The Tier Decides Everything

The Examiner’s Question · RBI’s MRMF 2026, read like evidence. By Dr. Anandkumar Prakasam

The Clause

Paragraphs 14 to 20 build the engine room of the framework. Paragraph 14 requires ongoing risk assessment at both entity and enterprise level, with action when the appetite is breached and reporting to the risk committee. Paragraph 15 installs three lines of defence, the owner, an independent model risk and validation function, and internal audit. Paragraph 16 demands continuous performance testing, backward and forward looking, with AI-specific evaluations where relevant. Paragraphs 17 to 20 then create the tiering system. Every model gets a risk tier. The tier is reviewed at least annually and on triggers. The tier drives everything downstream, validation intensity, approval level, control depth, monitoring frequency, documentation and continuity planning. Tiering weighs materiality, complexity and the difficulty of explaining the model, and paragraph 20 adds the quiet killer, factors cannot offset each other into a comfortable composite.

What It Actually Asks

The tier is the single most consequential number in the entire framework, because it prices every other obligation. Get tiering wrong and everything downstream is wrong in proportion, validation too shallow, approvals too junior, monitoring too sparse, and all of it defensibly documented against the wrong tier. Paragraph 19’s factor list is deliberately uncomfortable. Materiality and complexity are different axes, and the draft refuses to let one excuse the other. A simple model deciding something enormous is high tier. A complex model deciding something trivial may not be. Paragraph 20’s no-offsetting rule exists because every scoring committee in history has discovered the average, and the average is where risk goes to hide. The annual review plus trigger-based re-review in paragraph 17 means the tier is a living judgment, not a birth certificate. An examiner will test tiering the way an auditor tests depreciation, by finding the asset whose classification flatters the books.

The Incident File

In the summer of 2020, with examinations cancelled by the pandemic, England’s qualifications regulator Ofqual assigned A-level grades using a standardisation algorithm. As models go it was almost trivial, a statistical adjustment built on each school’s historical grade distribution and the rank order of its students. No neural networks, no black box, a method a competent analyst could explain in an afternoon. On results day it downgraded close to forty percent of teacher-assessed grades. Students from large state schools were pulled down by their schools’ past performance while small private school cohorts, statistically exempt from parts of the adjustment, kept their teachers’ optimism. University places evaporated on the basis of a postcode’s history. Within four days of public fury the government abandoned the algorithm entirely and reverted to teacher assessment. Judged by complexity, the model deserved the lowest tier in any framework. Judged by materiality, it was deciding the life chances of a national cohort in a single morning. Any tiering scheme that let its simplicity discount its consequence, exactly the offsetting paragraph 20 prohibits, would have signed off on the summer’s disaster in advance.

The Trap

Tiering fails through arithmetic and through appetite. The arithmetic failure is the weighted average, a scoring sheet where a five on materiality and a one on complexity settle into a soothing three, which is precisely the composite paragraph 20 forbids. The appetite failure is quieter. Tier assignments drift toward what the validation budget can afford, because every tier upgrade is an unfunded mandate for the second line. When the same committee owns the tiering decision and the validation resourcing, tiers become a budget instrument. An examiner finds this in an afternoon by sampling the models tiered just below each boundary and asking for the scoring worksheets. A suspicious cluster one point under the high-tier line is not a coincidence. It is a spreadsheet negotiating with a headcount plan.

The Fix

Make the tier a floor-taking function, not an average. The highest scoring factor sets the tier unless a documented, senior-approved rationale says otherwise, and that rationale is itself an exception the risk committee sees. Publish worked examples inside the methodology, including the uncomfortable ones, the simple high-stakes model, the complex low-stakes model, so scorers cannot claim ambiguity. Wire the triggers of paragraph 17 to the monitoring outputs of paragraph 16, so a model that starts touching more customers or drifting in performance re-enters tiering without a human remembering to send it. My book on decision centric governance, builds this as the graduated autonomy ladder, artifact A16, where authority is earned in degrees against demonstrated performance and the tier is simply the rung the evidence currently supports. Later episodes return to that ladder when the draft reaches AI autonomy, because the RBI has, perhaps without saying so, drafted the regulatory version of it.

The Examiner’s Question

List every model sitting exactly one point below your high-tier boundary, then pull their scoring sheets and check who benefits from the placement. If the cluster is dense and the rationales are thin, your tiers were priced, not assessed.

Commentary on a draft under consultation, not legal advice. Paragraph numbers may change at finalisation.