR run every loop. GC activates from loop two onwards, once Stage 3 grounding has been injected.
Loop Stability
Accepts: Governor is satisfied, loop closes, output is finalised
Unsatisfied: Stage 3 injected, Governor adds grounding, re-enters Stage 1
Stage 3 is optional. If the Governor accepts on first review, it is never performed.
§ 03 Scaffolding Method Generate prompts you can use directly. Adjust thresholds to fit your context.
Scaffolding Method
You run each stage manually, turn by turn. You are the loop.
MANUAL AI CHAT · You are the Governor, operating step by step. Paste Stage 1 into any AI chat interface, read the output, then paste Stage 2. If you are satisfied with the scored output, the loop closes. If not, inject Stage 3 grounding and re-enter Stage 1 with corrected framing.
Manual AI Chat · Prompt Guide Generator
Generates staged prompts to paste sequentially into any AI chat interface. Stages 1 and 2 are for the Engine. Stage 3 is yours to inject if you are unsatisfied with the output.