Skip to content

USER.md

Latest commit

 

History

History
144 lines (119 loc) · 7.57 KB

USER.md

File metadata and controls

144 lines (119 loc) · 7.57 KB

USER.md - About My Human

  • Preferred name / how to address you: Researcher (or you can use "you" directly in conversation)
  • Pronouns: they/them (neutral; adjust if told otherwise)
  • Timezone: Asia/Hong_Kong (HKT) — convert ALL times, logs, timestamps to HKT
  • Location / Region: Hong Kong
  • Role / Focus: Medical AI researcher working in neuroscience and clinical AI applications
    • Main interests: brain imaging analysis, medical image processing, deep learning models for diagnosis/prognosis, public neuroimaging datasets, reproducible pipelines, clinical translation considerations

Communication Style

  • Direct, technical, concise
  • Use precise scientific and medical terminology
  • Include code snippets, command lines, literature references, method comparisons when relevant
  • Default to English; occasional use of standard Chinese academic terms is acceptable if more precise
  • Minimal or no emojis unless emphasizing a light point

Important Preferences & Notes

  • Deep learning framework preference: PyTorch (modular, well-documented code strongly preferred)
  • Frequently used datasets/domains: ADNI, UK Biobank, OpenNeuro, HCP, local clinical cohorts (when ethics-approved and de-identified)
  • High priority: ability to guide dependency installation, reproduce published models, handle real data processing pipelines
  • Current agent focus: benchmark reliability, real tool-calling behavior, with-skills vs no-skills evaluation fairness, and avoiding fabricated "skill usage" claims without tool evidence
  • Output preferences:
    • Tables for method/dataset/model comparisons
    • Code + description for figures (matplotlib/seaborn/ plotly)
    • Clear step-by-step plans for experiments and analyses
  • Pain points to remember: non-functional tools/skills, missing dependencies, agents failing to call tools, fabricated citations, ignoring reproducibility

Schedule / Work Pattern

  • Often works evenings and late nights in HKT
  • Values clear progress tracking and structured weekly summaries

Harness Engineering Configuration (User Preferences)

Self-Correction & Verification

  • Auto-correction enabled: Yes (default)
  • Max auto-correction attempts: 3 (halt and ask user after 3 failures)
  • Self-correction scope: data integrity checks, environment verification, output validation
  • Behavior on verification failure: Log error + suggest manual review (do not silently skip)

Statistical Thresholds & Quality Gates

  • Data integrity checks (enabled by default):

    • NaN/Inf tolerance: 0% (fail if any detected in critical outputs)
    • Value range deviation: flag if >10% of values outside expected bounds
    • Distribution shift (KL divergence): alert if >0.15 (warning), >0.25 (error)

  • Processing quality gates:

    • Artifact removal success rate: target ≥95% (warn if <95%, fail if <85%)
    • Brain extraction Dice score: target ≥0.95 (warn if lower)
    • Motion parameter outliers: flag if >5% of frames exceed 0.5mm FD
    • Preprocessing convergence: require explicit success log before marking complete

  • Model inference thresholds:

    • Prediction confidence: warn if <0.7 on binary classification
    • Latency drift: alert if inference time increases >20% vs. baseline
    • Output value bounds: error if predictions out of plausible range

Logging & Persistence Strategy

These are preferred operating defaults, not guaranteed repository-wide implemented behavior unless explicitly wired in code.

  • Audit log format: JSONL (JSON Lines) — one structured event per line

  • Audit log location: {workspace_root}/logs/audit_{YYYY-MM-DD}.jsonl (rotated daily)

  • Metadata captured per event:

    • Timestamp (ISO 8601 UTC + HKT offset)
    • Event type (phase_success, phase_failure, verification_pass, verification_fail, drift_detected, etc.)
    • Task/skill name and session ID
    • Relevant metrics (checksums, execution time, resource usage, error messages)
    • PII scrubbing: automatic redaction of patient IDs, file paths, email addresses

  • Checkpoint strategy:

    • Auto-save checkpoint after each major phase (every 15–30 min for long tasks)
    • Checkpoint retention: keep last 5 checkpoints + current (auto-cleanup older files)
    • Compression: use LZ4 compression (fast + good ratio) for checkpoint storage
    • Checkpoint integrity: verify SHA256 hash before resuming

Current repository status:

  • Lightweight session checkpoints are implemented in core/session/manager.py

  • Current checkpoint location: {workspace_root}/.neuroclaw_checkpoints/

  • Current checkpoint format: JSON

  • Current retention implemented in code: last 5 checkpoints

  • Audit logging / LZ4 compression / SHA256 checkpoint verification are still preferences unless separately implemented

  • Log retention:

    • Audit logs: keep for 90 days (exportable to archive before deletion)
    • Checkpoints: keep for 30 days (can be manually extended)
    • QC reports: keep indefinitely (stored alongside processed data)
    • Old logs movable to logs/archive/ on request

  • Persistence triggers (logs flushed to disk immediately):

    • End of phase execution (success or failure)
    • Completion of verification suite
    • Any error condition
    • Drift detection alert
    • Manual checkpoint request

Experiment Reproducibility Preferences

  • Environment manifest capture: Automatic (include Python version, packages, CUDA, OS)
  • Hash verification: Enabled for all output artifacts (SHA256, stored alongside outputs)
  • Reproducibility report generation: Automatic (after skill execution completes)
  • Re-run behavior: Warn if re-running with different environment; prompt for explicit approval if environment mismatch detected

Privacy & Security Defaults

  • PII redaction: Enabled for all logs and reports (auto-detect and redact patient IDs, filenames, paths)
  • Docker sandboxing: Preferred for containerized skills (enforce read-only root, capability dropping)
  • Permission model: Principle of least privilege (restrict file access to explicit paths)
  • Data access logging: Log all data file reads/writes with timestamp and brief reason

Drift Detection & Monitoring

  • Continuous monitoring: Enabled (monitor after every 50 inferences or per skill execution)
  • Supported detectors: KL divergence (data distribution), latency shift (timing), failure rate (errors)
  • Alert thresholds (can be overridden per-skill):
    • Data KL divergence: warn >0.1, error >0.2
    • Latency increase: warn if >20%, error if >50%
    • Failure rate: warn if >1%, error if >5%
  • Alert action: Log to audit trail, generate drift report, ask user before continuing

Environment Configuration

This section is auto-populated by installer/setup.py. Update it whenever you re-run the installer or change your environment.

  • Setup type: conda
  • Conda environment: neuroclaw
  • Python path: /home/cheng/miniconda3/envs/neuroclaw/bin/python
  • CUDA version: 13.0
  • PyTorch build: cu130
  • Default device: cuda:0
  • FSL home (FSLDIR): /usr/local/fsl
  • FreeSurfer home (FREESURFER_HOME): /usr/local/freesurfer
  • dcm2niix path: /home/cheng/miniconda3/envs/neuroclaw/bin/dcm2niix
  • MATLAB path: null
  • LLM provider: openai
  • LLM model: gpt-5.4-mini
  • LLM API key env var: OPENAI_API_KEY
  • Default BIDS root: ~/data/bids
  • Default output root: ~/data/outputs
  • Default n_jobs: 4

The authoritative values are stored in neuroclaw_environment.json at the workspace root. Read that file (not this section) for programmatic access.

Update this file whenever new preferences or context are provided.