# Case Study Zero - From Document Chaos to Court-Leveraged Intelligence **Matter type:** Unlawful detainer + housing discrimination **Deployment environment:** Active litigation **Operator:** Self-represented tenant **Outcome:** Opposing party dismissed without prejudice --- ## The One Thing That Matters First Before deployment, the landlord's own documents contained three irreconcilable versions of the same rent figure across different notices and ledger records. The system found this automatically. The operator didn't search for it. It surfaced as a contradiction flag during indexing. That is what this system does. --- ## System Load | Metric | Value | |--------|-------| | Documents under management | 199 verified files | | Searchable text segments | 1,181 indexed chunks | | Exact duplicates identified and quarantined | 23 | | Document types | PDFs, emails, screenshots, ledgers, court filings, agency correspondence, payment records | | SHA-256 fingerprints | Every file, full chain of custody | | AI interfaces active | 2 (Claude Desktop via MCP, ChatGPT via FastAPI) | --- ## The Problem Before Deployment This was not a disorganization problem. Disorganization is survivable. This was a **source-of-truth collapse** under time pressure. The matter had: - Multiple document types arriving at irregular intervals from different sources - Three versions of the same rent figure across the landlord's own notices and ledger - Payment records showing acceptance, then reversal, with no corresponding credit in the notice amounts - Government agency correspondence with hard deadlines buried in multi-page PDF attachments - Court filings, minute orders, and hearing records with no cross-reference to the underlying documents they cited - Critical dates scattered across a dozen sources with no reliable chronology The practical risk was not losing the evidence. It was being unable to locate, sequence, or cross-reference it under pressure - which is operationally identical to not having it at all. --- ## Contradiction Proof - What the System Caught ### Contradiction 1: Three Irreconcilable Rent Figures The landlord's own documents produced three different numbers for the same alleged obligation across a three-day window: | Document | Figure | Date | |----------|--------|------| | Three-day notice | Amount A | Filed date | | Resident ledger printout | Amount B (differs from A) | Same period | | Portal balance screenshot | Amount C (differs from both) | Same period | **Impact:** A landlord's predicate notice grounded in a figure that his own ledger contradicts is a notice with a foundation problem. This contradiction did not require expert testimony. It required the documents to be placed next to each other. The system placed them next to each other automatically. ### Contradiction 2: Payment Accepted, Then Reversed, Notice Ignores Both The ledger showed a January payment accepted on intake, then reversed several days later. The subsequent notice amount failed to credit the accepted payment during the window it was held and also failed to explain the reversal. **Impact:** The accounting narrative required to support a default claim became internally inconsistent on the landlord's own records. Not a tenant's argument. The landlord's own documents. ### Contradiction 3: Notice Amount Does Not Reconcile to Any Ledger Entry After applying credits and reversals, the amount stated in the operative notice could not be derived from any arithmetic path through the ledger entries. The numbers simply did not add up. **Impact:** Creates a documented predicate failure independent of any tenant-side argument. --- ## Query Demonstration - What the System Returned All queries ran against the indexed vault. No documents were manually attached. --- **Query:** "Show all payments made in January and whether each was credited in the notices" **System returned:** - Payment records extracted from portal screenshots and bank records - Corresponding notice entries cross-referenced by date range - Mismatch between amount credited and amount held during reversal window flagged automatically --- **Query:** "What is the sequence of rent figures across all landlord-produced documents?" **System returned:** - Chronological listing of every rent figure appearing in landlord-produced records - Source document and date for each figure - Three non-reconciling values identified, source-cited --- **Query:** "What did the landlord's property manager communicate to the rental assistance program and when?" **System returned:** - All correspondence between property management and assistance program extracted - Timeline mapped across 73-day window - Specific dates of acknowledgment, silence, and subsequent adverse action sequenced --- **Query:** "Show the timeline from rental assistance application to eviction notice" **System returned:** - Application acknowledgment date - Property manager last contact date - Gap period with no cooperation activity - Eviction notice date - Days elapsed at each stage --- ## Before / After Delta | Dimension | Before | After | |-----------|--------|-------| | Document retrieval | Manual search across fragmented folders | Sub-second full-text query | | Contradiction tracking | None - relied on memory and luck | Automated cross-document flagging | | Timeline | Mental model, no source citations | Extracted, sequenced, sourced to document | | Exhibit readiness | Ad hoc assembly under deadline | Indexed, numbered, source-linked on demand | | AI reliability | Hallucination-prone without structure | Consistent sourced answers from indexed vault | | Duplicate risk | Unknown - no deduplication | 23 exact duplicates identified and quarantined | --- ## What This Made Possible The system made self-representation viable under active litigation conditions that included: - Emergency hospitalization mid-case - Court hearings on compressed timelines - Motion practice against represented opposing counsel - Simultaneous administrative proceedings with independent agencies - Attorney review requests that needed fast turnaround At no point did the operator lose track of where a document was, when an event occurred, or what the landlord's own records said. --- ## Strategic Output - Attorneys' Observations During the matter, several attorneys reviewed portions of the document record as part of potential representation conversations. Their observations, unsolicited: - The exhibit index and classification structure reduced their intake review time compared to unstructured file dumps - The timeline presentation made the chronology legible without reading every document - The contradiction map surfaced discrepancies they confirmed were legally significant - The missing-record tracker identified gaps they would have needed to discover independently None of them were told the system had found these things automatically. They reviewed the outputs and reached those conclusions themselves. --- ## System Architecture (Brief) Four layers: **Control layer** - Local filesystem + SQLite as canonical source of truth. Every file fingerprinted on intake. Classification states assigned (canonical, copy-exact, superseded). Real-time repository monitor watching for changes, new arrivals, and duplicate ingestion. **Storage layer** - Verified documents mirrored to private remote bucket. Large documents split into retrievable parts, tagged back to parent records. **Index layer** - Full-text search across 1,181 chunks via SQLite FTS5. Sub-second retrieval without requiring AI. **Access layer** - Claude Desktop via local MCP server. Custom ChatGPT interface via FastAPI endpoint. Both query the same index and return consistent results. --- ## What This Is Not This system does not provide legal advice. It does not make arguments. It does not predict outcomes. It makes the documents navigable, the contradictions visible, and the record queryable. What happens with that information is the attorney's job. --- ## This System Is Now Available for Deployment The architecture demonstrated here is deployable on client-controlled infrastructure for document-heavy matters. **Matter types this serves well:** - Tenant defense - UD, FHA, FEHA - Source-of-income discrimination - Administrative benefit disputes (APLA, HOPWA, Section 8, HUD) - Disability accommodation records - Document-heavy regulatory matters - Any matter where the other side's own documents contradict each other **The question to ask:** Does your matter have more than 50 documents, conflicting records, or a timeline that lives in your head? If yes - this is built for you. → **[Try the interactive demo - query the vault yourself](https://mandamus.pro/demo.html)** → [Founding Operator Program](CLIENT-DEPLOYMENT.md) → [Deployment scope and pricing](CLIENT-DEPLOYMENT.md) → [Technical architecture](ARCHITECTURE.md) --- > **One line:** mandamus.pro turned a live, document-heavy legal matter into an AI-operable evidence command system - and the landlord's own documents dismantled his case.