Issues

docs/API.md: the 0b/0c/7b/7c letter scheme implies sub-sections that do not exist

#417

· bbertucc opened

on Sep 5, 2026

page.md decides some printed conventions one case at a time and others not at all, so every model invents the rest

#374

· bbertucc opened

on Sep 3, 2026

Iris transcribes every page twice and the second pass is most of the bill — but writing the missing rules down removes only about a seventh of it

#369

· bbertucc opened

on Sep 3, 2026

The checker and the corrector both bill for working-out Iris throws away, and the clause that stopped it on a third agent is already in the checker's prompt — 61 lines from where it would bind

#365

· bbertucc opened

on Sep 3, 2026

Repoint the page agent from kimi-k2.5 to gpt-5.6-luna — one config line, 15.9% cheaper per 100 pages, lost pages 2 to 0

#344

· bbertucc opened

on Sep 3, 2026

Four page-agent defects survive on the model Iris ships, and on three of them it is worse than the arm I audited — two are fixable in code with no model call

#334

· bbertucc opened

on Sep 2, 2026

The free table-join path took 9 of 17 pairs one round and 4 the next with the join code unchanged: a $0.72/100-page swing caused by unstable header re-extraction

#326

· bbertucc opened

on Sep 2, 2026

The copy editor is 33.1% of the bill: a window that hits the 32,000-token ceiling costs 5.2x one that fits, and #300's salvage correctly recovered 0 of 2

#317

· bbertucc opened

on Sep 1, 2026

Quality regression: too many documents ship with unresolved issues

monitoring

#264

· github-actions opened

on Aug 29, 2026

Looking to contribute?

good first issue

#85

· bbertucc opened

on Aug 13, 2026

Navigation Menu

Uh oh!

All issues

docs/API.md: the 0b/0c/7b/7c letter scheme implies sub-sections that do not exist

page.md decides some printed conventions one case at a time and others not at all, so every model invents the rest

Iris transcribes every page twice and the second pass is most of the bill — but writing the missing rules down removes only about a seventh of it

The checker and the corrector both bill for working-out Iris throws away, and the clause that stopped it on a third agent is already in the checker's prompt — 61 lines from where it would bind

Repoint the page agent from kimi-k2.5 to gpt-5.6-luna — one config line, 15.9% cheaper per 100 pages, lost pages 2 to 0

Four page-agent defects survive on the model Iris ships, and on three of them it is worse than the arm I audited — two are fixable in code with no model call

The free table-join path took 9 of 17 pairs one round and 4 the next with the join code unchanged: a $0.72/100-page swing caused by unstable header re-extraction

The copy editor is 33.1% of the bill: a window that hits the 32,000-token ceiling costs 5.2x one that fits, and #300's salvage correctly recovered 0 of 2

Quality regression: too many documents ship with unresolved issues

Looking to contribute?

Uh oh!

Search results