Assessment Rules & Methodology

A composite score from 0-100 that represents the overall trustworthiness of the document based on all analysis modules.

A unique cryptographic fingerprint of the uploaded file, generated using the SHA-256 algorithm.

The AI's certainty in a classification result, expressed as a percentage.

Determines whether the document primarily serves the vendor's interests (upselling, lock-in) or empowers the reader to act independently.

Measures whether the document frames problems as needing external help (provider-favored) or as solvable by the reader's team (consumer-favored).

Measures dependency creation on specific vendors, proprietary tools, or locked-in methodologies.

Whether the reader can execute the recommendations independently without hiring the document's author.

Presence of 'Phase 2' mentions, premium tiers, 'contact us for more', or expansion opportunities that suggest revenue generation motive.

Who bears the implementation risk — the vendor (with SLAs/guarantees) or the client?

Detects the likely size of the organization that created the document — Solo/Boutique, Mid-tier, or Big 4/GSI.

Whether the document uses generic frameworks (SWOT, Porter's) or branded proprietary IP with ©/®/™ marks.

Whether data comes from desk research and secondary citations (smaller firms) or primary global surveys and benchmarking (GSI resources).

Whether the document appears template-based or has studio-grade branding with professional design elements.

Whether the document covers niche specialisation or end-to-end transformation capability.

Amount of disclaimer, NDA, liability, and legal language in the document.

The intended organizational audience — Startup (agile/MVP), SME (balanced), or Enterprise (governance-heavy).

Whether the document mentions small teams (startup) or steering committees and board approvals (enterprise).

Whether recommendations affect a single department or require enterprise-wide coordination.

Whether the document assumes greenfield (build new) or requires legacy system modernization/migration.

Whether costs suggest SaaS subscriptions (small scale) or multi-year capital transformation programs (enterprise).

Whether security mentions are basic or reference enterprise frameworks (SOC2, GDPR, ISO 27001).

Identifies who should read this — Developer, Manager, VP, or C-Suite — based on language and decision scope.

Balance between high-level strategic vision and hands-on implementation details.

Presence of financial language like EBITDA, NPV, CapEx, ROI, TCO that targets financial decision-makers.

Amount of technical terminology (APIs, Docker, Python, microservices) vs. business language.

Whether recommended actions are immediate ('do this today') or long-term ('3-year roadmap').

Whether decisions are tool-level (manager) or business-model-level (C-Suite).

Measures content uniqueness: Commodity (conventional wisdom), Differentiated (some original value), or Category-Defining (truly novel).

Whether data comes from original research (surveys, experiments) or aggregated secondary sources.

Whether the document challenges conventional wisdom or simply aligns with current hype.

Whether the document uses standard frameworks (SWOT, Porter's Five Forces) or introduces original models.

Whether predictions are vague ('AI will grow') or specific and falsifiable ('GPU costs will drop 40% by Q3 2026').

Whether examples use named, verifiable companies or anonymous placeholders ('a large bank').

Combined score (0-100) measuring the density and severity of manipulative language techniques in the document.

Accountability-destroying phrases like 'arguably', 'up to', 'may contribute', 'potentially' that make claims unfalsifiable.

Impressive-sounding metrics without baselines — e.g., '300% growth' without mentioning it went from 1 to 4.

Artificial time pressure like 'window is closing', 'act now before competitors', 'critical window of opportunity'.

Complex buzzword phrases that conceal simple ideas — e.g., 'synergistic paradigm shift' instead of 'working together differently'.

Number of logical fallacies detected per 1,000 words of text.

Attacking a weakened version of an opponent's argument rather than the actual position.

Binary framing like 'Adopt AI or go bankrupt' that eliminates nuanced middle-ground options.

Using 'As Gartner says...' or similar authority references without independent evidence.

Assuming causation from correlation — 'After implementing X, revenue grew 40%' doesn't prove X caused the growth.

Arguments to continue projects solely because of prior investment rather than future value.

Combined score (0-100) measuring the ratio of filler content to substantive information.

Estimates the years of formal education needed to understand the text on first reading.

The ratio of descriptive words (adjectives/adverbs) to action words (verbs) in the document.

Marketing-oriented words like 'synergy', 'leverage', 'cutting-edge', 'revolutionary' that add perceived sophistication without meaning.

Concrete action words like 'implement', 'deploy', 'configure', 'measure', 'test' that indicate practical content.

Count of concrete, verifiable numbers, percentages, dates, and statistics in the document.

Score (1-10) measuring the gap between conceptual description and executable implementation.

Checks for executable artifacts: code snippets, configuration files, checklists, architecture diagrams, templates, API definitions.

Whether specific roles and skills are defined (e.g., 'need 1 Data Engineer with 3+ years') vs. vague references.

Whether timelines include specific dates and durations or use vague references like 'future state'.

Clear statement of starting requirements (e.g., 'requires clean data lake and API gateway') vs. no mention.

Score (0-100) indicating how likely the document's recommendations will become outdated in the near term.

Technologies mentioned that are no longer current best practice.

Important current technologies and practices that the document doesn't mention.

Optimism score (0-100) measuring the balance between promotional language and honest risk acknowledgment.

Percentage of the document's language that is positive, promotional, or optimistic.

Count of references to risks, challenges, limitations, and potential failure modes.

Presence of sections or statements acknowledging where approaches might fail or have failed.

Combined score (0-100) assessing regulatory compliance awareness and ethical consideration in the document.

References to compliance frameworks like GDPR, EU AI Act, CCPA, SOC2, HIPAA, ISO 27001.

References to bias mitigation, fairness, transparency, explainability, and responsible AI practices.

Critical omissions like proposing data scraping without compliance or AI deployment without bias assessment.

Score (1-10) measuring the document's use of diagrams, charts, graphs, infographics, and visual elements.

Score (1-10) measuring the density of quantitative evidence: tables, citations, statistics.

Count of visual elements including charts, graphs, diagrams, and infographics in the document.

Count of academic-style citations and source references.

Count of tabular data references in the document.

Count of specific numerical claims, percentages, and financial figures.

Combined score (0-100) measuring the presence of cognitive biases that undermine document credibility.

Only presenting evidence that confirms a predetermined conclusion while ignoring contradicting data.

Only analyzing successful companies/projects while ignoring the many that failed using similar approaches.

Cherry-picking data, time periods, or examples that support a predetermined conclusion.

Over-weighting recent events and trends while ignoring historical patterns and cyclical behavior.

Over-reliance on expert opinions (Gartner, Forrester, Harvard) without independent empirical validation.

The most notable and surprising claims extracted from the document by AI analysis.

This finding contradicts conventional wisdom or current market consensus.

This finding includes specific numbers, percentages, or data that can be independently verified.

DocDetector uses a two-tier architecture: a heuristic engine for counting and pattern matching, plus Gemini 2.0 Flash AI for judgment, classification, and comprehension tasks.

Gemini analysis runs in 4 sequential layers where each layer's results are injected into the next layer's prompt, creating an increasingly informed analysis.

The standard scoring formula used across all classification modules: Score = Σ(rating × weight), where each driver has a predefined weight.

An individual analytical dimension within a module. Each classification module has 5 weighted drivers that contribute to its composite score.

A parameter controlling the randomness of Gemini AI's responses. Lower = more deterministic, higher = more creative.

Before analysis begins, Gemini performs a quick classification of the uploaded document to verify it falls within DocDetector's scope: technology vendor and advisory documents (consulting proposals, vendor pitches, training brochures, whitepapers, advisory decks) in the domains of AI, Data, Cloud, Digital Transformation, and Governance.

When Gemini AI is unavailable (no API key, rate limit, or API error), the system automatically falls back to a fully heuristic analysis engine.