drift observatory

calibration receipts rendered from committed export data (gaige-receipt-export/1). every number below carries the instrument fingerprint that produced it and a reproduce command a stranger can run. measurements, never verdicts.

3 receipts.

fast-detect-gpt on hc3-mini(n=100,seed=17)

instrument: model tiiuae/falcon-7b · quant 4bit · device cuda · Linux/x86_64 · receipt 20260722-163959-fast-detect-gpt (2026-07-22T21:40:31+00:00, gaige 0.0.2) · corpus sha256 7d2819d3e83bd10d…

AUROC 0.9720 [0.9458–0.9938] (bootstrap, n_boot=1000)

target FPRthresholdachieved FPRTPRTPR 95% CI
1.0%2.12291.0%86.0%[79.0%–92.0%]
5.0%1.83195.0%91.0%[85.0%–96.0%]

conformal (distribution-free FPR bound):

alphathresholdTPRguarantee
0.051.846890.0%P(human flagged) <= 0.05, marginal over calibration draws, under exchangeability (split conformal)
0.012.444676.0%P(human flagged) <= 0.01, marginal over calibration draws, under exchangeability (split conformal)
0.005refused: alpha=0.005 needs >= 199 human calibration samples, got 100. A tighter guarantee than your data supports is not a guarantee.

at target FPR 1.0% (threshold 2.1229):

axis length_bucket

stratumn humanFPRFPR CIn AITPRTPR CI
0-100w42withheld: n below the honesty floor (floor 20/stratum)7withheld
100-250w390.0%[0.0%–0.0%]8185.2%[77.7%–92.6%]
250-500w19withheld: n below the honesty floor (floor 20/stratum)12withheld

max FPR disparity across length_bucket: 0.0% (all reported groups equal)

at target FPR 5.0% (threshold 1.8319):

axis length_bucket

stratumn humanFPRFPR CIn AITPRTPR CI
0-100w42withheld: n below the honesty floor (floor 20/stratum)7withheld
100-250w392.6%[0.0%–7.7%]8188.9%[81.5%–95.1%]
250-500w19withheld: n below the honesty floor (floor 20/stratum)12withheld

max FPR disparity across length_bucket: 2.2% (worst 0-100w 4.8% vs best 100-250w 2.6%)

reproduce:

gaige run --corpus hc3-mini --n 100 --seed 17 --detector fast-detect-gpt --model tiiuae/falcon-7b --quant 4bit --device cuda --max-tokens 1024

binoculars on hc3-mini(n=100,seed=17)

instrument: model tiiuae/falcon-7b + tiiuae/falcon-7b-instruct · quant 4bit · device cuda · Linux/x86_64 · receipt 20260722-213952-binoculars (2026-07-23T02:40:53+00:00, gaige 0.0.2) · corpus sha256 7d2819d3e83bd10d…

AUROC 0.9992 [0.9974–1.0000] (bootstrap, n_boot=1000)

target FPRthresholdachieved FPRTPRTPR 95% CI
1.0%-0.78291.0%97.0%[94.0%–100.0%]
5.0%-0.87065.0%100.0%[100.0%–100.0%]

conformal (distribution-free FPR bound):

alphathresholdTPRguarantee
0.05-0.8706100.0%P(human flagged) <= 0.05, marginal over calibration draws, under exchangeability (split conformal)
0.01-0.754095.0%P(human flagged) <= 0.01, marginal over calibration draws, under exchangeability (split conformal)
0.005refused: alpha=0.005 needs >= 199 human calibration samples, got 100. A tighter guarantee than your data supports is not a guarantee.

at target FPR 1.0% (threshold -0.7829):

axis length_bucket

stratumn humanFPRFPR CIn AITPRTPR CI
0-100w42withheld: n below the honesty floor (floor 20/stratum)7withheld
100-250w390.0%[0.0%–0.0%]8196.3%[92.6%–100.0%]
250-500w19withheld: n below the honesty floor (floor 20/stratum)12withheld

max FPR disparity across length_bucket: 2.4% (worst 0-100w 2.4% vs best 100-250w 0.0%)

at target FPR 5.0% (threshold -0.8706):

axis length_bucket

stratumn humanFPRFPR CIn AITPRTPR CI
0-100w42withheld: n below the honesty floor (floor 20/stratum)7withheld
100-250w392.6%[0.0%–7.7%]81100.0%[100.0%–100.0%]
250-500w19withheld: n below the honesty floor (floor 20/stratum)12withheld

max FPR disparity across length_bucket: 7.0% (worst 0-100w 9.5% vs best 100-250w 2.6%)

reproduce:

gaige run --corpus hc3-mini --n 100 --seed 17 --detector binoculars --observer tiiuae/falcon-7b --performer tiiuae/falcon-7b-instruct --quant 4bit --device cuda --max-tokens 1024

fast-detect-gpt on raid-g2d2a2-n60-s17

instrument: model EleutherAI/gpt-neo-1.3B · quant fp32 · device cpu · Linux/x86_64 · receipt 20260727-193225-fast-detect-gpt (2026-07-28T00:44:59+00:00, gaige 0.0.2) · corpus sha256 54b786694f2e4c0d…

AUROC 0.9330 [0.9117–0.9521] (bootstrap, n_boot=1000)

target FPRthresholdachieved FPRTPRTPR 95% CI
1.0%2.39040.8%58.1%[54.2%–62.7%]
5.0%1.49575.0%77.7%[74.2%–81.5%]

conformal (distribution-free FPR bound):

alphathresholdTPRguarantee
0.051.572776.5%P(human flagged) <= 0.05, marginal over calibration draws, under exchangeability (split conformal)
0.012.735549.4%P(human flagged) <= 0.01, marginal over calibration draws, under exchangeability (split conformal)
0.005refused: alpha=0.005 needs >= 199 human calibration samples, got 120. A tighter guarantee than your data supports is not a guarantee.

at target FPR 1.0% (threshold 2.3904):

axis attack

stratumn humanFPRFPR CIn AITPRTPR CI
none1200.8%[0.0%–2.5%]24059.2%[52.9%–65.0%]
paraphrase0withheld: n below the honesty floor (floor 20/stratum)240withheld

axis decoding

stratumn humanFPRFPR CIn AITPRTPR CI
greedy0withheld: n below the honesty floor (floor 20/stratum)218withheld
none120withheld: n below the honesty floor (floor 20/stratum)0withheld
sampling0withheld: n below the honesty floor (floor 20/stratum)262withheld

axis domain

stratumn humanFPRFPR CIn AITPRTPR CI
abstracts601.7%[0.0%–5.0%]24053.3%[46.7%–59.6%]
reddit600.0%[0.0%–0.0%]24062.9%[57.1%–68.3%]

axis generator

stratumn humanFPRFPR CIn AITPRTPR CI
gpt40withheld: n below the honesty floor (floor 20/stratum)240withheld
human120withheld: n below the honesty floor (floor 20/stratum)0withheld
mistral-chat0withheld: n below the honesty floor (floor 20/stratum)240withheld

axis length_bucket

stratumn humanFPRFPR CIn AITPRTPR CI
0-100w2withheld: n below the honesty floor (floor 20/stratum)83withheld
100-250w1100.9%[0.0%–2.7%]33061.8%[56.4%–67.0%]
250-500w8withheld: n below the honesty floor (floor 20/stratum)67withheld

max FPR disparity across domain: 1.7% (worst abstracts 1.7% vs best reddit 0.0%)

at target FPR 5.0% (threshold 1.4957):

axis attack

stratumn humanFPRFPR CIn AITPRTPR CI
none1205.0%[1.7%–9.2%]24069.6%[63.7%–75.4%]
paraphrase0withheld: n below the honesty floor (floor 20/stratum)240withheld

axis decoding

stratumn humanFPRFPR CIn AITPRTPR CI
greedy0withheld: n below the honesty floor (floor 20/stratum)218withheld
none120withheld: n below the honesty floor (floor 20/stratum)0withheld
sampling0withheld: n below the honesty floor (floor 20/stratum)262withheld

axis domain

stratumn humanFPRFPR CIn AITPRTPR CI
abstracts605.0%[0.0%–11.7%]24074.2%[68.8%–79.2%]
reddit605.0%[0.0%–11.7%]24081.2%[76.7%–86.2%]

axis generator

stratumn humanFPRFPR CIn AITPRTPR CI
gpt40withheld: n below the honesty floor (floor 20/stratum)240withheld
human120withheld: n below the honesty floor (floor 20/stratum)0withheld
mistral-chat0withheld: n below the honesty floor (floor 20/stratum)240withheld

axis length_bucket

stratumn humanFPRFPR CIn AITPRTPR CI
0-100w2withheld: n below the honesty floor (floor 20/stratum)83withheld
100-250w1104.5%[0.9%–8.2%]33079.7%[75.5%–83.9%]
250-500w8withheld: n below the honesty floor (floor 20/stratum)67withheld

max FPR disparity across domain: 0.0% (all reported groups equal)

reproduce:

gaige run --corpus corpora/raid-g2d2a2-n60-s17.jsonl --n 100 --seed 17 --detector fast-detect-gpt --model EleutherAI/gpt-neo-1.3B --quant fp32 --device cpu --max-tokens 1024

This receipt was produced from a locally prepared corpus file. The corpus sha256 above pins the exact bytes; the file itself is not redistributed. RAID slices (Dugan et al., ACL 2024) are prepared from the public dataset with `gaige corpus prepare-raid` and are never redistributed.

rendered from the committed export index; the page recomputes nothing and is rebuilt whole on every refresh.