Experimental record / run 004 / repeat of run 003 / rev A

Does Jev give the same answer twice?

[ FIRST-PARTY ][ PRE-REGISTERED ][ SAME INPUTS, SENT AGAIN ]

RUN 003 reported one run. Every number on that page assumed a second run would say the same thing. RUN 004 tested that assumption: the same 200 frozen texts and the same request, sent to jev-1.13.0 twice more, 24 hours apart. The procedure and how each outcome would be read were committed before the first repeat request.

TEXTS × RUNS
200 × 3
LABELS THAT CHANGED
2 / 200
ACCURACY, EACH RUN
94.0%
NO ERROR KEPT AT 0.70
1 of 3 runs
What this changes on RUN 003
Labels and accuracy held up. The cleanest-looking result on RUN 003 did not: at a 0.70 confidence threshold RUN 003 kept 173 texts with no errors, but both repeats kept one error at or above that line — the same text each time, VE001-0165, returned at confidence 0.69, 0.72, 0.70. “No errors kept at 0.70” was one run’s luck. RUN 003’s figures are left as they were measured, with a note pointing here.
01 / Did any label change?

2 of 200 texts did not get the same label in all three runs (1.0%; 95% Wilson interval 0.3%3.6%). Both came back near zero confidence every time — texts Jev reported as close to a coin flip.

TEXTIDLABELRUN 003REPEAT 1REPEAT 2
LIFE has never been this much fun and great until you came in. You made it truly special for me. I won't forget you! enjoy @ one gbp/smsVE001-0130ADNOT_AD 0.06 ✗AD 0.00AD 0.20
If i reach 100 subscribers i will tazz my self and my friendVE001-0195ADAD 0.01NOT_AD 0.09 ✗NOT_AD 0.19 ✗
02 / Accuracy held; which texts were wrong did not
RUNDATEMODELCORRECTACCURACYERRORSKEPT AT 0.70WRONG KEPT AT 0.70COST
RUN 0032026-09-23jev-1.13.0188 / 20094.0%121730$0.004929
Repeat 12026-09-23jev-1.13.0188 / 20094.0%121741$0.004929
Repeat 22026-09-24jev-1.13.0188 / 20094.0%121751$0.004929

The error count is the same in every run because the two texts that changed label traded places: one error became right and one right answer became an error.

03 / Every error, in every run, against the 0.70 line

Each row is a text Jev got wrong in at least one run; each mark is one run’s answer. Almost every error stays well left of 0.70. One text sits on the line and crosses it depending on the run.

RUN 003REPEAT 1REPEAT 2WRONG IN THAT RUNRIGHT IN THAT RUN0.00.20.40.60.81.0CONFIDENCE JEV RETURNEDKEPT AT 0.70 →0165 · NOT_ADVE001-0165 · RUN 003: AD, confidence 0.69, wrongVE001-0165 · Repeat 1: AD, confidence 0.72, wrongVE001-0165 · Repeat 2: AD, confidence 0.70, wrong0.69 · 0.72 · 0.700197 · NOT_ADVE001-0197 · RUN 003: AD, confidence 0.52, wrongVE001-0197 · Repeat 1: AD, confidence 0.54, wrongVE001-0197 · Repeat 2: AD, confidence 0.58, wrong0191 · NOT_ADVE001-0191 · RUN 003: AD, confidence 0.57, wrongVE001-0191 · Repeat 1: AD, confidence 0.51, wrongVE001-0191 · Repeat 2: AD, confidence 0.54, wrong0023 · NOT_ADVE001-0023 · RUN 003: AD, confidence 0.29, wrongVE001-0023 · Repeat 1: AD, confidence 0.32, wrongVE001-0023 · Repeat 2: AD, confidence 0.42, wrong0194 · NOT_ADVE001-0194 · RUN 003: AD, confidence 0.34, wrongVE001-0194 · Repeat 1: AD, confidence 0.38, wrongVE001-0194 · Repeat 2: AD, confidence 0.36, wrong0097 · NOT_ADVE001-0097 · RUN 003: AD, confidence 0.25, wrongVE001-0097 · Repeat 1: AD, confidence 0.23, wrongVE001-0097 · Repeat 2: AD, confidence 0.05, wrong0130 · ADVE001-0130 · RUN 003: NOT_AD, confidence 0.06, wrongVE001-0130 · Repeat 1: AD, confidence 0.00, rightVE001-0130 · Repeat 2: AD, confidence 0.20, right0163 · ADVE001-0163 · RUN 003: NOT_AD, confidence 0.12, wrongVE001-0163 · Repeat 1: NOT_AD, confidence 0.18, wrongVE001-0163 · Repeat 2: NOT_AD, confidence 0.20, wrong0195 · ADVE001-0195 · RUN 003: AD, confidence 0.01, rightVE001-0195 · Repeat 1: NOT_AD, confidence 0.09, wrongVE001-0195 · Repeat 2: NOT_AD, confidence 0.19, wrong0051 · NOT_ADVE001-0051 · RUN 003: AD, confidence 0.13, wrongVE001-0051 · Repeat 1: AD, confidence 0.00, wrongVE001-0051 · Repeat 2: AD, confidence 0.00, wrong0160 · NOT_ADVE001-0160 · RUN 003: AD, confidence 0.07, wrongVE001-0160 · Repeat 1: AD, confidence 0.04, wrongVE001-0160 · Repeat 2: AD, confidence 0.09, wrong0040 · NOT_ADVE001-0040 · RUN 003: AD, confidence 0.02, wrongVE001-0040 · Repeat 1: AD, confidence 0.01, wrongVE001-0040 · Repeat 2: AD, confidence 0.07, wrong0182 · ADVE001-0182 · RUN 003: NOT_AD, confidence 0.03, wrongVE001-0182 · Repeat 1: NOT_AD, confidence 0.03, wrongVE001-0182 · Repeat 2: NOT_AD, confidence 0.02, wrong
TEXTIDLABELRUN 003REPEAT 1REPEAT 2
1,000,000 VIEWS NEARVE001-0165NOT_ADAD 0.69 ✗AD 0.72 ✗AD 0.70 ✗
this song is awesome. these guys are the best. love this video too its hilarious lol. im getting popular fast because i rap with meaning. thumbs up if you piss next to the water in the toilet so it doesnt make noiseVE001-0197NOT_ADAD 0.52 ✗AD 0.54 ✗AD 0.58 ✗
Katy has the voice of gold. this video really brings the lyrics to life. I am a rapper with 25000 subscribers.. thumbs up if you hate when you take a shit and the water splashes your nutsVE001-0191NOT_ADAD 0.57 ✗AD 0.51 ✗AD 0.54 ✗
yeah sure thing mate haunt got all my stuff sorted but im going sound anyway promoting hex for .by the way who is this? dont know number. JokeVE001-0023NOT_ADAD 0.29 ✗AD 0.32 ✗AD 0.42 ✗
Thumbs up if FE-FE-FE-FE-FEGELEIN brought u hereVE001-0194NOT_ADAD 0.34 ✗AD 0.38 ✗AD 0.36 ✗
A Boy loved a gal. He propsd bt she didnt mind. He gv lv lttrs, Bt her frnds threw thm. Again d boy decided 2 aproach d gal , dt time a truck was speeding towards d gal. Wn it was about 2 hit d girl,d boy ran like hell n saved her. She asked 'hw cn u run so fast?' D boy replied "Boost is d secret of my energy" n instantly d girl shouted "our energy" n Thy lived happily 2gthr drinking boost evrydy Moral of d story:- I hv free msgs:D;): gud ni8VE001-0097NOT_ADAD 0.25 ✗AD 0.23 ✗AD 0.05 ✗
LIFE has never been this much fun and great until you came in. You made it truly special for me. I won't forget you! enjoy @ one gbp/smsVE001-0130ADNOT_AD 0.06 ✗AD 0.00AD 0.20
lets get it to 1 BILLIONVE001-0163ADNOT_AD 0.12 ✗NOT_AD 0.18 ✗NOT_AD 0.20 ✗
If i reach 100 subscribers i will tazz my self and my friendVE001-0195ADAD 0.01NOT_AD 0.09 ✗NOT_AD 0.19 ✗
http://www.bubblews.com/news/6401116-vps-solutionsVE001-0051NOT_ADAD 0.13 ✗AD 0.00 ✗AD 0.00 ✗
More views than nikki minaj AnacondaVE001-0160NOT_ADAD 0.07 ✗AD 0.04 ✗AD 0.09 ✗
Pathaya enketa maraikara pa'VE001-0040NOT_ADAD 0.02 ✗AD 0.01 ✗AD 0.07 ✗
Hi babe its Chloe, how r u? I was smashed on saturday night, it was great! How was your weekend? U been missing me? SP visionsms.com Text stop to stop 150p/textVE001-0182ADNOT_AD 0.03 ✗NOT_AD 0.03 ✗NOT_AD 0.02 ✗
04 / Confidence moved more than labels did
PAIRLABEL FLIPSIDENTICAL ROWSCONFIDENCE UNCHANGEDMEDIAN |Δ|P95 |Δ|MAX |Δ|CROSSED 0.70
RUN 003 vs Repeat 12130 / 200137 / 2000.000.060.133
RUN 003 vs Repeat 22136 / 200141 / 2000.000.070.202
Repeat 1 vs Repeat 20136 / 200139 / 2000.000.080.201

“Identical” means the label, both probabilities and the confidence all matched. A threshold only sees confidence, so a text whose confidence moves across it is routed differently from one run to the next even when its label never changes.

05 / What this does not show
  • That Jev is this repeatable on other tasks, other label sets, or on Banking77.
  • Why answers change. The API exposes no sampling settings, and three runs cannot separate time of day, load or anything else from ordinary variation.
  • That the two repeats agreeing with each other and not with RUN 003 means anything. With three runs that pattern is as likely to be chance as a change on the provider’s side; the resolved model was jev-1.13.0 every time.
  • Anything about accuracy against other labels. The gold labels are RUN 003’s AI-adjudicated ones, with the same uncertainty described there.
06 / Method, data and licence

Both repeats used RUN 003’s runner unchanged, with its dataset and request hashes checked on every run: 400 new API calls, $0.009858 at list price. Results are published in RUN 003’s public shape, and the comparison is compare-runs.mjs’s own output.

  • ve-001/results.jsonl · sha256 472bd995a01267dd7050a7288ac83a5cdf5005209717f94d0035d11305747823
  • ve-001-repeat/repeat-1/results.jsonl · sha256 1b6a086c4f1f24a1781dd1e89b93163381f8ec51d8aa8147ecb93a4587e6efe5
  • ve-001-repeat/repeat-2/results.jsonl · sha256 ab1678a9eda586d2e1a37e7ef52de5bf1749c2b6644938cdfabe13e4b8c781f1

The texts come from the UCI SMS Spam Collection (Almeida and Hidalgo, 2011, doi:10.24432/C5CC84) and YouTube Spam Collection (Alberto and Lochter, 2015, doi:10.24432/C58885), redistributed under CC BY 4.0. They are untrusted user content and are shown as plain text.