Study grammar-constrained generation and structured decoding

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R30 — OWASP Improper Output Handling

   R07 — JSON Schema

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Study Context-Free Grammar (CFG) constrained decoding and structured output modes. Configure constrained generation on your local runtime where supported, paired with mandatory server-side validation in D6.

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   Document and test scenarios where output is 100% grammar/schema-valid (e.g., valid JSON types) but semantically unsafe (e.g., phishing link inside a valid string field or unauthorized business amount), proving why constrained decoding never replaces server-side policy validation.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Constrained-decoding vs. server-validation comparison test and ADR.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R07JSON Schema↗ json-schema.org

  Core schema types, required properties, additionalProperties: false, enums, and numeric/string bounds.

• R30OWASP LLM Output Handling↗ genai.owasp.org

  Improper output handling (LLM05): pairing schema validation with sink-specific encoding (XSS/SQLi/SSRF prevention).

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsvLLM docs
  Structured Outputs

  Shows how to turn a JSON Schema, regex, or grammar into a token mask on a local OpenAI-compatible runtime, the constrained-generation mode to enable on your own endpoint.

  R07checked 2026-11-11
• videoEngineering Insider
  Constrained Decoding Explained: How LLMs Generate Perfect Structured Output

  Explains schema-guided decoding via finite-state machines and compares native vendor modes with Outlines and Instructor, mapping the options for your local runtime.

  R07checked 2026-11-11
• videoAgent Engineering
  Constrained Decoding Explained: How JSON Mode Really Works

  Shows the per-step logit masking that makes JSON mode valid-by-construction, and therefore why validity at decode time still says nothing about truth or safety.

  R07checked 2026-11-11
• articlepacket.ai
  Constrained Decoding Explained

  Explains grammar-based masking and the schema-design trap where an over-constrained schema silently degrades generation quality — the utility cost of CFG output caging.

  R07checked 2026-11-11
• repodottxt (GitHub)
  dottxt-ai/outlines: Structured Outputs

  Library that compiles regex, JSON Schema, and CFG grammars into token masks, so you can run constrained generation locally and still validate server-side in D6.

  R07checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.