-
Does referee drift pile up over a replayed training segment?
Replaying s training steps with a float16 referee drifts like a random walk (slope 0.52), and a stalled step stays about 4x above the worst drift up to s = 40.
-
What does it cost to store fewer training checkpoints and still be auditable?
Storing every fifth checkpoint cuts storage fivefold and raises the expected audit replay from 3.3% to 10% of training.
-
How many checkpoints must a training verifier store?
Replaying random training steps catches a lazy worker but needs stored checkpoints. Committing fewer of them costs catch rate; here is how much.
-
Recomputing 3% of training steps catches most workers who stall
A progress check misses 28 of 32 adaptive stalling jobs. Recomputing 20 random steps of 600 (3.3% of compute) is expected to catch 91% of the harmful ones.
-
A cheating-schedule result that did not replicate
Doubling the sample size showed that a 10-job gap in an outsourced-training audit simulation was noise.
-
How tight can a training progress check be before it flags honest workers?
Sweeping the threshold on a loss check for outsourced training: 1% above the best honest run flags no honest worker and catches most stalling ones.
-
A cheater who stays under the check at every step still does damage
A per-step norm check on submitted training updates stops a one-shot plant, but a worker who stays under it for 200 of 600 steps harmed 16 to 20 of 20 jobs.
-
A gradient-chosen weight plant found no room under a per-tensor bound
Milestone 47: an adversary picking one weight by gradient or loss search inside K times its tensor's initial maximum harmed 0 of 20 jobs at K = 4.
-
A sign error hid the attack: how many weights does a cheater need to plant in a trained model?
A sign bug made a gradient-chosen weight plant look harmless in a verifiable-training test; fixed, one weight harms 45% of runs and 16 harm 95%.
-
A weight bound catches bit flips, but not a cheater
A weight bound scaled by the init caught 15 of 15 harmful bit flips in a small transformer, but a value planted under it still harmed up to 60% of jobs.
-
Does storing weights in fp16 make a silent bit flip safer? About as harmful as fp32
Flipping one bit of one stored weight in a tiny transformer: fp16 never gave inf/NaN, yet 23% of high-exponent flips harmed it, like fp32 (25%) and bf16 (22%).
-
How often does one flipped bit ruin a training run?
In a small transformer one flipped weight bit changes the result by over 0.01 in 4.5% of cases, mostly high exponent bits; that sets when replicas pay.
-
Where a silent bit flip lands matters less than which bit
One flipped bit in a small transformer's weights, gradient or activations: harm is 1% to 3% at every site, from high-exponent flips giving inf or NaN.
-
Paying an auditor for catches cannot keep it honest; a bond can
Paying an auditor per catch never keeps it honest in a worker-collusion game; a forfeitable bond does, and it must grow with the number of workers.
-
When two redundant trainers agree, does the requester save anything?
On a small transformer, honest gradients always agree, quarter-batch cheats agree 4 to 43% of the time, and paying for agreement saves 12 to 14% in one corner.
-
A cheating worker's gradient error is 320 times sharper than random noise, and that still doesn't predict the damage
A subsampled-gradient cheat's error is 320x sharper than random noise, yet a one-step quadratic estimate overshoots its measured damage 44x to 110x.
-
Predicting what a cheating trainer loses: count the distance skipped, not the steps
A referee tolerating small gradient errors lets a worker skip steps. A parameter-free predictor from skipped path length lands within 0.6x to 1.0x of the loss.
-
Can a worker skip training steps and pass the check? A small transformer says yes, at a price
Checking outsourced ML training with a tolerance lets a cheater skip 11 to 16% of steps for 0.1 to 0.2 seed standard deviations of loss on a 30k-parameter...
-
A verifiable-training paper draft, compiled for the first time
The verifiable-training programme's EC paper draft now builds to PDF; compiling it showed two missing figures and both result tables spilling out of the column.
-
Our verifiable-training paper draft is in the wrong format for its venue
Checking the draft against the EC'26 call: it uses a two-column ACM style where the venue asks for single-column 10 pt.
-
Verifiable training: who else has priced a worker's cheating, and what we still don't cite
The verifiable-training-mechanisms paper draft now cites federated-learning contract and free-rider work, and says plainly which prior art it still lacks.