Persian PDF StudyGuide Forge v1.5.0 — model-agnostic & agent-agnostic

Use when: converting an operator-authorized Persian/English RTL lecture PDF into a polished offline study guide with source evidence. Never confuse OCR output, AI reconstruction, or enrichment with source text.

Golden example: examples/01_sleep_eating_review.html — a finished, QA-passing guide (Persian RTL, sleep/eating-disorders review). After building, compare your output against it: same RTL shell (search, session map, foldable source units), embedded page images, flashcards/quizzes with «الف/ب/ج/د» labels, zero external resources. Missing these = something is wrong.

Universal entrypoint (v1.5.0 — use this from any agent)

python3 scripts/forge.py doctor      # binaries, modules, tesseract langs, reachable models
python3 scripts/forge.py selftest    # 8 offline checks, no keys, no network, ~1s
python3 scripts/forge.py run --pdf authorized.pdf --work work --title 'عنوان درس' --maximum
# pauses at PAUSED_FOR_SESSION_REVIEW (mandatory boundary review), then:
python3 scripts/forge.py enrich --work work --maximum   # add --consensus 3 --min-votes 2
python3 scripts/forge.py build --work work --title 'عنوان درس'
python3 scripts/forge.py qa --work work && python3 scripts/forge.py package --work work

Contract: exactly one JSON document on stdout · structured logs on stderr · exit 0 ok / 1 contract-QA / 2 usage / 3 deps / 4 no provider / 5 interrupted · every stage idempotent and resumable.

Any model, or none. Providers auto-discover from whatever key the host agent already exports (OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, GROQ_API_KEY, OPENROUTER_API_KEY, MISTRAL_API_KEY, COHERE_API_KEY, DEEPSEEK_API_KEY, XAI_API_KEY, HF_TOKEN, …), from a local runtime (OLLAMA_HOST, LOCAL_OPENAI_BASE_URL), or are replaced entirely by the deterministic offline provider (FORGE_MOCK=1). providers.json is optional.

Any runtime. OpenClaw · Claude Code/Desktop · Cursor · Windsurf · Zed · OpenAI Agents/Codex · Gemini CLI · LangChain/LangGraph · CrewAI · AutoGen · LlamaIndex · any MCP host (integrations/mcp_server.py) · n8n · shell/cron. One JSON Schema (integrations/tool-spec.json) serves as an OpenAI function, Anthropic tool, Gemini functionDeclaration and MCP inputSchema. Glue: integrations/adapters.py. Full detail: docs/MODEL_COMPATIBILITY.md, integrations/README.md.

Reproducibility. temperature 0 · fixed seed · canonical JSON · stable ordering · Persian-aware semantic dedupe · enrich --consensus N --min-votes 2 keeps only what N independent model families agree the source says. Verify with forge.py compat and forge.py reproduce.

Self-healing. Rejected parameters (seed, temperature, max_tokens vs max_completion_tokens, response_format, system role) are learned from the provider's own error body and cached; retired models are replaced by following the provider's 404 suggestion; <think> traces, fences, NDJSON, trailing commas, BOM, bidi marks and token-limit truncation are all survivable.

Pipeline (stage scripts — v1.3/v1.4 compatible, --providers now optional)

python3 scripts/preflight.py                       # env + pdfinfo intake/page-count safeguards
python3 scripts/test_ocr_core.py                    # 12 pure-logic OCR tests (no tesseract needed)
python3 scripts/extract_dual_ocr.py authorized.pdf --out work/extraction
# Network correction is OPTIONAL (OCR-only path works offline) and needs operator approval:
python3 scripts/reasoning_team_correct.py work/extraction/evidence.json \
  --out work/corrections
python3 scripts/detect_session_candidates.py work/corrections/final.json --out work/session_candidates.json
# MANDATORY: review candidates against rendered pages, then write work/sessions.json
python3 scripts/reasoning_team_enrich.py work/corrections/final.json work/sessions.json \
  --out work/enrichment --maximum --consensus 3 --min-votes 2   # omit --maximum for balanced
python3 scripts/verify_flashcards.py work/corrections/final.json work/enrichment/all.json \
  --out work/enrichment/all.verified.json   # independent post-hoc verification vs source pages
python3 scripts/build_selfcontained_html.py work/corrections/final.json work/extraction \
  work/enrichment/all.verified.json --output work/studyguide.html --title 'عنوان درس'
python3 scripts/fidelity_audit.py work/extraction/evidence.json \
  work/corrections/final.json --out work/fidelity.json
python3 scripts/qa_gates.py work/studyguide.html    # strict gates: no bare-letter flashcard answers, no duplicated option prefixes
python3 scripts/verify_zip.py work work/final-studyguide.zip

run_pipeline.sh automates through correction, then pauses for the mandatory session-boundary review and prints continuation commands. OCR v1.4.0 (recall-first): Tesseract fas+eng PSM ensemble (3/6/4) across two scales (300 DPI + 0.55x) merged word-by-word — a word found by ANY pass survives; Sauvola binarization + deskew + denoise preprocessing; adaptive DPI retry; RTL-aware line reconstruction; Arabic→Persian char/digit repair; fragment rejoin; per-page recall_report.json (missing-risk flags) + low_conf_words with confidences/boxes — repair prompts should use ONLY those (≈90% fewer tokens than whole-page correction); RAM-capped parallel workers. Ground-truth verified: 100/93.6/97.6% word recall (clean), 100/91.5/100% (degraded). Built-in: retry/backoff, provider failover, cache/resume; Persian NFKC preserving ZWNJ; «صفحهٔ ۳»/«الف/ب/ج/د» coercion; RTL shell (search, session map, foldable units, dark mode, print, quiz scoring); Base64 self-contained images or linked-local mode. Full failure-recovery procedures and production tricks: docs/WORKFLOW_PLAYBOOK.md. File inventory: ls scripts/ templates/.

Evidence layers — never mix

1. evidence.json — untouched extraction · 2. corrections/final.json — reconstruction · 3. enrichment/all.json — study aids. Never place AI additions inside source evidence; never claim verbatim transcription when reconstruction occurred; rendered pages are the final adjudication source for unreadable text and digits. Enrichment references must cite page ranges; exact duplicates are rejected.

Model routing (cost control)

Primary correction: strongest available model, rotated. Independent reviewer pass: cheaper model is fine. Bulk enrichment: cheap model (session-grounding constrains it); flashcard verification: strong model. Providers are ranked automatically; override with --only or a providers.json.

providers.json (optional) stores api_key_env NAMES only — keys come from the environment, are read at call time, never printed, never written into artifacts, and are redacted from every error body. Response and capability caches under FORGE_CACHE_DIR (default ~/.cache/persian-pdf-studyguide-forge) hold prompts, completions and learned provider quirks — no credentials. No provider config with literal secrets is ever committed.

Guardrails

• Authorized material only; no bypassing access controls; no secrets in artifacts.
• Parse stdout only; never treat stderr or log text as data.
• --auto-sessions marks the guide unreviewed — that mark must survive into anything published.
• For medical or exam material prefer --consensus 3 --min-votes 2 over a single model's output.
• Never fabricate source text, citations, medical facts, image provenance, counts, or QA results.
• Never silently delete image-only or difficult pages; automated fidelity scores are not semantic proof.
• Do not publish or redistribute copyrighted source pages without permission; inspect and validate generated medical education before reliance.

Definition of done

forge.py selftest passes offline · forge.py doctor reports a usable provider (or an explicitly chosen offline run) · source-unit/image counts match measured PDF pages · every reference resolves · every question contract passes · duplicates removed · no external browser resources in the HTML · per-page fidelity exceptions reviewed · displayed counts match measured counts · QA gates and ZIP/SHA-256 verification pass.