docs: how we improve tallyman from session reviews, the error corpus and the eval suite

CLAUDE.md

@@ -12,3 +12,9 @@ itself. There is no installed base and no third-party data on disk.
   The only catalogs are mine, and I'll rebuild them.
 - This is about not spending effort on migration/compat paths, not about
   lowering quality: still write correct code and tests for current behavior.
+
+## Reviewing sessions, hints and evals
+When asked to look at a past Claude session that used tallyman, to add or
+check error hints, or to run or extend the eval suite, read
+`docs/improving-from-sessions.md` first. It says where transcripts and error
+logs live, how a review is done, and what each kind of finding turns into.

docs/architecture.md

@@ -910,6 +910,10 @@ code wins.
 - [mcp-server.md](mcp-server.md): every MCP tool and the prompt, with
   parameters, return shapes and side effects. **Current.**
 - [installing.md](installing.md): install and run tallyman. **Current.**
+- [improving-from-sessions.md](improving-from-sessions.md): how we review
+  Claude sessions that drove tallyman, check hints against every recorded
+  error, and replay sessions in the eval suite, and what each kind of finding
+  turns into. **Current** as of 2026-09-29.
 
 ### Design records (`plans/ADR-*.md`)