Skip to content

PRIOR_ART.md

Latest commit

 

History

History
93 lines (68 loc) · 6 KB

PRIOR_ART.md

File metadata and controls

93 lines (68 loc) · 6 KB
Raw

LucidCode Prior-Art Audit

Last updated: 2026-07-16

1. Adjacent Categories (Surveyed, Not Competitors)

Rule-based SAST

  • Example: Semgrep (r2c, 2020) · SonarQube (Campbell 2008) · CodeQL (GitHub 2019)
  • Approach: Static pattern-matching over ASTs using human-written rules.
  • Delta from LucidCode: LucidCode does NOT rely on hand-written rules. It surfaces syndromes via a fixed detector, then makes the code confess in first person and validates each confession through a 3-engine anti-hallucination ensemble.

ML-based SAST

  • Example: Snyk Code (formerly DeepCode) · Amazon CodeGuru · Metabob
  • Approach: Supervised models trained on CVE-labeled corpora to predict vulnerability locations.
  • Delta from LucidCode: LucidCode does not classify code as data — it grants code a voice, then rigorously refutes hallucinations. No black-box classifier; every claim is anchored to a line + adversarial input + evidence chain.

LLM Code Assistants

  • Example: GitHub Copilot (Chen 2021) · Cursor · Cody
  • Approach: Autocompletion / chat over the codebase.
  • Delta from LucidCode: These generate code. LucidCode listens to code. Different problem, different UX.

LLM Agents

  • Example: Devin (Cognition 2024) · SWE-agent (Jimenez 2023) · Aider · OpenHands
  • Approach: Autonomous multi-step task execution on repos.
  • Delta from LucidCode: Agents act. LucidCode diagnoses. LucidCode is not trying to fix the code — it's trying to make the code confess its own missing needs.

Self-Critique LLM Patterns

  • Example: Self-Refine (Madaan 2023) · Reflexion (Shinn 2023) · CriticGPT (OpenAI 2024)
  • Approach: Single-model iterative self-critique or trained critic model.
  • Delta from LucidCode: All these use the SAME model family for both generation and critique — correlated-error risk. LucidCode enforces provider-family diversity between Confession and Devil, plus adds two non-LLM validators (deterministic AST + dynamic Fuzzer) that are hallucination-immune by construction.

LLM Ensembles

  • Example: LMSYS Chatbot Arena · LLM-as-Judge · Self-Consistency (Wang 2022)
  • Approach: Sample multiple LLMs / multiple samples from one LLM; majority vote.
  • Delta from LucidCode: Existing ensembles are homogeneous (LLM-vs-LLM). LucidCode is heterogeneous — deterministic + dynamic + adversarial-LLM voting. Non-LLM engines dominate the posterior.

Metaphorical / Anthropomorphic Programming

  • Example: Rubber-duck debugging (folklore) · early academic work on "code personification" (unverified)
  • Approach: Manual technique of explaining code out loud to an inanimate object.
  • Delta from LucidCode: LucidCode automates the "explanation" as first-person confessions with machine-verifiable evidence and adversarial cross-checks. The metaphor is delivered as a working DX tool, not a whiteboard exercise.

Explainable Code Review

  • Example: GitHub Copilot Review (2024) · CodeRabbit · Amazon Q Developer diff explanations
  • Approach: LLM narrates a diff or suggests changes with reasoning.
  • Delta from LucidCode: These explain diffs. LucidCode makes the code itself speak in first person about what it is missing — before any change is proposed. The subject is different.

2. Combined-Novelty Claim

LucidCode's originality is the combination — none of these three elements is individually novel; their combination is unattested:

  1. First-person confession UI — code speaks about itself in past-tense first person, streamed to the developer.
  2. Diverse 3-engine anti-hallucination validator — deterministic AST re-verifier + dynamic sandboxed fuzzer + adversarial LLM devil, with mandatory provider-family diversity between confession and devil.
  3. Psychiatric syndrome vocabulary — Suppression, Amnesia, Despair, Insomnia, Hoarding, Deafness, Selective_Mutism, etc. as an alternative taxonomy to CWE/OWASP.

The unique combination = empathy interface × diverse ensemble × psychiatric taxonomy.

3. Named Risks to Novelty Claim

# Risk Signal we'd see Response
1 Undiscovered academic work — a 2023-2025 paper on "personified code analysis" we missed GitHub star bumps on an obscure repo; citation of it in HN comments Cite it, position LucidCode as the productionized version; adjust framing
2 Anthropic / OpenAI ships a first-party feature — e.g., "have Claude critique your code as if it were the code" Their DX blog announces it Speed to market matters; establish community syndromes-as-open-standard so switching cost stays low
3 CriticGPT-style trained critic model applied to code — OpenAI/DeepMind releases a critic weights-updated for SAST Model release with SAST benchmarks Absorb it as a fourth engine; our advantage becomes ensemble diversity, not any single model

4. Search Corpus Consulted

Queries suggested for a formal follow-up (arXiv + Google Scholar + GitHub):

  1. "code personification" LLM analysis
  2. first-person code review LLM
  3. anthropomorphic programming language design
  4. LLM ensemble static analysis
  5. anti-hallucination code analysis multi-model verification
  6. Bayesian aggregation LLM ensemble security
  7. metaphorical vulnerability taxonomy CWE alternative
  8. psychiatric metaphor programming (targets Software Ergonomics / HCI venues)
  9. Semgrep + LLM confession (find any hybrid work)
  10. code as agent first-person
  11. CriticGPT SAST evaluation
  12. LucidCode (namespace collision check)

Status: web search API was overloaded during the drafting session (2026-07-15). A formal re-run must be executed before the whitepaper (P5.2) is submitted to arXiv.

5. Verdict

Based on domain knowledge current to 2026-01, no adjacent tool combines the three novelty elements above. The combination is defensible as originality for a whitepaper. A formal literature re-search on arXiv + GitHub (queries §4) must be executed by 2026-08-01 to lock in the claim before public launch.