LucidCode + Laundry

Founded by game_nexus_zone

LucidCode compels the truth. Laundry ships the change.

Two tools, one Charter. LucidCode makes legacy code confess its latent bugs. Laundry picks up where confession ends: it runs 12 specialised agents against the same repository and produces a signed evidence bundle a human engineer approves before anything ships.

• LucidCode finds the WHY — see lucidcode/ and the live landing page.
• Laundry ships the HOW — see the Laundry page with a live demo against psf/requests v2.7.0.

---

Original LucidCode

It doesn't find bugs. It compels the truth.

LucidCode is a developer-experience tool that grants your Python code cognitive self-awareness. Instead of running static rules against your source, LucidCode makes the code confess in first person — and rigorously refutes any confession that can't be proven.

[T1] Suppression @ line 8
  › I confessed: I caught an exception and said nothing.
     Every scream is muffled the moment it starts.
     What am I hiding?
  [VERDICT: TRUTH]  posterior=0.95  driver=ast_reverifier

---

Why LucidCode

	LucidCode	Semgrep	Copilot	CriticGPT
Approach	Confession + ensemble	Static rule match	Autocomplete	Self-critique
Anti-hallucination	Diverse 3-engine	n/a	n/a	Single LLM
Output style	First-person	Rule ID	Code	Chat
Runs offline	✅ subprocess	Partial	❌	❌

Full survey of adjacent work in docs/PRIOR_ART.md.

---

Access

LucidCode and Laundry are available by direct engagement — every deployment is scoped to your codebase, security posture, and review flow.

To request the service, contact: abdarahman10555@gmail.com

Response within one business day.

---

Usage

# Analyze a file or directory
python -m lucidcode.cli analyze my_project/

# CI gate — exit code 1 if any TRUTH/LIKELY finding
python -m lucidcode.cli analyze my_project/ --min-verdict LIKELY

# JSON output
python -m lucidcode.cli analyze my_project/ --json

# SARIF 2.1.0 for GitHub Code Scanning
python -m lucidcode.cli analyze my_project/ --sarif out.sarif

# Read from stdin
cat suspicious.py | python -m lucidcode.cli analyze --stdin

---

Architecture

lucidcode/
├── pipeline.py                ← analyze() top-level entry point
├── core/
│   ├── trauma.py              ← Trauma / EngineVote / ValidatedTrauma / Verdict
│   ├── prompt_shield.py       ← prompt-injection defense
│   └── provenance.py          ← confession sentence provenance tags
├── syndromes/registry.py      ← 12 syndromes (decorator-registered)
├── sandbox/runner.py          ← docker → subprocess fallback
├── validators/
│   ├── ast_reverifier.py      ← weight 0.90 (deterministic, hallucination-immune)
│   ├── fuzzer.py              ← weight 0.70 (runs adversarial code in sandbox)
│   ├── devil.py               ← weight 0.40 (LLM, mandatory provider diversity)
│   ├── bayesian.py            ← calibrated posterior probability
│   └── orchestra.py           ← parallel engine dispatch
├── enrich/network_osint.py    ← threat-intel enrichment via cyberlab.osint
└── cli.py                     ← click-based CLI

benchmarks/
├── dataset.py                 ← 22 CVE-inspired Python fixtures
├── runner.py                  ← scoring: precision / recall / F1 per syndrome
└── RESULT.md                  ← latest scores

tests/regression/              ← 15 unit-level fixtures, pytest-driven
docs/PRIOR_ART.md              ← originality defense
landing/index.html             ← lucidcode.dev

---

The 12 syndromes

Severity	Name	Detected when
CRITICAL	Blind_Trust_SQLi	SQL built via f-string / concat
CRITICAL	Selective_Mutism	bare except: or except BaseException
HIGH	Suppression	except handler is empty / bare pass
HIGH	Network_Blindspot	requests.get() etc. without timeout=
HIGH	Hoarding	open()/socket() outside with, never closed
HIGH	Deafness	signal.signal() with empty handler
MEDIUM	Amnesia	generic error log ("wrong", "error")
MEDIUM	Insomnia	while True with no break/return
MEDIUM	Split_Personality	function returns literals of ≥ 2 distinct types
MEDIUM	Compulsion	retry loop with no backoff / sleep
LOW	Despair	hedged string return ("probably", "maybe")
LOW	Impostor_Syndrome	assert with no message

Adding a new syndrome takes ~30 lines — see the decorator pattern in syndromes/registry.py.

---

Anti-Hallucination Story

Every confession must survive an ensemble of validators before it's shown:

1. AST Re-Verifier — re-runs the syndrome's own AST predicate. Deterministic, cannot hallucinate. Weight: 0.90.
2. Fuzzer — synthesizes an adversarial harness and runs it in a locked sandbox (docker --network none --read-only --cap-drop ALL when available; otherwise subprocess -I with wiped env). If the sandbox reports VULN_TRIGGERED or times out, the trauma is confirmed. Weight: 0.70.
3. Devil's Advocate (optional) — LLM defense attorney under a 12-rule constitution. Mandatory provider-family diversity: if the confession came from Mistral, the Devil MUST come from a different family (Groq/Kimi/ Llama/DeepSeek) — this defeats within-family agreement bias. Weight: 0.40 (lowest — LLM votes count least).
4. Bayesian aggregator — combines engine votes as log-odds and returns a calibrated posterior. Verdicts:
   • TRUTH (posterior ≥ 0.85)
   • LIKELY (≥ 0.60)
   • DISPUTED (≥ 0.35)
   • HALLUCINATION (< 0.35) — logged to logs/rejected_confessions.jsonl, never shown to the user.

---

Benchmark

Latest run against 22 CVE-inspired fixtures:

fixtures=22   precision=1.0   recall=1.0   f1=1.0
clean_fp_rate=0.0             median_latency=27ms

Full report: benchmarks/RESULT.md. Reproduce with python -m benchmarks.runner --markdown.

---

Regression suite

pytest tests/regression -q
# 15 passed in 0.04s

Auto-generated report: tests/regression/REPORT.md.

---

License

MIT — core is free forever. Pro tier (VS Code inline confessions + team dashboard + Firecracker sandbox) is planned for $9/mo.

---

Status

• v4 (2026-07-16): 20/29 planned tasks shipped. Core pipeline, CLI, SARIF, benchmark, landing page all working.
• Pending: CodeQL data-flow engine · Tree-sitter multi-language · VS Code + GitHub App integrations · Whitepaper · Launch.

Track progress in STATUS_v4.md.