fix: climb the Pro ladder as GPT-6 Pro, GPT-5.6 Sol Pro, then Latest at Extra High #81

Merged

xicv merged 2 commits from fix/ladder-order into main

2026-09-25 06:11:24 +00:00

xicv commented

2026-09-25 06:11:15 +00:00

(Migrated from github.com)

Summary

Corrects the "Pro only" ladder to the order the user intended: GPT-6 Pro → GPT-5.6 Sol Pro → Latest at Extra High → stop. v0.2.41 tried Latest at Extra High before GPT-5.6 Sol Pro.

Rung 1 is unchanged: the strongest route (Latest) at Pro.
Rung 2 selects GPT-5.6 Sol and accepts it only at a proven Pro maximum. An unreadable Power description still stops as effort_unreadable.
Rung 3 accepts Latest only when it is named exactly Latest and its maximum reads Extra High. When rung 2 selected GPT-5.6 Sol, rung 3 selects Latest again, raises Power to the maximum and proves Extra High anew. When the menu had no GPT-5.6 Sol option, Latest is still selected and is used as it is.
The stop message and evidence.proLadder follow the new order.
Browser contract 36, runtime generation 2026-09-25.3.

Verification

The tests now encode the new order:
- at the GPT-6 Pro limit the ladder sends at GPT-5.6 Sol Pro;
- with Sol Pro used up it returns to Latest at Extra High and pins the readback to it;
- with no Sol option it stays on Latest;
- the stop's evidence lists the rungs in order.
- The return-to-Latest test fails when the re-selection is removed.
npm test: 1524 tests, 1523 pass, 1 skipped. npm run lint clean. A3K fixture regenerated.
Live on a dev broker (2026-09-25, GPT-6 Pro capped): a Pro only exchange on the new layout went out at route sol_pro (GPT-5.6 Sol, Pro) and was answered by gpt-5-6-pro, with a tail-v2 head.
Codex review (read-only): the first pass flagged a node_modules symlink committed by mistake in the fixture commit; it was removed before push. The second pass found no actionable regression. (Codex's sandbox cannot run the suites; they were run locally.)

## Summary Corrects the "Pro only" ladder to the order the user intended: **GPT-6 Pro → GPT-5.6 Sol Pro → Latest at Extra High → stop**. v0.2.41 tried Latest at Extra High before GPT-5.6 Sol Pro. - Rung 1 is unchanged: the strongest route (Latest) at Pro. - Rung 2 selects GPT-5.6 Sol and accepts it only at a proven Pro maximum. An unreadable Power description still stops as `effort_unreadable`. - Rung 3 accepts Latest only when it is named exactly `Latest` and its maximum reads Extra High. When rung 2 selected GPT-5.6 Sol, rung 3 selects Latest again, raises Power to the maximum and proves Extra High anew. When the menu had no GPT-5.6 Sol option, Latest is still selected and is used as it is. - The stop message and `evidence.proLadder` follow the new order. - Browser contract 36, runtime generation 2026-09-25.3. ## Verification - The tests now encode the new order: - at the GPT-6 Pro limit the ladder sends at GPT-5.6 Sol Pro; - with Sol Pro used up it returns to Latest at Extra High and pins the readback to it; - with no Sol option it stays on Latest; - the stop's evidence lists the rungs in order. - The return-to-Latest test fails when the re-selection is removed. - `npm test`: 1524 tests, 1523 pass, 1 skipped. `npm run lint` clean. A3K fixture regenerated. - Live on a dev broker (2026-09-25, GPT-6 Pro capped): a Pro only exchange on the new layout went out at route `sol_pro` (GPT-5.6 Sol, Pro) and was answered by `gpt-5-6-pro`, with a `tail-v2` head. - Codex review (read-only): the first pass flagged a `node_modules` symlink committed by mistake in the fixture commit; it was removed before push. The second pass found no actionable regression. (Codex's sandbox cannot run the suites; they were run locally.)