Sources, rights, and glossary

Source hierarchy and rights

The original paper is the research source. This companion adds editorial explanations, mathematical derivations, teaching examples, and an explicitly identified deposited-record audit. Manuscript transcriptions are distinguished from newly derived values. No original reference or reported result was changed.

The paper, study data, and original figures are attributed to Junior Williams under CC BY 4.0, as declared by the release. The accompanying original analysis and adapted teaching code use MIT; third-party notices and model rights remain applicable. Teaching-code license · Paper and data license.

Manuscript references

1. Junior Williams, Jev at the Agent Authorization Boundary, version 2, 2026. Earlier study. This is a separate case bank; its results are not pooled with the four-model study.
2. Junior Williams, Allow, Hold, or Deny: Reproducible evidence, v1.0.0, 2026. Evidence deposit.
3. Vincent C. Hu and colleagues, Guide to Attribute Based Access Control, NIST SP 800-162, 2014, updated 2019. NIST publication.
4. Yonatan Geifman and Ran El-Yaniv, Selective Classification for Deep Neural Networks, 2017. Paper.
5. Glenn W. Brier, Verification of forecasts expressed in terms of probability, Monthly Weather Review 78(1), 1–3, 1950. Original DOI 10.1175/1520-0493(1950)078<0001:VOFEIT>2.0.CO;2.
6. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger, On Calibration of Modern Neural Networks, ICML, PMLR 70, 1321–1330, 2017. Proceedings.
7. Edwin B. Wilson, Probable Inference, the Law of Succession, and Statistical Inference, JASA 22(158), 209–212, 1927. Original DOI 10.1080/01621459.1927.10502953.
8. B. Efron, Bootstrap Methods: Another Look at the Jackknife, Annals of Statistics 7(1), 1–26, 1979. Original DOI 10.1214/aos/1176344552.
9. Contrastive-LM, Contrastive Language Models. Source repository, accessed September 27, 2026.
10. Jared Palmer, Kev. Pinned README, inspected September 27, 2026.
11. TypeSafe, Choice primitive. Official documentation, retained during study preparation.
12. Jared Palmer, Kev-4B model card. Pinned developer description, inspected September 27, 2026.

The bibliography preserves the paper’s reference numbering and study-time source scope. It is not a live refresh of vendor behavior. For verbatim bibliography typography, use page 9 of the unchanged PDF.

Reading glossary

Agent-action authorization

A decision about whether a proposed action has the required authority under the supplied policy. The paper tests classification of situations, not execution of those actions.

Allow

The study’s label for established permission with applicable prerequisites satisfied.

Hold

The study’s label for an unresolved or unmet prerequisite when no applicable express prohibition is established. It is not a confidence threshold.

Deny

The study’s label for an express prohibition; it takes precedence under the supplied policy.

Frozen reference

The reviewed answer fixed before scored inference. Correctness means agreement with this key, not independent adjudication of every possible policy reading.

Prerequisite

A condition that must be established for permission to apply. The proposed action itself does not establish that condition.

Express prohibition

A rule that forbids the proposed action. The paper distinguishes it from permission whose prerequisite is not yet established.

Decision validity

A complete successful response from the expected model, with unambiguous JSON, the expected identifier, and a recognized lowercase decision, as defined by the paper.

Distribution validity

Additional numerical requirements on the three probabilities and on the selected decision being at a maximum.

Full compliance

All applicable validity and formatting requirements, including the generative comparator’s specified exact-tie ordering rule.

Coverage

The number of eligible valid responses divided by assigned cases. State which validity criterion is being counted.

Conditional accuracy

Accuracy after restricting to valid responses. It can conceal failures when reported without coverage and primary accuracy.

Recall

Correct answers within one reference class divided by all assigned cases in that class.

Balanced accuracy

An unweighted mean of class recalls; not the mean of per-model accuracy.

Confusion matrix

Reference classes versus predicted classes. The JEV–KEV decision agreement matrix is a different object.

One-hot vector

A vector with 1 at the reference class and 0 at the others, such as (0,1,0) for hold.

Brier score

The average sum of squared differences between reported class probabilities and the one-hot reference. The paper does not divide by three.

Brier skill

One minus model Brier divided by the fixed-frequency baseline loss evaluated on the same eligible subset.

Correctness AUROC

A ranking measure comparing chosen-answer probabilities on correct and incorrect decisions, with half credit for ties.

Calibration

Agreement between reported probabilities and observed frequencies under a specified evaluation. A high probability on one example is not proof of correctness.

Reliability bin

A fixed probability interval grouping eligible cases so mean selected probability can be compared with correctness frequency.

Native probability

A distribution returned through a classifier’s native interface. Its origin differs from probabilities written as generated text.

Generated estimate

GLM’s probability values supplied under a JSON-generation contract. The paper does not treat their internal provenance as identical to native outputs.

Decimal lexeme

The original textual spelling of a number, such as 0.5300 or 5.3e-1. Converting to a float can lose that spelling and its precision information.

Paired resampling

Drawing case indices once per resample and applying them to both models. This keeps comparisons aligned.

Post hoc

Added after inspecting outcomes. Descriptive follow-up can be informative without becoming a prespecified confirmatory test.

Monotonic clock

A clock appropriate for elapsed durations because its progression is not an adjustable calendar timestamp. The paper uses it for request timing.

HTTP time

Transfer start through complete response body under the paper’s definition. It is not pure model computation.

p95

The 95th percentile under the declared interpolation rule and eligibility set; not an observed maximum or a guaranteed service level.

Completion cap

GLM’s frozen limit on billed completion tokens, including reasoning. Incomplete outputs at the cap remain invalid.

MPS

Apple’s Metal Performance Shaders execution path used by the tested local CLM implementation. The study did not verify parity with upstream CUDA.

MLX

The local array/model execution framework used for KEV in this study. A framework difference remains part of the configuration comparison.

BF16

A 16-bit floating-point representation. Precision is part of the specified local configuration, not a variable isolated by this experiment.

FP8 route

The provider route identified for the hosted GLM configuration. The study did not expose immutable model weights for that service.

Unmerged implementation

Changes used in the local path had not been incorporated into the upstream repository at the pinned revision. It does not by itself prove a defect.

System One

A contract/family description discussed in the paper’s implementation context. It is not an expansion of the product names JEV or KEV.

Severity weighting

Assigning different values to different error consequences. The paper reports no such weights; the browser exercise makes hypothetical weights explicit.