DossierMap builds a source-grounded evidence map before asking any model to reason over criteria. Uploaded PDFs are checked, indexed into line- or page-addressed source components, compressed into document-specific evidence units, and only then sent through a gated citation proposal step. Every proposed citation still has to pass deterministic verification before it can affect a gap report or matrix export.
Overview
The current pipeline
Each analysis run follows the same source-grounded sequence. Most stages are deterministic. Provider calls are explicitly gated and receive compact, source-addressed evidence context rather than raw PDF files or unrelated workspace data.
1
Criteria loaded from matrix template
The full set of promotion criteria rows is read from the institution-specific, track-specific, version-controlled matrix template before any document analysis begins.
2
PDF preflight and validation
Uploaded PDFs are checked for file type, readability, encryption, scope requirements, line-number quality, page references, and extraction readiness. Failed preflight blocks the run and does not consume an analysis credit.
3
Evidence indexed by source component
Documents in the selected scope are parsed into source components with document hashes, page spans, verified line records when available, text hashes, provenance, and component type. Documents outside the selected scope are reported as ignored and cannot support final citations.
4
Source-specific evidence prepared
CV entries, cover-letter claims, claim-support hints, supplemental page/table/figure units, and prompt-injection screening artifacts are prepared before criteria matching. This lets the model reason over the right kind of evidence instead of treating every document as a generic text block.
5
Retrieval-ranked citation proposals generated
The configured proposal provider receives official criteria rows plus a relevance-ranked evidence pack for each criteria section or batch. The preferred network path sends selected chunks, fallback chunks for under-supplied rows, and compact line/page inventories, not the entire workspace.
6
Deterministic citation verification
Every proposed citation is checked against the original evidence index. A citation is accepted only if its quote, document type, document hash, matrix item, page, and line range are grounded to the uploaded source material and selected analysis scope.
7
Gap report and synthesis generated
Each criteria row receives a status based on verified citations. The run also records multi-document synthesis, false-negative audit planning, and analysis quality readiness so reviewers can see which rows need more human attention.
8
Run audit and quality checks
The run records quality readiness, prompt-screen findings, false-negative audit planning, visual-review warnings, letter-claim risks, and other review flags without changing the deterministic evidence artifacts.
9
Human review of possible matches
Verified possible matches, unsupported attempts, not-found rows, and visual-review warnings are presented for faculty judgment. Faculty approvals, exclusions, manual evidence, and no-support acknowledgements are stored as a separate decision layer over the immutable run artifacts.
10
Matrix export
Only verifier-accepted citations are eligible for export. Strong matches are included by default, accepted possible matches and manual evidence may be added, and faculty-excluded citations are left out.
Step 1
Criteria loading
DossierMap reads criteria rows from the selected Personalized Promotions Matrix template for your institution, school, track, rank, and criteria version. Criteria are loaded before any document is uploaded or analyzed.
The criteria version date is displayed on every gap report and matrix export. DossierMap does not use generic academic promotion criteria or infer criteria from document content.
The criteria set is fixed at the start of each run. It cannot change during analysis, and the AI proposal step cannot expand or reinterpret criteria rows beyond what the template specifies.
Step 2
PDF preflight
Before any analysis begins, each uploaded file is checked against the requirements for its document type. The preflight step rejects files that would produce unreliable results.
CV and cover letter PDFs must have visible line numbers. Supplemental materials must have stable page references. All files must be unencrypted, readable, and in PDF format. Files that fail preflight are returned with a specific reason, are not sent to provider-backed matching, and do not consume an analysis credit.
All uploads
PDF format only.
Readable, unencrypted, and not password protected.
Text must be extractable enough to support verification.
The selected analysis scope must include every document type required for the run.
Files outside the selected scope are not analyzed for that run.
CV
Visible sequential line numbers are required.
Line numbers must be dense enough for the verifier to map citations reliably.
The CV should be the current academic CV intended for promotion review.
Scanned or image-only CVs require OCR before upload.
Cover letter
Visible sequential line numbers are required.
The letter should be uploaded separately from the CV and supplemental packet.
Letter citations are treated as narrative claims, not independent verification.
Summaries or paraphrases are not accepted as quotes.
Supplemental materials
Stable page references are required.
Visible line numbers are optional and used only when they are reliable.
Exhibits should be paginated and grouped in the order faculty will review them.
Image-only pages may need OCR or manual review before they can support quoted evidence.
Step 3
Evidence indexing
Each in-scope document is parsed into source components: discrete, addressable segments of evidence with source locations, document hashes, text hashes, provenance, and component type. A CV component might be a dated career entry with verified lines 84 through 91. A supplemental component might be a page-backed exhibit or a line-backed page segment when visible line numbers are reliable.
CVs are grouped around detected academic career entries. Cover letters are grouped around narrative claim blocks. Supplemental packets use page, table, figure, exhibit, and line-backed components depending on what can be extracted reliably. The evidence index remains the source of truth for final verification.
Documents outside the selected analysis scope are reported as ignored documents. They are not silently used as support for the final report.
Step 4
Source preparation before matching
Before citation proposals, DossierMap builds a proposal-facing evidence index. This is a compact matching layer derived from the original evidence index. It helps the model reason over document-specific evidence while preserving the original source index as the verification authority.
Cover-letter paragraphs are converted into verified claim chunks when a cover letter is in scope. Claim-support hints can link those claims to CV or supplemental evidence, but the hints do not replace separate citations from the supporting source documents.
Supplemental materials are converted into evidence units for text pages, tables, figures, document excerpts, mixed visual pages, and review-only visual pages. The current pipeline classifies supplemental visual material locally and flags visual pages for review; it does not send supplemental images to an external vision provider in the normal run.
DossierMap also runs a prompt-injection screen over uploaded source material. Uploaded dossier text, tables, figures, links, and future visual summaries are treated as untrusted source content, not instructions to the system or model.
Document-specific mapping
How each source maps to the matrix
DossierMap uses one official criteria matrix, but it does not treat every uploaded document the same way. CVs, cover letters, and supplemental packets are prepared with different source logic because they carry different kinds of evidence. The matrix row is the merge point: a single criterion can receive support from the CV column, the cover-letter column, the supplemental-materials column, or all three.
The important design choice is that the model does not receive a raw, unstructured PDF and simply answer from memory. DossierMap first builds a local evidence index with source locations, document hashes, page spans, line records when available, component metadata, and extraction provenance. Model prompts operate over that indexed source material, and final citations still have to prove back to the original evidence index.
CV
Career-record evidence
The CV is mapped as a structured record of activities, roles, dates, products, awards, committees, grants, publications, teaching, clinical work, service, and other career entries. The CV process is deliberately line-first because the matrix export needs source locations that faculty can inspect.
File and readability checks. The CV must be a readable, unencrypted PDF with extractable text. Scanned or image-only CVs are blocked until OCR is performed because there would be no dependable source text to verify.
Workload limits. The CV is checked against page and text-volume limits before indexing. This prevents unusually large or malformed files from producing unstable extraction or runaway provider costs.
Line-number validation. DossierMap looks for visible line numbers using plain text, layout-preserving text, and visual-margin extraction. It then checks line-number density, order, missing numbers, reversals, restarts, large jumps, and year-like false positives. A CV with years that look like line numbers should not pass by accident.
Line record creation. Once a line-number source is selected, each usable line becomes a record with a visible line number, page, extracted text, position, and source-line index. These line records become the citation authority for the CV.
Career-entry chunking. The CV is not sliced into arbitrary token windows. The indexer groups line records around known section headings, heading-like all-caps sections, dated entries, and continuation lines. A publication list, award entry, committee role, grant, appointment, or teaching activity should remain together as a career-record component.
Component metadata. Each CV component receives a deterministic component ID, component type, section heading, title, date range when detected, line count, document hash, page span, line span, text hash, and provenance marker. This metadata travels with the chunk into later matching.
Criteria matching. The matching prompt asks the model to identify the concrete academic activity, role, product, date, or outcome in the CV entry before choosing a matrix row. A topically related CV line is not enough; the quote has to directly support the criterion.
Verification and export. The verifier accepts only CV citations whose quote, page, line range, document name, document hash, scope, and matrix item prove back to indexed CV line records. Accepted CV locations write only to the CV column.
What this means for users: a CV is strongest for proving that an activity, product, role, date, or appointment exists. It may be weaker for explaining impact, quality, scope, or significance unless those details are written directly in the CV.
Cover letter
Narrative claim evidence
The cover letter is mapped as the faculty member's narrative argument. It is not treated as another CV chronology.
Line-backed preflight. Like the CV, the cover letter must be readable and line-numbered. This lets DossierMap cite narrative claims precisely instead of citing an entire page or paragraph loosely.
Claim block creation. The letter indexer groups lines into claim blocks, preserving sentence and paragraph boundaries where possible and separating headers, salutations, transitions, and closings from substantive claims.
Claim extraction. When the Haiku claim-extraction path is enabled, a low-cost model pass extracts discrete claims from the verified letter components. The prompt explicitly says to extract claims only, not to map them to matrix criteria.
Support hints. DossierMap can link letter claims to related CV or supplemental evidence by text overlap or matching terms. These links are hints for review and later matching, not proof by themselves.
Matrix matching. The main matrix prompt uses verified letter-claim chunks when available, so the model reasons over concise claims instead of repeatedly rereading the whole letter for every row.
Verification and export. A cover-letter citation still has to quote exact letter text from exact lines in the original evidence index. Accepted letter locations write only to the cover-letter column.
What this means for users: a letter is strongest for explaining significance, leadership, independence, growth, philosophy, or impact. Letter-only claims are review-sensitive because they may be self-attested unless corroborated by CV or supplemental evidence.
Supplemental materials
Exhibit and corroboration evidence
Supplemental packets are mapped as exhibits: pages, tables, figures, screenshots, certificates, appendices, letters, or other supporting artifacts.
Page-stable preflight. Supplemental packets must be readable PDFs with stable page references. Line numbers are optional because many exhibits are tables, figures, screenshots, certificates, or appended documents.
Line use when reliable. If supplemental line numbers are detected and pass usability checks, DossierMap can use line-backed supplemental components. If they are missing or inconsistent, supplemental evidence falls back to page references.
Evidence-unit classification. The supplemental-unit step classifies material as text pages, tables, figures, mixed table/figure pages, document excerpts, or visual-only pages. The goal is to preserve the exhibit's role instead of pretending every page is ordinary prose.
Visual-review flags. Visual-only or figure-heavy pages are flagged for human review. The normal run does not automatically send supplemental images to an external vision provider, and visual-only placeholder text should not become a strong matrix citation.
Gated vision eligibility. A separate Haiku vision path can be configured for sparse visual pages, figures, or mixed table/figure pages, but it requires explicit approval, strict caps, and a dedicated network gate. Its allowed output is a page-anchored visual summary for human review, not an automatic matrix citation.
Corroboration matching. Supplemental units often corroborate a CV entry or cover-letter claim. The matching prompt may use them as supporting evidence for a row, but it must cite the supplemental source separately from any CV or letter citation.
Verification and export. The verifier accepts supplemental line citations only when verified line records exist. Otherwise it accepts page-backed citations. Accepted supplemental locations write only to the supplemental-materials column.
What this means for users: supplemental materials are strongest for corroborating outcomes, evaluations, quantities, screenshots, awards, and documents that do not fit naturally in the CV or letter. Visual-only material can guide review, but it should be checked by a person before relying on it.
What a CV component contains
A CV component is the unit DossierMap asks a model to reason over. It contains the source text, but it also contains enough structure to keep the model grounded.
Source identity: run ID, analysis scope, document name, document type, document hash, page span, and text hash.
Line authority: visible line start, visible line end, and individual line records with line number, page, and line text.
Career context: component type, section heading, title, detected date range, line count, and the CV career-entry strategy.
Provenance: extraction method and source kind, so later stages know whether the component came from a CV entry, letter claim block, supplemental page, or supplemental unit.
This is why DossierMap can show users exactly where a proposed citation came from and why impossible line references can be rejected automatically.
Matching is criteria-batched, not document-batched. DossierMap loads a set of official matrix rows, gathers the most relevant CV entries, letter claims, and supplemental units for those rows, asks for proposed source locations, and then verifies each proposed citation against the original upload. The document type determines the export column; the matrix item determines the row.
Model prompting
Different models are used for different jobs
DossierMap does not use one generic prompt for the whole dossier. The current architecture separates lower-cost extraction, broad matrix matching, and skeptical second-reader review. Each model role receives a different prompt, different evidence payload, and different output schema.
Haiku
Extraction and preprocessing
Haiku is used for narrow, lower-cost extraction and preprocessing tasks when explicitly enabled. The main current use is cover-letter claim extraction, with a separate gated policy for possible supplemental-vision summaries.
The prompt says: extract claims only; do not map claims to matrix criteria.
The input is verified letter claim-candidate components, not the whole workspace.
The output is a claim artifact with line ranges, claim type, supporting quote, confidence, and parse rationale.
For supplemental vision, the allowed input is capped visual units selected by policy; the allowed output is a page-anchored summary for human review.
The result is an intermediate artifact for later matching. It does not approve evidence and does not write to the matrix.
Sonnet
Retrieval-first matrix matching
Sonnet is the preferred model for the broad citation proposal pass. It receives official criteria rows plus a retrieval-ranked evidence pack.
The prompt is criteria-batched by matrix section or section slice.
For each batch, local retrieval selects row-specific chunks, domain-broad chunks, fallback chunks for under-supplied rows, and evenly spaced audit chunks.
The model must return structured citations and one row assessment for every supplied criteria row.
The prompt emphasizes high precision: cite direct criterion support, mark merely related evidence as insufficient, and keep possible matches sparse.
Exact quotes are repaired back to indexed source lines locally when possible, then deterministic verification makes the final accept/reject decision.
Opus
Second-reader audit
Opus is reserved for more expensive, skeptical review after the first pass. The normal run creates a privacy-safe audit plan by default; a live audit requires explicit network approval and provider configuration.
The prompt targets only high-risk rows: not-found rows, weak support, unsupported attempts, verifier rejections, and retrieval-under-supplied rows.
The prompt tells the model to treat the prior pass as fallible and search for missed direct support.
It must return only newly found citation candidates for supplied high-risk rows, not restate already verified citations.
Audit proposals still go through the same deterministic citation verification before they can affect any review workflow.
The model names may change over time as providers release new versions, but the role separation is the important part: extraction prompts, matrix-matching prompts, and audit prompts are different tasks with different evidence boundaries and different verification expectations.
Step 5
AI citation proposals
The main AI-assisted step is citation proposal generation. In local demo modes, this may use fixture data or a limited deterministic baseline without a network call. In network-enabled runs, DossierMap uses the configured provider only when the required environment gates and API key are present.
The preferred provider path is retrieval-first. DossierMap sends one structured request per criteria section or section batch. Each request includes the official criteria rows for that batch plus relevance-ranked evidence chunks, domain-broad chunks, fallback chunks for under-supplied rows, and compact line/page inventories. The request does not include unrelated workspace data, payment data, admin logs, or documents outside the selected analysis scope.
The model must return a structured response for every supplied criteria row: a row assessment status, proposed citations with exact quoted evidence, document type, page or line references, evidence strength, and a short rationale. Rows without direct support should be marked not found or insufficient evidence rather than forced into a citation.
For CV evidence, the prompt tells the model to work from career-record components. It must identify the concrete activity, role, product, date, or outcome recorded in the CV before proposing a matrix row. For cover-letter evidence, it must reason from narrative claims and avoid treating self-attested explanation as independent proof. For supplemental evidence, it must treat pages, tables, figures, and exhibits as corroborating support and avoid inventing line numbers or table coordinates.
The prompt also includes a document evidence contract. That contract tells the model which document types are available, which official source column each document type can fill, which reference mode should be used, and which citations require exact line ranges. When several document types support one matrix row, the model must return separate citations rather than blending the evidence into a single summary.
There are two strength classifications a proposed citation can receive:
A citation classified as a strong match means the quoted evidence directly and specifically supports the matrix criterion. A citation classified as a possible match means the quote is relevant but requires faculty judgment or corroboration before it should be used in the matrix.
The AI proposes. It does not decide. Every proposal passes through deterministic verification in the next step before it can influence the gap report or matrix export.
Step 6
Deterministic verification
Every proposed citation is checked against the original evidence index, not the compact proposal-facing version. The verifier confirms that the quoted text appears at the referenced source location and that the citation uses a valid matrix item, document type, document hash, analysis scope, page, and line range.
CV and cover-letter citations require verified line ranges. Supplemental citations use verified lines only when the original supplemental component has line records; otherwise they must cite a stable page. The verifier rejects invented line numbers, unsupported quotes, wrong document names, wrong matrix rows, out-of-scope documents, deleted sources, and structurally invalid citations.
Rejected citations are logged in the run audit trail and are visible in the gap report. They do not silently disappear, and they cannot reappear in the matrix export.
The verification step exists specifically to catch plausible-sounding but fabricated, mislocated, or overbroad model output. No ungrounded citation reaches the output.
Step 7
Gap report statuses
After verification, each criteria row in the gap report receives one of four user-facing statuses based solely on verifier output.
Verified
At least one verifier-accepted strong-match citation exists for this row. The evidence is grounded to a real source location and directly supports the criterion.
Weak support
At least one verifier-accepted possible-match citation exists, but no strong-match citation. The evidence is grounded to a real source location but requires faculty judgment before it should appear in the matrix.
Unsupported
The AI proposed citations for this row, but the deterministic verifier rejected all of them. The proposals were not grounded to real source locations in the uploaded documents.
Not found
No verified citation is available for this criterion in the selected analysis scope. This does not mean the candidate failed the criterion; it means DossierMap has not verified evidence for it in the uploaded materials.
The distinction between unsupported and not found matters. Unsupported means a citation was attempted but deterministic verification rejected it. Not found means no verified citation is available in the selected analysis scope. A not-found row may reflect a genuine gap, incomplete supplied evidence, an out-of-scope document type, or a source format the pipeline cannot read yet.
Step 8
Run audit and quality checks
The run manifest ties every completed analysis to its deterministic artifacts: evidence index, letter claims, supplemental units, prompt-injection screen, citation proposals, citation verification, gap report, analysis quality audit, and matrix export.
The quality audit does not approve matrix content. It summarizes what needs human attention before relying on an export: missed-evidence risk, weak or unsupported rows, letter-only self-attested claims, supplemental pages that need visual review, broad CV line ranges, duplicate source ranges, and prompt-screen findings.
DossierMap also creates a false-negative audit plan for higher-risk rows, such as not-found rows, weak support, rejected citation attempts, and under-supplied retrieval packs. By default this audit plan is privacy-safe and does not send dossier text to another model.
Step 9
Human review gate
Verified citations classified as possible matches are held in a dedicated review screen before they can reach the matrix export. Faculty see the quoted evidence, source location, and rationale, then make an explicit decision to accept or exclude it.
Faculty decisions are stored separately from the immutable run artifacts. Review actions can include approving possible matches, excluding verified citations, adding manual evidence, or acknowledging that a no-support row was reviewed. These decisions change the draft export layer; they do not rewrite preflight, evidence index, verifier, or gap report artifacts.
Possible matches do not populate matrix source-location cells by default. The default state is conservative. Faculty must take an action to include a possible match in the export.
Step 10
Matrix export rules
The matrix export writes verified citations to the correct source-location columns in the selected Personalized Promotions Matrix template. Each citation's document type determines its column: CV evidence writes to the CV column, cover letter evidence to the cover letter column, and supplemental evidence to the supplemental materials column.
By default, only strong-match citations that passed verification are written into export cells. Possible matches accepted during the human review step may be included. Faculty-excluded citations are omitted, and reviewer-added manual evidence belongs to the review decision layer.
The exported matrix is editable. Faculty and mentors retain full control over the final submission document.
DossierMap does not claim the exported matrix is complete, submission-ready, or institution-approved. The export is a draft starting point for faculty and mentor review.
Data sent for matching
Provider data boundary
When a network provider is enabled, DossierMap sends the evidence needed for citation proposal: official criteria rows, selected source components or evidence units, exact source text needed for matching, document type, page and line metadata, line inventory summaries, and prompt instructions that require structured output.
DossierMap does not send billing records, payment card data, admin dashboards, unrelated files, private activity logs, or documents outside the selected analysis scope as part of citation proposal requests.
Admin-visible telemetry is limited to operational and billing-safe fields such as provider name, model key, token counts, cost estimates, run sequence, readiness status, and aggregate counts. It must not include uploaded file names, extracted dossier text, citation quotes, line references, citation IDs, or generated matrix contents.
Disclaimer
Assistant, not authority
DossierMap uses AI and deterministic verification to help identify possible evidence matches. It can miss relevant material, misclassify support, or surface evidence that still requires academic judgment.
The faculty member is always responsible for reviewing the source text, verifying accuracy, correcting errors, and deciding what belongs in a promotion submission. DossierMap is an assistant for organizing and checking evidence; it is not a substitute for the faculty member, mentor review, faculty affairs guidance, or committee judgment.
Before downloading a matrix export, users must acknowledge that the exported file is a draft aid and that they are responsible for validating all content before use.
Privacy
Privacy and data handling
Uploaded documents may be processed locally and, when network providers are enabled, excerpted source text may be sent to the configured AI model provider for citation proposal. DossierMap sends source-addressed evidence context for matching, not the whole workspace or unrelated account data.
DossierMap does not intentionally use uploaded documents for model training. Provider processing, retention, and deletion practices depend on the deployed provider and any written customer agreement. DossierMap's own retention controls, account data handling, and deletion practices are described in the Privacy Policy.
If you have questions about whether DossierMap is appropriate for your institution's data policies, consult your faculty affairs office or institutional privacy officer before uploading real dossier materials. See the Privacy Policy or contact DossierMap support for full details.
Limitations
What DossierMap does not do
DossierMap maps evidence to criteria. It does not score candidates, predict promotion outcomes, compare candidates to each other, or generate written content for your dossier.
The gap report reflects what is verifiable in the materials you upload in the scope you select. A not-found status does not mean the criterion is unmet; it means verified evidence was not found in the uploaded documents. Evidence in documents outside the selected scope, deleted documents, image-only pages without readable text, or formats the tool cannot parse will not appear as verified support.
AI citation proposals can be wrong, and retrieval can miss relevant material. The deterministic verification step is designed to catch ungrounded proposals, but verification confirms source grounding, not academic merit. Whether a cited passage genuinely satisfies a promotion criterion is an academic judgment that belongs to the faculty member, their mentor, and the promotion committee.
DossierMap is an organizational and verification aid. It does not replace mentor review, institutional guidance, or committee judgment at any stage of the promotion process.
Ready to build a verified evidence map?
Upload your line-numbered PDFs, select your criteria set, and review source-grounded findings before exporting.