thesis_coherence_audit v1.0.0

name: "thesis_coherence_audit" description: "Audit a thesis or dissertation as one connected system before submission: build a chain map that traces each research question through theory, hypotheses, variables, operational definitions, method, data, and analysis to results and conclusions, flag drift, renamed variables, orphan results, and conclusions stronger than the design supports, run change impact analysis before any late edit, check supervisor feedback against what the text actually shows, and finish with a separate submission readiness pass. Ships two tested CLIs (chain_audit.py and feedback_tracker.py), per-link checks, a readiness reference, templates, and a worked audit on a full fictional sample thesis with seven planted faults, their fixes, and a clean re-run. Use when a thesis is near submission, when chapters were written months apart, after a method or question changed mid-candidature, when a supervisor asks whether the thesis holds together, or before an examiner ever sees it. Not for reviewing someone else's manuscript (use peer review), not for checking whether citations support claims (use citation integrity audit), and not for drafting a literature review (use academic related work)."

Thesis Coherence Audit

Purpose

Answer one question with evidence: does this thesis hold together as one system, so that the study the introduction promises is the study the methods ran, the results report, and the conclusions claim. Examiners grade exactly this. The University of Sydney's examiner guidelines require a thesis to form a cohesive and united whole. Walden University's prospectus rubric asks whether the components of the research plan align overall. This audit turns those judgements into checks. It produces a chain map, a findings list with severities, a change impact record for any late edits, a supervisor feedback state, a readiness verdict, and a fix order. It does not grade the quality of the research itself, rewrite chapters, or judge whether the topic matters.

Workflow

1. Set the mode, the document, and the institution. Say at the start whether this run is report only or fixes applied, which thesis version is being audited (file name and date), and which institution's regulations govern submission. Institutional rules are an input the candidate supplies. Never assert a word limit, page order, declaration wording, or deadline from memory of another university.

1. Extract the chain map. Start from assets/chain-map-template.csv. One row per research question. Fill each cell with the thesis's own wording, quoted or tightly paraphrased from the text, with the chapter or page noted in your working notes: question, hypothesis, variables, operational definitions, theory, method, data, analysis, results, conclusion, limitations. Extraction is the audit. A cell you cannot fill from the text is not a gap in your spreadsheet, it is a finding about the thesis. For a thesis with no research questions (theoretical or argumentative work), map problem and thesis statement through argument steps to conclusions instead, and say that is the chain you used.

1. Run the chain audit. Run bin/chain_audit.py <chain.csv> --out-dir <audit-folder>. It writes audit.md and findings.csv and checks missing links, orphan results, variable term drift between the setup chapters and the method, analysis, results, and conclusion links, causal conclusion wording on non-causal designs, and statistical wording inside qualitative questions. Exit code 0 means no major findings, 1 means major findings exist, 2 is a usage error. Read every finding against the text before reporting it. The script reads the map you extracted, so a finding can mean the map is wrong. That distinction goes in the report.

1. Walk the links the script cannot see. Follow references/chain-links.md link by link: problem to purpose term match, questions identical everywhere they are restated, theory constructs traced into instruments and discussion, population terms identical in questions, sample, and conclusions, analyses in the methods matching analyses in the results in both directions, discussion covering every result including null results, conclusions numbered against questions one to one, limitations and recommendations traced to findings, and the abstract audited last against the finished chain. Record each link as checked and clean, checked with finding, or not checked, with the evidence. Name the drift pattern when one fits: renamed variable, shrinking population, vanishing theory, method swap without a paper trail, orphan results, conclusion inflation, abstract drift, or a chapter 1 fossil describing the proposal instead of the study that ran.

1. Run change impact before any late edit lands. Any change after the audit starts (a renamed variable, a dropped question, a changed design, a new sample) goes through bin/chain_audit.py <chain.csv> --changed "<term>" first. The output lists every question and link the term touches and the downstream links to recheck. Work that list, rerun the audit, and record both in the report. A change that edits one chapter and skips the downstream list is how drift is created, and the audit treats an unrun impact list as an open finding.

1. Check supervisor feedback against the text. Load the feedback tracker (assets/feedback-tracker-template.csv shows the shape) and run bin/feedback_tracker.py <feedback.csv> --out-dir <audit-folder> --today <date>. It counts items by status, lists what is not yet verified by chapter, and flags overdue items and items marked addressed with no action recorded. Then verify a sample of addressed items against the thesis text itself. A tracker is a claim. The sample in examples/worked-audit.md shows an item marked addressed that the chain audit proved was never fixed, because the action field was blank and nobody checked the chapter.

1. Run the submission readiness pass separately. Follow references/submission-readiness.md against the institution's current regulations: title identical to the registered title, declarations present and true (authorship, AI use where required, ethics approval numbers cited in the body where required, co-authorship and editing statements where required), contents and lists regenerated from the final text, references matching citations one to one in both directions, tracked changes and comments cleared, numbers consistent across abstract, results, and conclusions, PDF export checked as examiners will see it, examiner appointment lead time observed, and the submitted version frozen and recorded. Report readiness apart from coherence. Neither verdict hides the other.

1. Build the fix order and, in fixes mode, prove each fix. Order fixes by chain damage, not by chapter order: orphan results and unanswered questions first, then term drift and population drift, then overclaims, then missing minor links, then readiness items. Each fix names the chapters it touches and the recheck that proves it landed. In fixes mode, apply the fix to the text and the chain map, rerun both tools, and report the new counts next to the old ones. A fix without a clean re-run of the affected checks is claimed, not done.

1. Check the worked example when unsure. examples/worked-audit.md audits a full fictional sample thesis end to end. Its chain map (examples/sample-chain.csv) produces 7 findings, 4 major and 3 minor, verdict NOT READY, exit code 1. Its feedback tracker produces 2 findings with 1 overdue item. Its fixed chain map (examples/sample-chain-fixed.csv) produces 0 findings, verdict CLEAN, exit code 0. Rerun those commands before trusting the format of your own report.

Output Contract

An audit folder containing: the completed chain map CSV, audit.md and findings.csv from the chain audit with every finding read against the text and marked confirmed or extraction error, the manual link checklist from step 4 with evidence per link, any change impact runs with their downstream recheck results, tracker.md from the feedback tracker plus the result of verifying addressed items against the text, the readiness pass results against the named institution's regulations, and a finished report built from assets/audit-report-template.md giving the two verdicts separately, the fix order, and a plain list of what was not checked. No verdict is reported without the chain map and the findings file behind it.

Operating Rules

1. The thesis text outranks every tracker and summary. Supervisor trackers, candidate summaries, and your own chain map are claims about the text. When one disagrees with the text, the text wins and the disagreement is itself a finding.
2. Exact terms are the chain. Alignment work at Walden, Capella, and elsewhere rests on the same point: problem, purpose, questions, methods, and conclusions carry the same words for population, variables, and phenomenon. A synonym is not a style choice in a thesis, it is a possible second construct. Flag it and let the candidate prove it is the same one.
3. Claim strength never exceeds design strength. Correlational, cross-sectional, observational, and descriptive designs support association and description. Causal wording past that line is a method error. The fix is always one of two honest moves: weaken the claim or strengthen the design, and the report says which was chosen.
4. Every conclusion answers a question, every question earns a conclusion. Count both sets and match them in order, including questions whose results were null or inconclusive. An unanswered question and an unasked conclusion are both major findings.
5. No new material at the end. Conclusions introduce no data, no literature, and no claims that have no chain behind them. Findings that appear first in a conclusion get moved back into the chain or cut.
6. Institutional rules come from the institution. Readiness items are checked against the regulations and submission pages the candidate supplies for their own university, accessed during the audit. Another institution's numbers never appear in the report as if they applied.
7. Audit the abstract last. The abstract is the most retold version of the study and the most common place for the promised study and the run study to differ. Check it against the finished, fixed chain, never against chapter 1.
8. Do not rewrite the thesis inside an audit. Report only mode stops at findings and the fix order. Fixes mode applies the agreed fixes and proves them by re-run. Either way, authorship decisions (which term wins, whether a question is dropped or restored) belong to the candidate and the supervisor, and the report records who decided.
9. Separate the three verdicts. Coherence, feedback state, and readiness are reported as three results. A thesis can pass any one and fail the others, and merging them into one grade hides the failure that matters.
10. Report counts, files, and exit codes, not adjectives. Findings by severity, questions checked, items overdue, links not checked, and the commands rerun. A reader with the audit folder should be able to reproduce every number in the report.