[MODEL] Question generation loses the answer to assumption decision chain
Preflight Checklist
- [x] Existing open and closed issues were searched for duplicates.
- [x] This payload excludes credentials and private identifying details.
Type of Behavior Issue
Other unexpected behavior.
What You Asked Claude to Do
I work on AI adoption and systems design. This is a historical product failure report assembled from my local incident log; an independent controlled replay was not run. The task: annotate a set of source answers and generate follow up questions, each linked to the assumption and decision tree it tests. The private task prompt is not reproduced.
What Claude Actually Did
Claude annotated the source answers but generated the follow up questions without linking each answer to the assumption and the decision tree. A long working guide turned into a changelog. The analysis did not show the recorded contradictions at the point where the next decision needed them.
Expected Behavior
Trace each question and decision to the actual source answer and assumption. The working guide shows the decision logic and the unresolved contradictions, with complete evidence links.
Files Affected
Local artifacts are represented as <workspace>/<artifact>. Exact private filenames and paths are omitted.
Permission Mode
Not recorded in the incident log.
Can You Reproduce This?
Haven't tried to reproduce in an independent clean session. The local log records the historical occurrence.
Steps to Reproduce
- Give Claude Code a set of answers plus an assumption and decision tree, and ask for follow up questions, each linked to the assumption it tests.
- Let it produce the questions and the working guide.
- Check whether every question traces back to a source answer and an assumption.
These are historical conditions and a proposed verification procedure, not a claim of deterministic reproduction.
Claude Model
Opus 5.5, as recorded in the incident log.
Relevant Conversation
The sequence is described above with private context removed. No assistant self explanation is treated as proof of an internal mechanism.
Impact
Medium - Extra work to undo changes
OBSERVED raw tokens: 12,362,935, counted once per message ID. ESTIMATED waste share: 8.7M to 11.1M; this share is an estimate without a complete derivation. ESTIMATED source time: about 280 minutes, from 13:26 to 18:19; not all elapsed time is proven waste.
Claude Code Version
Not recorded in the incident log. This is a historical record, not a test of the latest release.
Platform
Not recorded in the incident log.
Additional Context
Evidence: private mistake log entry 168, dated 2026-09-23.
Related existing issue: #95826, same family (pasted assumptions promoted into user decisions). This report differs: the failure is in linking generated questions to an existing decision tree.
Raw usage is distinct from attributable waste and from billing. Private transcripts, people, project names, credentials and paths are omitted.