The Framework
Bernan's Law Rubric v1
Put a deterministic gate on a guessing machine.
A universal, modular standard for grading any AI build. It does not care whether the build estimates a job, drafts an email, or runs outreach. It asks the same structural question every time: how much of the work is gated, and how much is still a confident guess.
Locked v1, 2026-06-23. Cite it by version: Bernan's Law Rubric v1.
Why It Exists
It grades the build, not the subject.
A large language model is a probabilistic engine. It will produce a plausible answer to anything, including the things it has no business answering. A build earns its keep by the gates it wraps around that engine: the rules, lookups, formulas, schemas, thresholds, and verify checks that turn a confident guess into a defensible result.
Those properties are subject-blind. The same nine dimensions score a construction estimator and an email skill identically, because the judgment lives in the evidence, not in the grader.
The Nine Dimensions
Weighted toward determinism, on purpose.
Weights sum to 100. The determinism family, the first three, carries 44 of those points, because that is where Bernan's Law lives.
Determinism Coverage
18Determinism family
What fraction of the build's consequential decisions are resolved by a deterministic gate, a rule, a lookup, a formula, a threshold, versus left to free model judgment.
Run-to-Run Reproducibility
14Determinism family
Whether the same input yields the same answer across runs, inside a declared tolerance, proven by actual re-runs rather than inferred from the architecture.
Gate Correctness and Validity
12Determinism family
Whether the gates encode the right rules and hold up against ground truth. A gate that fires the wrong answer is worse than no gate.
Failure-Mode Behavior
14What the build does when an input is missing or wrong. It either halts and surfaces the problem, or it silently guesses. Only one of those is safe.
Self-Verification Depth
12Whether the build checks its own work against the real result before it declares done, using a different primitive than the one that produced the output.
Source Grounding and Truthfulness
10Whether consequential claims trace to a named, re-fetchable source, and whether a missing source fails closed instead of inventing a plausible answer.
Blast Radius and Reversibility
8How far a wrong answer can travel before a human catches it, and whether irreversible actions are proposal-gated and isolated from live state.
Variable Coverage
6How many of the variables that actually move the outcome the build ingests and weighs, versus the ones it silently leaves out.
Auditability and Traceability
6Whether a stranger can reconstruct why the output is what it is, from a durable trail, without having to re-run the build.
The Tiers
From guessing machine to reference build.
The weighted score maps to one of five tiers.
Deterministic-Grade
The gate is dense, stable, correct against ground truth, fails closed, verifies the real result, and is fully reconstructable. A reference build.
Operator-Grade
A trustworthy working build. It gates the load-bearing decisions, holds still across runs, fails loud, and is auditable, with named gaps an operator can live with.
Assisted-Grade
Useful with a human in the loop. Real gates exist, but consequential decisions still float or correctness is unproven. Do not run it unattended on high-stakes work.
Draft-Grade
A scaffold. The shape of a gated build is present, but most of the consequential surface is still free judgment. Fine for low-blast drafting.
Guessing-Grade
A guessing machine in a gate costume. It fails the central test of Bernan's Law.
The hard floor
One ungated irreversible action, a silent auto-send, an unverified live write, caps the whole build at Assisted-Grade. No score can buy its way past it.
The Method
Grade any build in five steps.
- 1
List the consequential decisions the build makes and every point where it writes, sends, or commits. That declared manifest is what graders score over, so two graders count the same things.
- 2
Tag each decision as gated or guessed, and tag how much a wrong answer costs: cosmetic, low, high, or irreversible.
- 3
Score each of the nine dimensions 0 to 100 against its anchors, and cite the evidence for every score. A claim without evidence is an opinion, not a grade.
- 4
Weighted-sum the nine scores to a single 0 to 100 build score.
- 5
Map the score to a tier, apply the irreversible-defect floor, and publish the per-dimension evidence so a second grader can replay the number.
Want to find the operational leak worth grading first?
Run the free operator scan. Three taps identify the bottleneck most likely to justify a build. The rubric above is the separate standard used to grade that build once it exists.
Find the first workflow →Bernan's Law Rubric v1, locked 2026-06-23. A standard you can cite, dispute, and grade against.