← All evidence

Three-way email classification: Jev prompt variants vs TF-IDF logistic regression

AIndependent · Reproduction artifacts availableSecurity detectionMain paired set (author's own assembly) · n = 5,733date not stated by source

Label an email as legitimate, spam or phishing.

Accuracy
93.6%
Jev jev-1.13.0, text-only
no interval reported
98.0%
Jev jev-1.13.0, enriched evidence
no interval reported
98.6%
Jev jev-1.13.0, enriched + evidence wording
no interval reported
98.9%
TF-IDF + logistic regression enriched
no interval reported
supervised — trained on this dataset, not zero-shot
99.3%
Jev + TF-IDF ensemble
no interval reported
Every recorded metric
ModelAccuracyRecallFPRAUROCECELatencyCost
Jev
93.6%85.7%not recorded
Jev
98.0%98.4%not recorded
Jev
98.6%98.4%not recorded
TF-IDF + logistic regression
98.9%75.3%not recorded
Jev + TF-IDF ensemble
99.3%95.7%not recorded
Protocol

Three-way classification. Jev accessed as jev-1.13.0 in three prompt configurations: text-only, enriched with extracted evidence, and enriched with evidence-focused wording. The TF-IDF logistic regression baseline used fixed, untuned settings. Five-fold cross-validation with near-duplicate grouping. The full paired run issued 19,772 requests in about 252 seconds at concurrency 16, consuming 32.06M input tokens.

Dataset
Main paired set (author's own assembly)
n = 5,733
1,911 legitimate / 1,911 spam / 1,911 phishing
Limitations (7)
  • The source does not state a run date, so these figures cannot be placed against a specific model version or price period.
  • No confidence intervals are reported at any sample size.
  • The author is explicit that 'no task-specific fitting is not no supervision': the detailed binary criteria used in the enriched prompts were informed by labelled errors, so the enriched configurations are not cleanly zero-shot.
  • The author notes pretraining exposure to these messages is unknown and the labels are confounded, and describes the set as 'not a locked benchmark'.
  • Calibration was not measured. The author notes that probability-shaped outputs still require calibration checks.
  • The repository does not state the author's affiliation or who funded API costs.
  • No per-call cost is recorded. The source reports about $1.35 of measured spend, but that total covers the full paired run across every prompt configuration and every set, so it cannot be attributed to a single configuration on this dataset.
Note on this run
The most consequential result recorded here. Jev's phishing recall moves from 85.7% to 98.4% purely by changing the prompt configuration on the same data — which is why a recall figure quoted without its protocol means very little.

What this benchmark does not prove

  • 01That Jev has a phishing recall of 98.4%. That is the enriched configuration on this dataset; the text-only configuration on the same data scored 85.7%, and a different dataset scored lower still.
  • 02That prompt enrichment is free. The author states the enriched criteria were informed by labelled errors, so part of the gain is supervision that a genuinely cold-start deployment would not have.
  • 03That the TF-IDF baseline is the better choice because it has the highest accuracy. It has the worst phishing recall of the serious configurations, and on a detection workload that is usually the number that decides it.
  • 04That the ensemble's 99.3% is production performance. It is an unvalidated combination on a set the author describes as not locked, with no intervals reported.
Source and artifacts

jev-spam-eval — three-way email classification: Jev variants vs TF-IDF logistic regression bitnovus, verified 2026-09-19

Reproduction artifacts: github.com/bitnovus/jev-spam-eval

Independent: yes · Vendor-affiliated: no · Independently reproduced: not yet

Public code and stated protocol. The repository does not state a run date, the author's affiliation, or who funded API costs; it does report measured spend of about $1.35 for the full paired run.