How it works
Methodology
The Global Productivity Index compares what humanity did in a given hour with what it normally does in that hour on that weekday. 'Normal' is the median of 4 comparison weeks. An index of 100 means business as usual; 90 means measurable output is ten percent lower than expected. We use public data, publish our weights, and release every number. The method is real. The panic is not. Correlation is not causation. It is, however, extremely funny.
No score until a baseline supports one
The index stays unlisted while the comparison weeks cannot support it. A missing score means the method is waiting, not that activity is normal or low. No score is inferred from missing instruments.
Direct comparisons
Local metrics and instrument cards can compare one completed UTC bucket with the same bucket 168 hours earlier, falling back to 24 hours earlier when enabled. Every comparison names its actual date, source and basis. Missing and zero comparison values remain unavailable. These comparisons do not wait for the composite index baseline below.
The baseline
Each signal is compared with the same hour in 4 earlier comparison weeks. The weeks freeze at 2026-11-18 00:00 UTC, before the event window opens.
| Week starting | Clock shift (hours) |
|---|---|
| 0 | |
| 1 | |
| 1 | |
| 1 |
Signals marked provisional rest on a short, recent history: weekdays 2026-11-16, 2026-11-17, Saturday 2026-11-14 and Sunday 2026-11-15. They use a wider noise floor of 0.05.
The statistic
Normal is the median of the comparison weeks, which resists one odd week. Deviation is measured against the typical spread of those weeks, never less than a floor of 0.03, and clipped at 8 in either direction so no single signal dominates.
Weights
Only signals with a nonzero weight enter the index; the rest are shown for context. The weights are provisional values, not a validated model.
| Signal | Role | Weight | Baseline | Cadence |
|---|---|---|---|---|
| GitHub pull requests | Work | 47.1% | History | Hourly |
| Wikipedia edits | Work | 29.4% | Provisional | Hourly |
| Hacker News posts | Work | 23.5% | History | Hourly |
| Steam players | Displacement | Context only | Provisional | Hourly |
| Launch article page views | Attention | Context only | History | Hourly |
| Internet traffic | Displacement | Context only | History | Hourly |
| GTA VI article views | Displacement | Context only | History | Daily |
| Twitch | Displacement | Context only | Provisional | Hourly |
| npm downloads | Control | Context only | History | Daily |
| PyPI downloads | Control | Context only | History | Daily |
Weights of scored signals total 100%.
When a reading is withheld
If the instruments that are reporting carry less than 50% of the total weight, the reading is withheld and no number is shown. A signal older than 3 hours is stale and drops out, and the remaining weights are renormalized. A source with no update for 90 minutes is marked delayed.
Alert levels
The level comes from the weighted deviation score. Above an index of 98 the site reports no measurable effect, whatever the level.
| Level | Name | Score range |
|---|---|---|
| 0 | SUSPICIOUS DILIGENCE | below โ2 |
| 1 | NOMINAL | โ2 to 1 |
| 2 | ELEVATED SLACKING | 1 to 2 |
| 3 | SIGNIFICANT ABSENTEEISM | 2 to 3 |
| 4 | WIDESPREAD DERELICTION | 3 to 4.5 |
| 5 | TOTAL COLLAPSE | 4.5 and above |
The change figures on the board
Each tile compares its latest reading with the average of its earlier readings in the same snapshot window. At least 3 readings are needed, and a change within 0.5% is shown as unchanged. This is not a comparison with the baseline and does not measure a launch effect.
What this does not measure
These limits apply to every number on this site.
- It does not measure a representative share of global economic output. Each instrument is a narrow, observable subset of activity.
- It does not separate people from automation. Pull requests, commits, edits and tokens can come from bots and agents running without a person at the keyboard, and a bot-name filter alone does not establish human origin.
- Steam totals include every game, not only the launch title, and say nothing about time taken away from work.
- A GitHub count describes public activity. It is not a count of distinct contributors, reviews or human participation until those measures are validated.
- Daily sources cannot support any claim about a lunchtime effect. A launch-day change and a local lunch window are separate hypotheses.
- It does not locate people. Contributor location is not inferred from repository ownership, and gamers are not identified from a traffic category.
- It does not establish causation. Ordinary weekday patterns and provider outages are not labeled launch effects, and a drop is only called one after it persists and recovers.
- Platform figures describe that platform, not all model usage. An organization's usage report is not provider-wide usage.
- A missing instrument is not read as zero. A null result is a valid result, and a quiet chart is not evidence of a collapse.
- Observed activity is not a measure of economic loss. This is not financial advice.