Skip to main content

Simulation run on this device. Scenario values are not sent anywhere — the engine is fully local.

research game · free

Reproduction Sprint

Sequence environment, data, training, and eval against the clock.

Seven days to reproduce the headline number — plan it.

Engine 1.0.0 · game
7
Days
7
Paper
not named
Failure modes
3
Sprint plan
DayFocusTask
Day 1Environment + dataPin versions, download dataset, reproduce the reported split counts
Day 2Baseline firstRun the paper's simplest baseline; record metric to compare later
Day 3Core methodImplement the headline component; unit-test one forward pass
Day 4Train / scale upLaunch a small run, check the loss curve against the paper's figure
Day 5Full evalRe-run the eval protocol; capture variance across seeds
Day 6Diagnose gapsCompare deltas to the paper; isolate the first divergence
Day 7Write it upRepro notes, what matched, what did not, and the next experiment
  • medium
    Silent data mismatch

    the split or preprocessing differs from the paper → checksum + count the exact files on day 1

  • medium
    Undocumented hyperparameters

    a missing LR/optimizer detail moves the result → ablate the missing knob before scaling

  • medium
    Eval protocol drift

    metric definition differs (micro vs macro, decoding) → re-read the eval section before comparing numbers

One week is enough to reach a defensible first comparison, not a full replication.

Method

  • Distributes a seven-stage reproduction plan (environment → baseline → method → train → eval → diagnose → write-up) across the days available.
  • Ships the three failure modes that kill most repro attempts, each with a pre-emptive move.