SOFT CAT.ai
FIND SOMETHING USEFUL
← The build notebook

THE LONGER RECORD

Thoughts & corrections.

Earlier writing that still has a useful point, revised where its claims went beyond the evidence. Corrections explain what changed and link to the earlier text.

For documented experiments and maintenance, start with the notebook.

CHANGING OUR MINDS IN PUBLIC

The correction record

Shorter prompts still need a fair test

Removed an unsupported 80 per cent prompt-reduction claim and the conclusion that prompting had become obsolete.

Originally published 3 Jul 2026

Measure cost changes before assigning motives

Withdrew unsupported claims about hidden monitoring, a 40 per cent token increase and deliberate price manipulation.

Originally published 2 Jul 2026

Hybrid routing should show where a request goes

Removed the mental-health metaphor and unsupported claims about how all hybrid systems behave. The useful issue is request routing and disclosure.

Originally published 6 Jun 2026

Agent controls need tests at the boundary

Withdrew blanket claims that security frameworks are useless and that a confused agent will inevitably escape a sandbox.

Originally published 19 Mar 2026

The original SOFT CAT pipeline

Reframed the launch article as project history. Removed claims that all six bots ran this morning, corrected the treatment of schedules and linked the current records. The separate server is not verified by a new website build.

Originally published 13 Mar 2026

Context size needs a retrieval test

Removed an undocumented whole-project test and the blanket conclusion that a larger context window beats a more capable model.

Originally published 14 Feb 2026

Document parsing still needs tests

Withdrew the solved-problem claim and the unsupported claim of equal cost and latency across document lengths.

Originally published 25 Jun 2026