Adam,
The recorded study generated 5,840,131,702 output tokens (5.840 billion) across 32,800,844 saved answers. This counts the actual saved native output-token IDs, including repetitive answers that are later excluded from clustering.
| Quantity | Verified total |
|---|---|
| Generated output tokens, all saved study records | 5,840,131,702 |
| Generated output tokens, primary analysis dataset | 5,827,948,453 |
| Generated output tokens, auxiliary runs | 12,183,249 |
| Distinct neuron interventions | 884,736 |
| Distinct layer–neuron coordinates | 294,912 |
| Prompts per intervention | 36 |
| Primary neuron-perturbed answers | 31,850,496 |
| Primary answers including controls and baselines | 32,735,268 |
| Additional auxiliary answers | 65,576 |
The neuron total is 32 layers × 9,216 neurons × 3 strengths (−0.7, +0.5, +0.8) = 884,736 distinct interventions. Each was tested on all 36 prompts. The original four-layer, 20-prompt run is included in these totals, not added a second time.
Including 4,096 auxiliary repeats, there are 31,854,592 saved generations with a neuron intervention. Those repeats introduce no additional layer–neuron–strength combinations.
An edit is counted by its recorded layer, neuron and strength. Each run retains its own hardware and calibration metadata; matching condition labels across cohorts do not imply identical numerical calibration.
Why this is not four billion edits: Qwen3.5-4B has approximately four billion learned scalar parameters. A neuron has many associated weights. We perturbed one MLP neuron's output direction at a time, across all layers; we did not change every individual weight, sweep attention components, or enumerate simultaneous neuron combinations. A separate three-modification sweep over roughly four billion individual weights would have about 12 billion variants, or 432 billion answers across 36 prompts.
The primary dataset contains 31,850,496 neuron-perturbed answers, 884,736 random-direction control answers and 36 baseline answers. Auxiliary records are separate. The 20 expanded baseline bridge records were checked against original job IDs and record hashes before inclusion, so they are not double-counted copies of the initial records.
Original saved runs contributed 437,879,099 output tokens; expanded saved runs contributed 5,402,252,603. The token census covers the collected study records, not uncollected failed attempts, calibration/warm-up runs or unrelated live chats. Input and embedding tokens are different quantities and are not included in this generated-token total. The 256-token limit was not used as an estimate.
Attached are the combined totals, exact source breakdowns and counting-scope notes.