NainWrite ES — Fail-Closed Spanish Rewriting Workbench

NainWrite ES is a local, fail-closed rewriting agent for natural, clear and professional Spanish from Spain (es-ES). It rewrites drafts produced by people or language models while preserving meaning, uncertainty, conditions, negations, attribution and protected literals.

Interactive Online Demo: Try the editorial review workbench directly at write.naindev.com with sample contracts, technical proposals, diff analysis, and human approval gates without installing local models.

Current pipeline

1. Qwen3.8-27B produces a structured rewrite through local Ollama.
2. A deterministic validator compares protected and automatically detected invariants.
3. One constrained repair is allowed after deterministic drift.
4. Local LanguageTool checks spelling, grammar and typography. One constrained linguistic repair is allowed; it repeats the invariant and language checks.
5. Qwen and a separately configured Gemma auditor assess meaning and style. Both must pass before the result is accepted; otherwise it is review_required.

sequenceDiagram
    autonumber
    actor User as User / Client
    participant Orchestrator as NainWrite Engine
    participant LLM1 as Rewriter (Qwen 2.5/3.8)
    participant Invariant as Invariant Verifier
    participant LangTool as LanguageTool Engine
    participant LLM2 as Dual Auditor (Qwen + Gemma)

    User->>Orchestrator: Ingest Draft (PDF/MD/TXT) + Protected Invariants
    Orchestrator->>LLM1: Request Structured es-ES Rewrite
    LLM1-->>Orchestrator: Candidate Rewrite
    
    Orchestrator->>Invariant: Verify Invariants & Negations
    alt Invariant Drift Detected
        Invariant-->>Orchestrator: Drift Detected
        Orchestrator->>LLM1: 1x Constrained Repair Pass
        LLM1-->>Orchestrator: Repaired Candidate
    end

    Orchestrator->>LangTool: Grammar, Typography & Orthography Check
    alt Linguistic Faults Detected
        LangTool-->>Orchestrator: Rule Violations
        Orchestrator->>LLM1: 1x Constrained Linguistic Repair Pass
        LLM1-->>Orchestrator: Repaired Candidate
    end

    par Dual Auditor Consensus
        Orchestrator->>LLM1: Qwen Semantic Equivalence Audit
        Orchestrator->>LLM2: Gemma Independent Style & Invariance Audit
    end
    
    alt Both Auditors Pass
        LLM1-->>Orchestrator: Equivalence Verified
        LLM2-->>Orchestrator: Style & Integrity Verified
        Orchestrator-->>User: Status: ACCEPTED (Fail-Closed Validated)
    else Auditor Divergence or Flaw
        Orchestrator-->>User: Status: REVIEW_REQUIRED (Human Approval Gate)
    end

Loading

Automated acceptance is not proof of perfect equivalence. Legal, contractual, financial and other high-impact text requires human approval.

Current validation is preliminary: the corrected unassisted semantic bank exposes one false acceptance (SEMA-022), and some valid rewrites still trigger proofreading review. The strict acceptance gate has not passed. See measured validation results before using the agent.

The historical document baseline passed 165 software tests. Its known-control run detects all six deliberate meaning changes but requests review for one valid control because an auditor lists correct translations as critical changes. The Spanish smoke corrects the targeted calques and repetition; broader editorial acceptance remains open. The 40-case external semantic baseline exposes substantial disagreement with inherited corpus labels, including questionable translated positive pairs. It does not establish maximum quality or a better model. See independent evaluation and its separate human-review gate.

Local browser application

On the configured Windows workstation, double-click Start-NainWrite.cmd, or run:

pwsh -NoProfile -File tools/open-workbench.ps1

Open http://127.0.0.1:8765. Submit PDF, Markdown or TXT; review the extraction; edit and approve the exact Spanish; then review US and UK English. PDF, Markdown and a Markdown/images ZIP are available for each candidate. The single queue can pause and resume, with limited, intensive or optional night operation.

The launcher reuses installed local components. On another machine, install the pinned optional document dependencies and configure runtimes, OCR, LanguageTool and the existing model aliases first; see the runbook. It does not download components or contact a paid service during document work.

New model clients use dedicated Ollama at 127.0.0.1:11436; the shared Ollama service is unchanged. Historical jobs with a different endpoint identity cannot be silently migrated. The original short-text CLI and its quality limitations remain. Current implementation and verification are recorded in specification 011.

PDF reading is conservative and needs page-by-page confirmation, particularly for scans, columns, tables and footnotes. Exact PDF layout, DOCX, arbitrary structural editing and editing text embedded in images are not implemented. English can be reviewed in full and downloaded for manual editing; external edits do not receive automatic approval or silently replace retained versions.

The short-text API remains bounded. The document workflow segments UTF-8 Markdown/TXT, produces editable Spanish first and pauses for explicit revision approval. British and American English are then generated independently from the same approved Spanish, with additional original-source checks. Completed batches are checkpointed. Prompts above 12000 UTF-8 bytes and incomplete length-limited responses are rejected. See the active delivery handoff for verified versus pending work.

Rewrite a document

python tools/rewrite_document.py --input "C:/path/draft.md"
python tools/rewrite_document.py --resume "C:/path/draft.rewritten/job-id"

The command creates a unique job folder beside the original and stops with an editable es-ES copy and separate review. After reviewing the exact revision, use --resume JOB --approve-spanish SHA256 to authorise both English variants. Review-required Spanish additionally needs explicit --acknowledge-review; that never clears its findings. The original and previous revisions are retained. Later Spanish edits invalidate current approval and mark earlier translations outdated.

Exit 3 is a normal approval pause; 0 means three-version completion and automatic acceptance, 2 means completed candidates require review, and 1 means an error. Human approval is never automatic. Unit tests use synthetic approvals; actual document approval belongs to the user.

Use --clean-notes only when removing explicit VERIFY/DAVID preparation markers is intended; removed spans stay in local evidence. --omit-lines FIRST:LAST records an explicit preparation range. Neither option performs fact-checking or fills missing fields. This legacy CLI does not ingest PDF or DOCX; use the browser intake for PDF. HTML blocks and complex Markdown constructs remain unsupported. Editing is wording-only and rejects ambiguous span ownership or structural changes. See the document runbook.

Requirements

• Windows or another platform with Python 3.12 or later.
• A loopback Ollama service.
• Installed model qwen3.8-local-fast2:latest or an explicitly selected compatible model.
• Installed independent auditor gemma4:latest or an explicitly selected different model.
• Java 17 and a local LanguageTool server on 127.0.0.1:8082.

The Python core uses only the standard library. PDF/OCR/browser document delivery additionally uses requirements-documents.txt, portable Tesseract and installed local language data.

These are the measured workstation model aliases, not public download names. A new installation must resolve the original model package and verify its digest; do not assume that ollama pull can retrieve these custom aliases.

Rewrite one text

python tools/rewrite_text.py --text "El texto original necesita una redacción más clara." --plain

Protect exact literals by repeating --protected:

python tools/rewrite_text.py `
  --text "La entrega podría realizarse si Aitor aprueba ADR-007." `
  --protected "Aitor" `
  --protected "ADR-007"

Exit code 0 means every configured automated gate passed. Exit code 2 means human review is required. See the runbook for full operation and failure handling.

Verification

The software contract is validated by 207 automated tests across 30 test suites, executed on every push via GitHub Actions CI:

python -m unittest discover -s tests -v
python tools/validate_bank.py
python tools/validate_semantic_mutations.py
python tools/evaluate_corporate.py --bank benchmark/corporate_review_controls.json
python tools/evaluate_corporate.py --bank benchmark/corporate_calibration_controls.json
python tools/run_semantic_audit_benchmark.py
python tools/run_rewrite_benchmark.py --runs 2

Benchmark reports are written under artifacts/ and intentionally excluded from Git because they are generated and may contain supplied text.

The document path uses its own versioned corporate instructions and reproducible es-ES editorial checks. Its calibration must not be confused with the historical short-text API benchmark. Tests passing establish software behavior; draft synthetic model results do not guarantee flawless writing. The runbook explains resumption, version compatibility and review-required outputs.

Documentation

• Charter
• Requirements
• Architecture
• Architecture Decision Records (ADRs)
• Security Policy
• Traceability
• Runbook
• Primary-source research
• Measured validation results
• Active specifications

The first browser translation smoke exposed a conditional-ability change missed by both model auditors. document-fidelity-v2 adds a narrow deterministic guard; see specification 012. Earlier jobs keep their outputs but cannot be resumed under changed fidelity rules; use a new import.