One of the first objections credentialing organisations raise when AI is proposed for grading is the black box problem. You put a submission in. A grade comes out. Nobody can see what happened in between. When an auditor asks, when a board member questions, when a candidate appeals, there is no trail to follow. The grading is opaque, and opacity is incompatible with the defensibility professional credentialing requires.
This is the right objection. Most AI grading tools are vulnerable to it. Assess for Learning is not, and the reason is architectural. The grading process does not run on a single opaque model. It runs on a layered rules engine where every evaluation criterion, every threshold, every aggregation step is explicit, inspectable, and auditable. The rules are text. The logic is structured. The execution is traceable. Transparency is not bolted on after the fact. It is how the platform is built.
“Transparency is not a feature, it is the architecture.”
What the rules engine actually is
The rules engine is the execution environment inside Assess for Learning that runs the grading process for every assessment. When a submission comes in, the engine walks through a layered set of evaluation rules, starting from the most atomic checks and building up through task-level and assessment-level aggregations until a final grading outcome is produced.
The rules themselves are the detailed evaluation criteria that were either written by the assessment designer or generated by the evaluation copilot and approved. They are stored as structured, text-based logic. Each rule specifies what it is checking in the candidate’s submission, what the expected pattern or reasoning looks like, how to score the result against the rubric, what feedback to produce if the criterion is met, partially met, or not met, and any dependencies on upstream tasks via short-term memory slots.
Every rule is visible. Every rule is editable. Every rule can be traced back to the assessment design decision that created it. There is no hidden layer of implicit judgement. The judgement is encoded in the rules, the rules are explicit, and the execution of the rules is recorded.
The layered architecture and why it matters
The rules engine is layered deliberately. Grading does not happen in a single pass. It happens in sequence, layer by layer, from the bottom up.
The four layers of the rules engine
- Layer one — Atomic evaluations: individual checks of the candidate’s reasoning at the sentence, paragraph, or formula level
- Layer two — Criterion-level evaluations: atomic checks combined into rubric criterion judgements
- Layer three — Task-level aggregations: criterion judgements combined into a task-level outcome, with weighting, dependencies, and short-term memory applied
- Layer four — Assessment-level aggregations: task outcomes combined into the final grading outcome, with minimum thresholds, mandatory criteria, and holistic overrides applied
This structure matters for two reasons. First, it produces explainability by construction. When someone asks why the candidate got the grade they did, the answer is a walk through the layers, showing which rules fired at which level and how the results combined. Second, it produces targeted auditability. If a specific grading decision is being challenged, the engine can show exactly which rules were involved and what their outputs were. The audit does not have to work with the final number. It can work with the full chain of reasoning that produced it.
Why this is the answer to the AI black box problem
The fear of AI in credentialing is based on the assumption that AI grading is inherently opaque. A neural network produces a score and nobody can explain why. That is true for some AI systems and it is a legitimate concern in high-stakes contexts. But it is not an inherent property of using AI in grading. It is a property of using AI without a structured rules layer around it.
Assess for Learning uses AI inside the rules engine, not as a replacement for it. The AI is the mechanism by which individual evaluation rules are applied to candidate submissions. It does the pattern recognition and reasoning assessment that would otherwise require a human grader for every rule. But the rules themselves are explicit, the aggregation is explicit, and the execution path is recorded. The AI is doing the work inside a frame that stays transparent.
The result is a grading system that has the speed and consistency benefits of AI with the explainability and auditability required by professional credentialing. The black box concern does not apply, because the box is not black. It has walls that are made of explicit rules you can read, edit, and audit.
Why this matters at the governance level
For governance and compliance leadership, the rules engine is the technical foundation that makes everything else in Assess for Learning’s governance story work. The precision report captures grading outcomes with psychometric rigour. The examiner’s report summarises cohort performance from grading data. The grading copilot assists human graders using the same rules. The evaluation copilot generates the rules that will be executed. None of these would be defensible if the underlying execution were opaque. They all depend on the rules engine being what it is: a transparent, structured, auditable layer where the grading actually happens.
This is what good AI governance looks like in practice. Not a separate compliance document written after the fact. Not a set of promises about how the system should behave. A technical architecture where the transparency is built into how grading is executed, so compliance is a property of the system rather than a statement about it. That approach survives an audit, a regulatory update, and a board challenge, because the evidence is in the structure itself.
The advanced layer: Excel analysis, diagnostics, and more
Inside the rules engine sit several advanced capabilities that sophisticated assessment designs can call on. Excel analysis extends the rules to examine formulas inside spreadsheet submissions. Diagnostic tagging maps evaluation outcomes to pedagogical frameworks and competency models. Short-term memory passes context between dependent tasks. Rule composition and logical operators allow complex grading conditions to be modelled explicitly.
These are not separate subsystems. They are extensions of the core rules engine, operating under the same transparency principles. Every advanced capability is configured through the same editable rules, runs through the same layered execution, and produces the same auditable trail. Sophistication does not come at the cost of transparency. It comes from the depth of what the transparent rules can express.
From trust me to show me
Credentialing governance is moving from a trust-me model to a show-me model. Regulators, awarding bodies, and boards are increasingly unwilling to accept assurances about AI behaviour without technical evidence to back them up. The platforms that will thrive in this environment are the ones that can show the grading in detail, rule by rule, layer by layer, submission by submission. The rules engine in Assess for Learning is built for that environment.
If your credentialing programme is being asked to defend its grading to increasingly demanding stakeholders, the architecture of the platform you use is about to matter more than the marketing around it. The rules engine is how Assess for Learning meets that new bar.
Ready to run AI-assisted grading on a transparent, auditable foundation?
Talk to us about how the Assess for Learning rules engine delivers the explainability your credentialing governance actually requires.