One paired length comparison
Fourteen text types each supplied a short public-domain collage (“Brief,” 105–219 words) and a generated output padded toward the source length (“Keep,” roughly 1,000–1,300 words).
Archived experiment · not the current product
LENGTH WAS A CONFOUND.
Pangram 4.0 · 14 types · brief vs keep · 2 September 2026
LENGTH WAS A CONFOUND.
SO WE KEPT THE FAILURE.
This early experiment compared short public-domain collages with outputs padded back to source length. Repeated generated asides dominated the result. It is preserved here as a lab record, not as evidence that shortening text makes it human.
- Brief condition
- 12 / 14
- Entire-text scans labelled Human
- Keep condition
- 0 / 14
- Entire-text scans labelled Human
- Best Keep result
- 53%
- Human share for Speech
- Observed Keep length
- 103–106%
- Of source word count
What this run actually isolated
Attested openers often stayed human while repeated generated filler shifted the scan. That finding killed the padding strategy; it did not validate a detector-bypass product.
Human share · Brief vs Keep
Pangram 4.0 human-written percent. Green is the short collage. Orange is the same type at Keep length. These are detector outputs, not authorship judgments.
Brief (105–219 words)
Keep (~1,000–1,300 words)
Full matrix
Credits: Brief 2905→2876 (29). Keep 2876→2718 (158). Detector: Pangram 4.0 dashboard, Individual plan. “Human” and “AI” below reproduce its recorded labels.
Rendering 14 recorded rows…
| Type | Source words | Brief words | Keep words | Brief human | Keep human | Keep AI |
|---|---|---|---|---|---|---|
| Rendering recorded outcomes… | ||||||
Method & provenance
Enough detail to interpret the archive without turning an exploratory dashboard run into a product claim.
Counts, hashes, and recorded outcomes
Source rows, generated artifacts, and manually recorded detector outcomes are linked by hashes in the research fixtures. This page publishes the aggregate record, including the failed condition.
Diagnostic, not causal
This archive does not report a randomized, blinded, or powered estimate. Length and repeated filler changed together, so the run cannot assign causality or establish a pass rate.
Evidence map
The source rows, generated artifacts, and manually recorded detector outcomes remain linked by hashes in the research fixtures. The visual below is an editorial illustration, not a detector screenshot.
