Documentation only: a dated note under 0044's thresholds, and the status line's open item.

The note says:

• The tokens-per-document threshold (about 6k) is not met. With thinking off in every call (Every model call uses the workspace's reasoning effort, and none is a value: alignment no longer thinks on its own #1066) a fresh base spends 19k–22k on deepseek-flash, about three times the threshold; with alignment thinking it was 31k–48k.
• What is left is not thinking. Extraction and time interpretation alone come to the threshold, and the rest is prompts that restate the contract, the properties and the candidates on every call, paid per new phrase signature. Signatures grow with documents, so a warm base saves only 13–18%.
• Prompting a general-purpose model differently does not close a gap of three times; each cut so far took tens of percent.
• The planned route is a model trained for these steps. It is not built yet and nothing about it is measured. Until it is, the threshold stays as written and is recorded as not met.

🤖 Generated with Claude Code