vak-dissertation-review

name	vak-dissertation-review
description	VAK dissertation review: audit dissertation and autoreferat.

Рецензирование диссертации и автореферата (ВАК)

Methodology for reviewing a Russian dissertation + autoreferat ahead of a scientific-degree defense. The output is a set of рекомендательных (advisory) review documents — never a go/no-go verdict. You are an assisting reviewer / нормоконтролёр; the qualification decision belongs to the диссертационный совет and ВАК. The review includes нормоконтроль: formatting control and normative-compliance control.

This file is the map. Each references/ file holds the depth — read the relevant one when you reach that part of the work; don't load all of them up front.

• references/verification-discipline.md — the non-negotiable verification rules (read FIRST).
• references/review-dimensions.md — the full checklist: what to inspect, dimension by dimension.
• references/vak-norms.md — ПП РФ № 842, ГОСТ Р 7.0.11, ГОСТ Р 7.0.5, паспорт специальности, publication criteria. These change — verify the current redaction before relying.
• references/deliverables-and-pipeline.md — the 6-document deliverable set and the docx-js + MathType + table-borders build pipeline, with its gotchas.

What you produce

A clustered set of advisory docx review lists (one per dimension-cluster), each carrying severity labels and a proof under every note. The reference deliverable set is 6 documents:

1. Замечания к диссертации — formulas (render-checked), numbers, figures/tables, structure, references, numbering.
2. Замечания к автореферату — same dimensions for the автореферат + PDF/metadata passport + автореферат↔диссертация identity.
3. Новизна / ВАК-аудит — научная новизна (form + substance), паспорт-специальности fit, ПП 842 compliance, publication/patent registry, § степень разработанности (ДО → приращение).
4. Русский язык и плагиат — language proofread + borrowings (заимствования) + internal repeats.
5. Сводная рецензия — executive summary + достоверность table + priority recommendations.
6. Предложения по усилению работы — constructive, forward-looking "how to strengthen" (outside the defect review).

Tone throughout: рекомендательный. Phrase findings as "рекомендуется…", not "обязан/провал". Mark severity [КРИТ] / [ВАЖНО] / [КОСМ] (red / amber / green).

Non-negotiable discipline (the hard-won rules)

These exist because each was learned by getting it wrong once. Full detail + the war stories are in references/verification-discipline.md; the short form:

• Proof under every note. No finding ships without a verifiable anchor: страница, формула (N.M), a named норма (ГОСТ/ПП пункт), a DOI/URL, or an instrumental check. A note without a proof is a hypothesis, not a finding — verify it or cut it.
• Judge formula STRUCTURE by visual render, never by the OCR/text layer. Cyrillic + math text layers are garbled; misreading them produces false "errors". Render the page (pdftoppm) and look.
• No fabricated references. Every reference you propose (DOI, first source for a borrowing, a предшественник) must be one you actually retrieved this session. Target: zero fabrications.
• Verify arithmetic with a calculator/Wolfram, not by eye — units, conversions, percentages, trig. Cite the provenance triad for any computed value (formula/model · code path · inputs).
• Prior-art separates OWN from OTHERS'. A result already published by the author is legitimate priority (required for a defense), NOT disqualification; a result owned by others is. Always read the dissertation's own список литературы — predecessors are often cited there already.
• Adversarial review-loop at the boundary. Before merging any batch of findings into the documents, run an independent skeptic pass (a second model / agent) that tries to refute each finding and catch overclaim/wrong anchors. Merge only what survives.
• Cyrillic search via Bash grep, not PowerShell Select-String (Select-String mangles Cyrillic and em-dashes). pdftotext also mangles Cyrillic — grep Latin tokens only.

Workflow

1. Ingest. Extract the dissertation and autoreferat text (per chapter), and the список литературы, into plain-text working files. Locate the source PDFs for render checks.
2. Pass per dimension (see references/review-dimensions.md). For each finding, capture the anchor as you go — you cannot reconstruct proofs later.
3. Verify before writing. Render formulas, run arithmetic, confirm each reference exists, check each norm against its current text. Treat agent / external-model output as a hypothesis until you verify it.
4. Boundary review-loop. Send each finding batch through an independent adversarial gate; keep only confirmed, correctly-anchored items.
5. Assemble the docx via the generator pipeline (references/deliverables-and-pipeline.md): edit the canonical gen_*.js generators (never the output .docx), rebuild, add table borders, convert MathType placeholders to OLE for the formula-bearing lists, copy finals out.
6. Verify the build. Grep the extracted output text for the new markers (convergence by class-exhaustion: a class is closed when its grep count equals the expected count and stray classes are 0 — not after an arbitrary number of LLM passes). Confirm 0 unresolved @@ formula placeholders and the expected OLE object counts.
7. Reconcile and report. State what was covered, what stays out of scope (see below), and keep the tone advisory. When the current task has work-items/active/<slug>/, persist only the canonical VAK artifact there and return concise result/provenance for the root agent-runs.jsonl; do not duplicate it in .reports/ or .plans/. With no active work-item, a meaningful standalone VAK result MAY use one .reports/YYYY-MM/ summary; do not create a .plans/ duplicate.

Scope honesty

Some checks are principally outside a text/render/web review and need closed access or the author — say so plainly instead of faking them: a full Антиплагиат.ВУЗ run, re-running the work's own simulations or experiments, regenerating experimental data, full closed-database registries (РИНЦ/Scopus/WoS over every publication), and co-author contribution verification. For prior-art and reference existence, close them to the feasible limit (existence, completeness, open-source priority) and label the residual as ASSUMPTION (UNVERIFIED).

If the dissertation carries personal data, unpublished or embargoed results, or third-party confidential material, keep verbatim excerpts minimal — only as much as a finding's proof needs — and never copy them outside the user's own delivery channel (no external service, ticket, or log).