Thesis Control

/thesis-control - Thesis Drift Control

Purpose

Prevent AI-assisted writing from becoming fluent but distorted. Use this before and after substantive thesis or manuscript edits when the risk is not spelling or style, but loss of author control: project-level reframing, deletion of the real-world task or intended use, a primary domain becoming a secondary example, a changed research object or question, auxiliary analyses becoming primary, widened claims, blurred section purpose, missing caveats, unsynchronised adjacent paragraphs, or local edits that weaken the paper spine.

Trigger Words

This skill activates on: thesis control, drift audit, edit contract, spine card, claim drift, author control, loss of control, scope creep, rewrite risk, /thesis-control.

Core Rule

Do not edit thesis prose until the project intent, current manuscript contract, global thesis audit, section spine, and intended local change form one explicit and traceable contract chain.

The contract must answer:

This edit is allowed to change [specific local issue] in [specific unit], while preserving [spine sentence], [scope boundary], [core claims], and [do-not-change items].

If this sentence cannot be written, stop and diagnose the section instead of rewriting it.

The control hierarchy is:

Author-approved Project Intent
→ Author-approved Manuscript Contract
→ Passed Global Thesis Audit
→ Section Spine Card
→ Edit Contract
→ Post-edit Drift Audit

A lower layer cannot amend a higher one. If the title, abstract, primary domain, research object, research question, contribution scope, or manuscript structure no longer matches the approved intent, stop. Revise or roll back the manuscript, or create a new explicitly approved intent version that preserves the earlier row as history.

Control Files

Choose one control profile and name it as canonical.

For a lightweight single-manuscript workflow, read references/author_control_lightweight.md and use:

• 00_AUTHOR_INTENT.md
• 01_EVIDENCE_AND_CLAIMS.md
• 02_REVISION_LOG.md

Create and check the bundled templates with:

python scripts/scaffold-author-control.py <project_root>
python scripts/check-author-control.py <project_root> --strict

The lightweight checker validates structure, approval state, and unresolved placeholders. It does not infer semantic alignment. Do not use the lightweight profile to bypass a gate that already requires the durable packet.

Use a thesis_control/ directory when the project needs durable tracking:

• project_intent.csv
• manuscript_contracts.csv
• global_thesis_audits.csv
• spine_cards.csv
• edit_contracts.csv
• drift_audits.csv
• revision_escalations.csv

Run the optional validator when Python is available:

python {skill_dir}/scripts/check_thesis_control.py <project_root> --strict

Strict validation requires the project-intent layer. It blocks approved or applied edit contracts unless they reference a passed global thesis audit for the active author-approved intent and manuscript contract. Non-strict mode can still inspect legacy packets that do not yet have this layer.

The validator checks packet structure and recorded gate consistency. It does not infer semantic alignment or judge scholarly truth. The author or reviewer must compare the manuscript with the intent and record each alignment field honestly; the validator then prevents an unresolved or drifted audit from being used as authorisation.

Strict validation requires revision-tracking schema v3. Upgrade a complete legacy packet without guessing historical revision families:

python {skill_dir}/scripts/upgrade_thesis_control_revision_tracking.py <project_root>

To create a draft packet from a real Markdown unit before editing prose:

python {skill_dir}/scripts/scaffold_thesis_control.py <project_root> \
  --source chapters/ch1_introduction.md \
  --start-line 71 \
  --end-line 104 \
  --revision-issue-id ri-ch1-gap-clarity \
  --attempt-no 1 \
  --copy-source

The scaffold writes schema v4 draft project-intent and manuscript contracts, a pending global thesis audit, human_approved=false, status=draft, and AUTHOR_REVIEW_REQUIRED fields. Replace those fields with concrete author judgement before applying a substantive edit. A scaffolded packet may be structurally valid while remaining non-executable. Its default contract id includes the attempt number, so attempts 1 and 2 become ec-<unit>-001 and ec-<unit>-002. Reuse an explicit revision_issue_id for retries.

The migration helper stops without writing when revision metadata is partial or when a legacy escalation cannot be classified from current contracts and resolved audits. It preserves named extension columns and converts one- or two-trigger legacy rows to early_diagnostic; a three-trigger row becomes a cycle_gate only when it already matches one completed failure group.

Upgrade a complete schema-v3 packet into a deliberately blocked schema-v4 draft without guessing author intent:

python {skill_dir}/scripts/upgrade_thesis_control_project_intent.py \
  <project_root> --json

The helper adds manuscript_id and global_audit_id links, preserves named extension columns, and creates AUTHOR_REVIEW_REQUIRED draft intent, manuscript, and global-audit rows through one atomic batch. Previously approved or applied edits remain blocked. Replace the draft fields with real author judgement, approve the active intent and manuscript contract, and resolve the global audit before strict validation can pass. A partial project-intent schema stops without mutation.

Workflow

0. Establish The Project Intent And Manuscript Contract

Before section-level planning, record:

• the real-world problem and intended user or beneficiary
• the intended application and the present method or software task
• the primary scholarly domain
• the research object
• the core research question
• the primary experiment that directly answers that question
• supporting, robustness, exploratory, failed-development, and out-of-scope analyses
• the strongest evidence-licensed headline claim
• the current validation and evidence boundaries
• the target venue or audience
• concepts that must remain visible in the title or abstract
• reframes that require fresh author approval
• the current title, abstract focus, contribution scope, and structure
• concrete approval evidence and the active version ids

Keep one active author-approved project intent and one active author-approved manuscript contract. A later intent version must identify the immediately previous version in supersedes_intent_id, record the amendment reason, and leave the earlier version as superseded. Do not overwrite the original row.

Run a global thesis audit whenever the title, abstract, primary domain, research object, research question, contribution scope, or overall structure changes. Record each dimension as aligned, drifted, or not_assessed. Only a fully aligned audit with detected_reframe=false can have status=passed and human_decision=accept.

If any dimension is drifted, set human_review_required=true and use needs_review or failed. The author must choose to revise the manuscript, roll back, or approve a versioned intent amendment. Merely accepting the audit cannot authorise the reframe.

Keep intended use and current validation separate. Narrow evidence may narrow the empirical or headline claim, but it must not silently delete a legitimate application problem or recast the evidence boundary as the paper's research object. Record future application as intended use and untested hardware, clinical, causal, deployment, or transfer outcomes as unvalidated boundaries.

Before admitting a completed analysis into the paper, record whether it directly answers the core question, whether the main conclusion survives its removal, its one-sentence argumentative function, its role, destination, and author decision. Completion alone does not make an analysis a main contribution. Use the role and placement defaults in references/author_control_lightweight.md.

1. Establish Or Read The Spine Card

Before editing a chapter, section, or paragraph cluster, identify:

• unit id
• source path
• section title
• spine sentence
• scope boundary
• core claims
• do-not-change items
• the active manuscript contract id

The spine sentence should be narrow:

This unit argues that [specific claim] by showing [specific basis], so the chapter can [specific function].

If the current text does not support a clear spine sentence, produce a diagnosis and ask for author direction before changing prose.

2. Create The Edit Contract

For every substantive edit, state:

• target unit and file range
• change scope: local_patch, section_restructure, or full_reframe
• allowed changes
• forbidden changes
• evidence baseline: the ref, artifact, data, configuration, table, or frozen numbers inherited
• argument baseline: the author-approved intent and manuscript version inherited
• adjacent context that must be checked
• acceptance checks
• whether human approval is required before editing
• the passed global thesis audit id that covers the spine card's manuscript contract

When using the scaffold helper, treat its output as a draft control packet, not as approval. A generated contract becomes actionable only after the author has replaced the AUTHOR_REVIEW_REQUIRED fields and explicitly approved the scope.

Always require human approval for:

• changing the section spine
• adding or broadening claims
• deleting caveats or limitations
• moving evidence between sections
• rewriting more than one paragraph
• merging or splitting sections
• changing the title, abstract, primary domain, research object, research question, contribution scope, or manuscript structure
• changing the real-world problem, intended use, primary experiment, analysis prominence, headline claim, or evidence boundary

Treat any change to the title, abstract thesis, research object, core question, primary experiment, contribution order, evidence chain, application purpose, or paper-wide structure as a full_reframe, even when the request calls it polishing. Before editing, show an old-versus-proposed spine comparison and obtain explicit author approval.

3. Apply Only Approved Changes

After approval, edit only the approved scope. Do not apply the edit if the linked global thesis audit is pending, failed, stale, drifted, or attached to a different manuscript contract.

Keep mechanical fixes separate from argument changes. Do not bundle style, structure, evidence, and claim changes into one patch unless the contract explicitly allows it.

4. Run The Drift Audit

After editing, compare the new prose against the contract and report:

• changed claims
• changed boundaries or caveats
• new unsupported claims
• deleted evidence anchors
• missed adjacent updates
• section-spine change
• research-object or core-question change
• deleted, generalised, or demoted real-world task or intended use
• promoted auxiliary analysis
• evidence boundary rewritten as the paper topic
• loss of application meaning caused by over-cautious wording
• title, abstract, Introduction, Results, and Conclusion alignment
• decision: accept, partial accept, revise, or rollback

If any claim, boundary, or caveat changed, the result needs human review even if the prose is smoother.

Use audit status=needs_review only while the author's post-edit decision is pending. Strict validation blocks an applied contract in that state. After the author decides, record status=passed for accept or partial_accept, and status=failed for revise or rollback. Do not treat a pending audit as a completed unsuccessful attempt.

5. Record Human Gate Outcome

The author decides whether to accept, partially accept, revise, or rollback. Do not mark a high-risk edit as accepted without explicit human approval.

6. Run Post-Spine Readability Gates When Relevant

Only after the research spine is stable, use /logic-review to audit repeated argument functions and /self-review to prepare the unfamiliar-reader packet. Do not solve repetitive AI prose by generating synonyms. Remove duplicated problem, gap, evidence, interpretation, or boundary functions while preserving essential local qualifiers. A model simulation cannot pass a human unfamiliar- reader gate; record it as advisory or not_run until an actual reader responds.

Revision Escalation Rule

Treat three unsuccessful attempts on the same revision issue as an operational escalation threshold, not as evidence that every task fails after three turns. Use revision_issue_id to keep successive contract versions attached to that issue. Only count an attempt when its drift decision is revise or rollback and its audit status is failed. Only applied contracts count as unsuccessful attempts. Record author rejection as one of those decisions. Multiple failed audits of one contract still count as one attempt; contradictory passed and failed resolved audits are invalid. Clarifying discussion, pending human reviews, and unexecuted proposals do not count.

After three unsuccessful attempts, stop. Do not apply a fourth prose patch. Record a row in revision_escalations.csv; a later contract may become approved or applied only after the matching escalation has human_approved=true and status=approved.

An approved escalation closes only that group of three unsuccessful contracts. If three later contracts also receive revise or rollback, require a new escalation before another contract can proceed.

Only a cycle_gate whose three triggers exactly match one completed group of unsuccessful contracts, in attempt order, may close that group. Set approved_after_attempt to the final attempt number in that group. The gate is effective only with human_approved=true and status=approved. One gate cannot close more than one group. Do not repeat a trigger contract within a row or create multiple rows for the same issue and trigger set.

Only an escalation whose trigger set exactly matches one completed group of three unsuccessful contracts may close that group. One escalation cannot close more than one group.

Record an earlier warning as early_diagnostic with one or two unique triggers and an empty approved_after_attempt. It may be author-approved as a diagnosis, but it never closes or pre-authorises a later completed group.

An earlier escalation with fewer than three trigger contracts does not close or pre-authorise a later completed group.

Escalate earlier than three attempts when any of these signals is already visible:

• the section spine cannot be stated consistently
• the requested claim lacks supporting evidence
• a revision changes a claim, caveat, or scope boundary outside the contract
• the latest author-approved version cannot be identified
• old assumptions, duplicated explanations, or conflicting requirements indicate version contamination

Required Escalation Check

Before editing again:

1. Consolidate the currently valid requirements into one brief.
2. Compare that brief with the spine card, evidence boundaries, current contract, and latest author-approved version.
3. Classify the failure as one primary category:
   • underspecified or conflicting intent — the target, audience, venue, constraint, or acceptance condition is missing or inconsistent, or the feedback is evaluative but not operational, such as “weak”, “unclear”, or “still not right” without a concrete change target
   • local execution failure — the contract is clear, but the edit did not implement it correctly
   • structural mismatch — the problem affects the section purpose, research question, gap, contribution, evidence chain, or manuscript structure
   • evidence gap — the requested claim is not supported by the available sources, data, experiments, or files
   • version contamination — accumulated patches mix incompatible assumptions, duplicate reasoning, or obscure which prose the author approved
4. Classify the writing scope:
   • local patch — wording or presentation changes that preserve the spine, claims, evidence, and adjacent-section relationships
   • section-level restructure — changes confined to one section without changing the research question, contribution, or evidence chain
   • full reframing — changes to the title, abstract, research question, gap, contribution, methods-results alignment, evidence chain, or discussion framing
5. Recommend the smallest valid next action and wait for author approval.

Use these default actions:

• For a local execution failure, create a corrected local contract.
• For underspecified or conflicting intent, ask for the missing decision before editing.
• For a structural mismatch, propose a section-level restructure or full reframing plan before editing.
• For an evidence gap, narrow, qualify, or remove the unsupported claim unless the author supplies more evidence.
• For version contamination, restore or copy the latest author-approved version, then apply a consolidated contract. Create a separate branch or manuscript version only when the approved scope requires structural work.

For full reframing, hand off a brief that states the target venue, old and proposed real-world problem, intended use, research object, research question, primary experiment, contribution order, headline claim, evidence boundary, evidence baseline, argument baseline, available evidence, claims that must not be made, and proposed new structure. Do not rewrite the manuscript until the author approves that brief.

Return the escalation check in this form:

## Revision Escalation Check

Revision issue:
Contract:
Unsuccessful attempts:
Trigger contracts:
Primary category:
Writing scope:
Why the revisions did not converge:
Valid requirements:
Missing or conflicting information:
Latest author-approved version:
Recommended next action:
Author decision required:

Output Patterns

Audit Only

Return:

• current spine diagnosis
• likely drift risks
• control gaps
• recommended edit contracts
• blocked items needing author decision

Pre-Edit Contract

Return:

## Edit Contract

Unit:
Spine sentence:
Scope: local_patch / section_restructure / full_reframe
Evidence baseline:
Argument baseline:
Allowed changes:
Forbidden changes:
Adjacent context to check:
Acceptance checks:
Human approval required:
Proceed only after approval:

Post-Edit Drift Audit

Return:

## Drift Audit

Contract:
Changed claims:
Changed boundaries:
New unsupported claims:
Missed adjacent updates:
Research object or question changed:
Intended use deleted or demoted:
Auxiliary analysis promoted:
Cross-section paper identity aligned:
Decision:
Human review required:
Recommended next action:

Stop Conditions

Stop and ask for author direction if:

• the section spine cannot be stated clearly
• the requested edit would broaden a claim without evidence
• a local edit requires adjacent updates outside the approved scope
• the user asks for a full-chapter rewrite without a spine map
• previous AI edits cannot be distinguished from author-approved text
• the edit would remove caveats, limitations, or uncertainty language without explicit approval
• the active project intent or manuscript contract cannot be identified
• the real-world task, intended use, primary experiment, evidence baseline, or argument baseline cannot be identified
• a global thesis audit is missing, unresolved, stale, or records project-level drift
• a proposed local contract would preserve a section spine that conflicts with the author-approved project intent
• a full reframe lacks an approved old-versus-proposed spine comparison