A/B/C writing test: Fable takes all three
A/B/C writing test: Fable takes all three (2026-08-13)
Ruling on 13 August: Fable writes the wiki regeneration. Retired for routing on 1 September (Grok writes). Three pages, three arms each — the page as it stands, Grok 4.6, and Fable — all working from identical research banks and a byte-identical brief. Wedge’s verdict on reading them: Fable better in all three cases, “and it’s not even close.”
What was tested
| Page | Why it was chosen |
|---|---|
| Fixed vs Growth Mindset | Reader owns the term, so a definition delivers nothing. The operating claim sits buried at line 203 of the control — could an arm find it and lead with it? |
| ICS System | Reader does not own the term, so the definition is the payload — but the page is tangled in paid-course material a reader cannot reach. |
| Attention Management — Preserving Flow | Ordinary competence against heavy evidence. The control case. |
Both arms were blind to the original page and to each other, wrote from the same bank, and were held
to the same constraints. Shared brief:
01 - Workbench/regen-2026-08/abc-test/SHARED-BRIEF.md
Mechanical layer
All six regenerated pages passed: no links lost, no composed prose carried over. The mechanical layer did not separate the arms at all, which is the expected result — it verifies compliance, not quality.
The single sharp divergence was length on ICS System: Grok cut the body from 1,113 words to 512, Fable took it to 1,293. Grok’s cut followed the bank’s own instruction to shrink the page to a door; the ruling went against it anyway.
Two flaws in the test, on the record
The blind was broken by the board. The comparison board printed each pane’s file path, and the
paths read abc-test/fable/ and abc-test/grok/. The letters were shuffled independently per page
and the shuffle worked — Fable landed B, B, A and the control moved every row — but the paths gave
the arms away regardless. Whether the ruling was made before or after noticing them is unresolved.
The arms were not equally exposed. Grok’s research lane had written roughly 260 banks against this vault before its arm ran; Fable came in cold. That cuts in Grok’s favour, so it does not undermine the result — it means a narrow Grok win would have been discountable, and a Fable win is not.
What it settles, and what it does not
Settles: Fable writes. This is the second independent comparison pointing the same way — the earlier Flow State bakeoff had Fable’s opening ruled perfect against two Opus arms on the same page from the same bank.
Does not settle: whether Grok is a weaker writer generally, or weaker at this vault’s register. It had the standard, the specimens and the accepted openings in front of it, same as Fable. That is the honest limit of a three-page sample.
Unchanged: Grok keeps the research lane, where it has been productive — 287 banks on disk against a 255-page queue. Research is agent-agnostic; writing is not.
Artifacts
- Board:
01 - Workbench/regen-2026-08/abc-test/ABC-BLIND-BOARD.html - Arms:
01 - Workbench/regen-2026-08/abc-test/{grok,fable}/ - Shared brief:
01 - Workbench/regen-2026-08/abc-test/SHARED-BRIEF.md - Writing handoff:
02 - System/Wiki Regeneration Handoff.md
Next comparison, if there is one
Strip the paths off the board. Copy each arm to a neutral filename before boarding, and keep the map outside the artifact the ruling is made from.