Two strengths of AI, one trusted platform
How MetTutor blends generative language intelligence with exact deterministic computation — so every explanation teaches and every number is right.
Language models explain beautifully. They don’t compute reliably.
Large language models generate each number from patterns learned during training — not from a validated algorithm. The same prompt can return different results on different runs, and even plausible-looking answers can carry errors invisible at a glance.
In metrology, that matters more than in most fields. A measurement result is a value with a stated uncertainty, traceable to the SI through an unbroken chain of calibrations. Every number in that chain must be reproducible. A result that is right most of the time is not a measurement result.
NIST’s Generative AI Profile identifies “confabulation” — the production of confidently stated but erroneous content — as a core risk of generative models. A wrong uncertainty presented with a plausible derivation is the worst case in measurement science.
Research has shown that language models achieve only moderate accuracy on multi-digit arithmetic. When errors compound through chained calculations — an uncertainty budget with a dozen contributors, squares, a square root, a Welch-Satterthwaite ratio and a table lookup — the probability that every step is correct drops rapidly.
MetTutor’s answer: keep both AI strengths, but assign each to the role it was built for.
Two complementary strengths, each doing what it does best
Rather than choosing between generative AI and rule-based computation, MetTutor uses both — with a clear boundary between them. Code computes every reported value. The language model explains the finished results.
Generative LLM
Explains, teaches, guides a technician through a procedure, and answers the question behind the question — in natural language, grounded in MetLibrary’s foundational standards and principles.
- Explains why νeff is truncated rather than rounded
- Describes which contributor dominates a budget and why
- Walks through decision rules in plain language
- Adapts explanations to the learner’s level
- Connects the calculation to its standard-based rationale
Deterministic Computation
Computes every reported number from fixed, validated algorithms using exact rational arithmetic — the same input gives the same output, every time, on every device.
- Exact rational arithmetic: 0.1 + 0.2 = exactly 0.3
- Same card on every run and every device
- Validated case-by-case against independent references
- Refuses ambiguous input rather than estimating an answer
- Every export shows the same strings as the card
What each approach contributes
The distinction is not theoretical. Each property below has practical consequences for a technician, a laboratory, or an assessor.
| Property | Deterministic Computation | Generative Language Model |
|---|---|---|
| Same input, same output | Always — by design | Not guaranteed, even at the most repeatable settings |
| How a number is produced | Computed by an explicit, validated algorithm | Generated token by token from learned text patterns |
| Error behavior | Systematic and reproducible; found by testing | Confabulation: fluent, plausible, and wrong |
| Can be validated case by case | Yes — one correct output per input | Only by sampling; a pass does not carry to the next run |
| Explains its result | Only what it was programmed to state | Fluently, in natural language |
| Role in MetTutor | Compute every reported number | Explain, teach, and guide |
From input to explanation — how a request flows
Every calculator request passes through five stages. The first four are deterministic code; only the final stage uses the language model — and it receives finished, verified numbers it cannot alter.
Guided Form
Fixed fields compose every request into one consistent structure. Nothing depends on phrasing.
Parser
Code reads each field as an exact quantity. Anything ambiguous is refused with a clear fix — never guessed.
Calculation
Exact rational arithmetic computes the result. Values are held as integer fractions, rounded once at display.
Presentation
One function produces every display string. Card, PDF, Word and spreadsheet cannot disagree.
Explanation
The model receives verified results and explains them. It quotes the computed numbers — never recomputes.
Exact arithmetic where it matters most
Standard floating-point arithmetic introduces tiny rounding errors that can change a displayed digit, push a value across a threshold, or alter a coverage factor. MetTutor’s calculators use exact rational arithmetic wherever a result can be exact, and confine floating-point computation to the handful of values that are inherently irrational.
Rational Numbers
Values are held as exact fractions of integers built from the digits the user types. Ratios, sums, unit conversions and the Welch-Satterthwaite effective degrees of freedom are computed exactly — so 0.1 + 0.2 is 0.3, not 0.30000000000000004.
Exact Square Roots
A displayed root — a standard uncertainty, uc, or U — is decided by comparing squares of integers, so the last displayed digit is never off by one through floating-point error.
Exact Truncation
νeff is truncated by integer division on an exact fraction, so a value that is exactly 4 can never be computed as 3.9999999999999996 and truncated to 3 — a silent error that would inflate U by 14%.
Doubles Only Where Required
Only inherently irrational quantities — π-based unit factors, normal-distribution risk integrals and exponentials — use IEEE-754 double precision. Each is validated against independent high-precision references.
Conservative rounding — a shown value never overstates
Each display rule is chosen so that no number on the screen claims more than the true value supports. A reported uncertainty is never smaller than the one computed. A ratio never appears to reach a threshold it doesn’t reach. An acceptance limit is never wider than the one computed.
| Quantity | Rule & Rationale | |
|---|---|---|
| TUR, TAR, ratios | Truncated to 2 decimals — a ratio never appears to reach a threshold (1:1, 2:1, 4:1) it does not reach | |
| Expanded uncertainty U | Rounded up to 2 significant figures — the reported uncertainty is never smaller than the one computed (consistent with GUM 7.2.6) | |
| uc, u, ui | Rounded up to 3 significant figures — one guard digit finer than U | |
| νeff | Truncated to an integer, as GUM G.4.1 permits; the raw value is truncated to 1 decimal so it never reads above the next integer | |
| PFA (false acceptance risk) | Rounded up to 2 significant figures — the risk shown is never smaller than the risk computed | |
| Guard bands, acceptance limits | Guard band rounded up, acceptance limit rounded inward — the displayed acceptance zone is never wider than the computed one | |
Eight validated calculators, covering core calibration mathematics
Each calculator is a fixed, validated algorithm that handles one core calculation a calibration laboratory or metrology student needs. Each follows the same five-stage architecture, uses the same exact arithmetic, and has been independently validated.
TUR Calculator (Simplified)
Computes the test uncertainty ratio from the tolerance and expanded uncertainty, with Z540.3 Method 6 guard bands when below 4:1.
TUR Calculator (Full)
Goes beyond the ratio to measurement decision risk: PFA and PFR for four guard-band options, judged against the Z540.3 2% requirement.
TAR Calculator
Computes the test accuracy ratio between a device’s tolerance and the standard’s accuracy, placed against the legacy 4:1 rule.
SI Unit Converter
Converts between units of the same quantity across 24 unit domains using exact factors from their defining constants.
Type B Uncertainty Estimator
Turns a bound — half-width, full span, resolution or specification — into a standard uncertainty using the chosen probability distribution.
GUM Uncertainty Budget
Builds a complete budget: sensitivity coefficients, combined standard uncertainty, effective degrees of freedom, coverage factor and expanded uncertainty.
Welch-Satterthwaite
Combines uncertainty contributions, computes νeff by the Welch-Satterthwaite formula, looks up k from Table G.2, and reports U with dominant contributors.
Decision Rules & Guard Band
Evaluates four decision rules side by side — simple acceptance, Method 6, ISO 14253-1 and stringent — with conformity statements for each.
68,000 independent test cases. Zero mismatches.
Every calculator was validated against independent reference implementations written from the standards — not from the production code — using exact rational arithmetic. Random differential testing, published known answers, edge cases, invariant checks and live-application tests were applied across all eight calculators.
Independent References
Each reference implementation was written from the standards, in exact rational arithmetic, sharing no code with production. A disagreement between production and a reference is investigated on both sides.
Before/After Comparison
Every earlier validation input — 69,667 in total — was re-run before and after each change and compared field by field. Only values the change was designed to move had moved.
Published Known Answers
Results from GUM H.1, UKAS M3003, JCGM GUM-5, the Z540.3 Handbook and NIST SP 811 were reproduced to their printed precision — confirming alignment with the standards themselves.
Why metrology’s standards require deterministic computation
Every requirement that gives a measurement result its authority assumes a calculation that can be repeated exactly. Metrology’s governing documents do not use the word “deterministic,” but each of them depends on it.
Traceability Is a Chain of Calculations
Each link in a traceability chain carries a computed value and a computed uncertainty. A link whose arithmetic cannot be reproduced is not documented — and the chain is broken at that point.
Software Must Be Validated
ISO/IEC 17025:2017 requires software used for calculating results to be validated before use and after changes — which is only meaningful if the output for a given input is fixed.
Uncertainty Is a Defined Procedure
The GUM specifies how uncertainties are combined and coverage factors chosen. Once choices are documented, two competent evaluators of the same budget must obtain the same result.
Conformity Follows an Agreed Rule
ISO/IEC 17025:2017 requires the decision rule to be agreed and stated. Applying a stated rule to stated values must give one answer — a decision that could differ on a second run is not the result of a rule.
Results Must Be Reviewable
An assessor must be able to trace every reported value from the input to the result through a documented algorithm. Deterministic computation is what makes that audit trail complete.
Aligned with AI Governance Frameworks
The NIST AI Risk Management Framework places “valid and reliable” at the base of trustworthy AI. The OECD AI Principles call for traceability of processes and decisions. Deterministic computation meets both.
Grounded in MetLibrary’s foundational standards
Every calculation, every explanation, and every teaching response in MetTutor is grounded in MetLibrary — a curated knowledge base of foundational metrology standards, recommended practices and reference guidance.
The deterministic calculators implement the procedures defined in these standards. The language model’s explanations draw on MetLibrary’s content to ensure that the reasoning behind each result is consistent with the standards it applies.
This means a student learning about uncertainty evaluation, a technician checking a TUR, and a manager reviewing a decision rule all receive guidance aligned with the same recognized international standards — explained in natural language and backed by exact computation.
MetTutor’s dedicated MetLibrary manages and controls this information to ensure consistent, accurate guidance aligned with recognized international standards, minimizing the presence of erroneous data.
JCGM 100:2008 (GUM)
Guide to the expression of uncertainty in measurement
ISO/IEC 17025:2017
General requirements for calibration and testing laboratories
ANSI/NCSL Z540.3-2006
Requirements for calibration of measuring and test equipment
ILAC-G8:09/2019
Guidelines on decision rules and statements of conformity
NIST SP 811 & SP 330
Guide for the use of the SI and the International System of Units
NIST AI 100-1 & AI 600-1
AI Risk Management Framework and Generative AI Profile
Numbers you can trust. Explanations you can learn from.
See MetTutor’s deterministic calculators in action — validated, exact and grounded in the standards that matter.
Try MetTutor View Plans →NCSL International (NCSLI), the American Association for Laboratory Accreditation (A2LA), the American Society for Quality (ASQ), the American National Standards Institute (ANSI), the International Organization for Standardization (ISO), the International Electrotechnical Commission (IEC), the Joint Committee for Guides in Metrology (JCGM), the National Institute of Standards and Technology (NIST), and all other organization names, product names, and trademarks referenced on this page are the property of their respective owners. MetTutor is not affiliated with, endorsed by, or sponsored by any of these organizations.
