Add lost-in-conversation writing control bench

README.md

@@ -29,6 +29,8 @@ python3 .claude/skills/release-governance/scripts/check_release_packet.py exampl
 
 Then open [`examples/demo-project/`](examples/demo-project/) in your agent host and ask it to explain the local workflow.
 
+The repository also includes a small public-safe writing-control fixture at [`examples/lost-in-conversation-bench`](examples/lost-in-conversation-bench). It compares normal multi-turn editing, consolidated prompting, and `/thesis-control` artifacts for author-control review.
+
 ## Common use cases
 
 - [Write a literature review](docs/use-cases/write-literature-review.md)

docs/product/lost-in-conversation-writing-control-design.md

@@ -144,3 +144,11 @@ The safe public claim is:
 ```text
 Academic Writing Toolkit turns scattered multi-turn writing intent into thesis-control artifacts before substantive edits are applied, helping authors inspect claim drift, evidence boundaries, and acceptance decisions.
 ```
+
+## Implementation Pointer
+
+The first implementation should use the public-safe fixture in `examples/lost-in-conversation-bench/` and the structural checker `scripts/check_lost_in_conversation_bench.py`. The treatment packet should also validate with:
+
+```sh
+python3 .claude/skills/thesis-control/scripts/check_thesis_control.py examples/lost-in-conversation-bench/treatment --strict --json
+```

examples/lost-in-conversation-bench/README.md

@@ -0,0 +1,18 @@
+# Lost-in-Conversation Writing Control Bench
+
+This fixture demonstrates a local, public-safe evaluation for AI-assisted thesis editing drift.
+
+The bench compares three workflows on the same desensitised thesis-style section:
+
+1. Baseline A: normal multi-turn chat editing.
+2. Baseline B: a consolidated single-turn prompt.
+3. Treatment: `/thesis-control` with a spine card, edit contract, bounded edit, drift audit, and author decision.
+
+The fixture is not a model benchmark and does not claim scientific validity. It is a product-control check: can the author inspect what changed, decide whether the edit stayed inside scope, and choose accept, partial accept, revise, or rollback without rereading a long chat history?
+
+Run from the repository root:
+
+```bash
+python3 scripts/check_lost_in_conversation_bench.py examples/lost-in-conversation-bench --json
+python3 .claude/skills/thesis-control/scripts/check_thesis_control.py examples/lost-in-conversation-bench/treatment --strict --json
+```

examples/lost-in-conversation-bench/baselines/baseline_a_review.md

@@ -0,0 +1,19 @@
+# Baseline A Review: Normal Multi-Turn Chat Edit
+
+## Workflow
+
+The editor receives the requirements over five turns and revises as the conversation evolves. This represents ordinary chat-based writing assistance where later constraints may arrive after an early answer attempt.
+
+## Metric Review
+
+| Metric | Review |
+| --- | --- |
+| Spine preservation | Medium risk. The section can become clearer while losing the narrow function of motivating edit boundaries. |
+| Claim drift | High risk. The request to make the prose stronger can broaden the claim into a general criticism of AI writing. |
+| Evidence boundary | High risk. Source A and Source B may be treated as support for model unreliability rather than workflow-control need. |
+| Scope discipline | Medium risk. Later requests can pull implementation architecture into a motivation section. |
+| Author control recovery | Weak. The author must reread the conversation to reconstruct what was allowed and what changed. |
+
+## Control Finding
+
+This workflow is useful for fluency but weak for author control. It does not produce a durable contract or drift audit.

examples/lost-in-conversation-bench/baselines/baseline_b_review.md

@@ -0,0 +1,19 @@
+# Baseline B Review: Consolidated Single-Turn Edit
+
+## Workflow
+
+The scattered requirements are consolidated before editing. This reduces the chance that the editor over-relies on an early assumption, but the output is still not bound to a durable author-control record.
+
+## Metric Review
+
+| Metric | Review |
+| --- | --- |
+| Spine preservation | Better than Baseline A because the spine is stated before editing. |
+| Claim drift | Lower risk than Baseline A, but the author still needs to inspect the final prose manually. |
+| Evidence boundary | Lower risk because the prompt names the source boundary explicitly. |
+| Scope discipline | Lower risk because implementation details are explicitly forbidden. |
+| Author control recovery | Partial. The consolidated prompt is reviewable, but there is no CSV contract, applied status, or post-edit audit. |
+
+## Control Finding
+
+Consolidation helps, but it remains a prompt technique. It does not create the reviewable artifacts needed for accepting, partially accepting, revising, or rolling back the edit.

examples/lost-in-conversation-bench/chapters/desensitized_section.md

@@ -0,0 +1,7 @@
+# Section 2.3: Evidence Boundaries In Assisted Revision
+
+The chapter argues that AI-assisted thesis revision is useful only when the author can inspect how local edits affect the section's claim boundary. A revision may improve sentence flow while also weakening the distinction between what the sources show and what the author infers from them. This matters because the section is not trying to prove that all AI-assisted writing is unreliable. It is trying to show why revision workflows need explicit controls around claims, caveats, and evidence use.
+
+The current section draws on two source notes. Source A describes a writing-support workflow where notes, claims, and revision decisions are kept as separate records. Source B discusses multi-turn assistant use and warns that later clarifications can be folded into earlier assumptions. Together, these sources support a bounded claim: thesis writers need visible edit boundaries when using AI across several turns. They do not support a stronger claim that any particular model is unsafe for all writing tasks.
+
+The next subsection will discuss implementation, so this section should not introduce product architecture in detail. Its role is to motivate the control problem and define what must not drift during revision. The key caveat is that the concern is not fluency by itself. The concern is whether a smoother paragraph still preserves the author's intended claim, the evidence boundary, and the relationship to neighbouring sections.

examples/lost-in-conversation-bench/requirements/consolidated_prompt.md

@@ -0,0 +1,11 @@
+# Consolidated Prompt
+
+Revise `chapters/desensitized_section.md` for clarity and reduced repetition while preserving the section's current role.
+
+Required controls:
+
+- Spine preservation: the section must still argue that AI-assisted thesis revision needs visible edit boundaries because fluency can hide claim drift.
+- Claim drift: do not broaden the claim into a statement that all AI writing or all models are unreliable.
+- Evidence boundary: Source A and Source B support the need for visible edit controls; they do not support universal model-safety claims.
+- Scope discipline: do not add product architecture or implementation details because the next subsection handles implementation.
+- Author control recovery: after revising, report whether the spine, claim boundary, evidence boundary, and caveat survived.

examples/lost-in-conversation-bench/requirements/multi_turn_requirements.md

@@ -0,0 +1,21 @@
+# Multi-Turn Requirements
+
+## Turn 1
+
+Please make this section clearer and less repetitive.
+
+## Turn 2
+
+Do not make the claim broader. The section should not say that all AI writing is unreliable.
+
+## Turn 3
+
+Keep the evidence boundary visible: Source A and Source B support the need for visible edit controls, not a universal model-safety claim.
+
+## Turn 4
+
+Do not move into product architecture yet because the next subsection handles implementation.
+
+## Turn 5
+
+After editing, tell me whether the section spine, claim boundary, and caveat survived.

examples/lost-in-conversation-bench/treatment/edited_section.md

@@ -0,0 +1,7 @@
+# Section 2.3: Evidence Boundaries In Assisted Revision
+
+This section argues that AI-assisted thesis revision is useful only when the author can inspect how a local edit affects the section's claim boundary. A paragraph can become smoother while also blurring the difference between what the sources show and what the author infers from them. The point is therefore narrow: the section does not claim that all AI-assisted writing is unreliable. It shows why revision workflows need explicit controls around claims, caveats, and evidence use.
+
+The section relies on two source notes. Source A describes a writing-support workflow that keeps notes, claims, and revision decisions as separate records. Source B discusses multi-turn assistant use and warns that later clarifications can be absorbed into earlier assumptions. Together, these sources support a bounded claim: thesis writers need visible edit boundaries when they use AI across several turns. They do not support the stronger claim that any particular model is unsafe for every writing task.
+
+The next subsection turns to implementation, so this section keeps product architecture out of scope. Its role is to define the control problem and state what must not drift during revision. The key caveat is that fluency is not the target by itself. The target is a revision process in which smoother prose still preserves the author's intended claim, the evidence boundary, and the relationship to neighbouring sections.

examples/lost-in-conversation-bench/treatment/review_report.md

@@ -0,0 +1,19 @@
+# Treatment Review: Thesis-Control Edit
+
+## Workflow
+
+The same multi-turn requirements are converted into a spine card, edit contract, bounded edit, and drift audit before the final prose is accepted.
+
+## Metric Review
+
+| Metric | Review |
+| --- | --- |
+| Spine preservation | Strong. The spine card states the section function before editing, and the edited section keeps that function. |
+| Claim drift | Strong. The edit preserves the bounded claim and avoids universal AI-writing unreliability. |
+| Evidence boundary | Strong. The edit keeps Source A and Source B tied to workflow-control need rather than broad model-safety claims. |
+| Scope discipline | Strong. Product architecture remains outside this motivation section. |
+| Author control recovery | Strong. The author can inspect `spine_cards.csv`, `edit_contracts.csv`, and `drift_audits.csv` without reconstructing the full chat history. |
+
+## Control Finding
+
+The treatment improves author control because the approval decision is tied to explicit artifacts: the spine card, applied edit contract, and drift audit. The prose is not accepted merely because it reads better.

examples/lost-in-conversation-bench/treatment/source_excerpts/lost-conversation-section.md

@@ -0,0 +1,7 @@
+# Section 2.3: Evidence Boundaries In Assisted Revision
+
+The chapter argues that AI-assisted thesis revision is useful only when the author can inspect how local edits affect the section's claim boundary. A revision may improve sentence flow while also weakening the distinction between what the sources show and what the author infers from them. This matters because the section is not trying to prove that all AI-assisted writing is unreliable. It is trying to show why revision workflows need explicit controls around claims, caveats, and evidence use.
+
+The current section draws on two source notes. Source A describes a writing-support workflow where notes, claims, and revision decisions are kept as separate records. Source B discusses multi-turn assistant use and warns that later clarifications can be folded into earlier assumptions. Together, these sources support a bounded claim: thesis writers need visible edit boundaries when using AI across several turns. They do not support a stronger claim that any particular model is unsafe for all writing tasks.
+
+The next subsection will discuss implementation, so this section should not introduce product architecture in detail. Its role is to motivate the control problem and define what must not drift during revision. The key caveat is that the concern is not fluency by itself. The concern is whether a smoother paragraph still preserves the author's intended claim, the evidence boundary, and the relationship to neighbouring sections.

examples/lost-in-conversation-bench/treatment/thesis_control/drift_audits.csv

@@ -0,0 +1,2 @@
+audit_id,contract_id,changed_claims,changed_boundaries,new_unsupported_claims,missed_adjacent_updates,drift_decision,human_review_required,status
+da-lost-conversation-section-001,ec-lost-conversation-section-001,none,none,none,none,accept,false,passed

examples/lost-in-conversation-bench/treatment/thesis_control/edit_contracts.csv

@@ -0,0 +1,2 @@
+contract_id,unit_id,change_scope,allowed_changes,forbidden_changes,adjacent_context,acceptance_checks,human_approved,status
+ec-lost-conversation-section-001,lost-conversation-section,Revise the three-paragraph section for clarity and reduced repetition while preserving the motivation role.,Improve sentence flow; reduce repeated phrasing; make the caveat and source boundary easier to inspect.,Do not broaden into universal AI-writing unreliability; do not add product architecture; do not introduce new evidence or model-safety claims.,Check that the next subsection remains responsible for implementation details.,Spine is preserved; claim remains bounded; evidence boundary is explicit; no unsupported universal claim is added.,true,applied

examples/lost-in-conversation-bench/treatment/thesis_control/spine_cards.csv

@@ -0,0 +1,2 @@
+unit_id,path,section_title,spine_sentence,scope_boundary,core_claims,do_not_change
+lost-conversation-section,source_excerpts/lost-conversation-section.md,Evidence Boundaries In Assisted Revision,This unit argues that AI-assisted thesis revision needs visible edit boundaries because fluent local edits can hide claim and evidence-boundary drift.,Motivate the writing-control problem only; do not introduce product architecture or universal model-safety claims.,Thesis writers need visible edit boundaries when using AI across several turns; Source A and Source B support workflow-control need rather than universal model unreliability.,Do not claim all AI writing is unreliable; do not claim any specific model is unsafe for every writing task; do not move implementation architecture into this section.