MethodBridge Local

Offline scientific-reasoning support for African researchers using ordinary laptops.

Comment

MethodBridge Local

One-line summary

An offline, CPU-first research-methods and scientific-reasoning assistant for postgraduate students and early-career researchers using ordinary laptops.

Project details

Team ID: methodbridge-local Domain: Math & Scientific Reasoning Cross-disciplinary pairing: Education Repository: https://github.com/MaPeL-LAB/MethodBridge Report: https://github.com/MaPeL-LAB/MethodBridge/blob/codex/ops-001-assurance-hardening/REPORT.md Demo video: https://youtu.be/pwILKBdyXaA

Problem

Students and early-career researchers often need help interpreting study design, causal claims, statistical results, bias, uncertainty, and reproducibility. Reliable internet, cloud-inference budgets, and specialist support cannot always be assumed. That can make high-quality methodological support harder to access precisely when it is needed.

Solution

MethodBridge Local runs a compact GGUF model through CPU-only llama.cpp on the learner's own computer. The response contract asks for a direct explanation, assumptions, missing information, limitations, and appropriate human review. Inference runs offline after a verified model download, so prompts remain on the local machine rather than being sent to a hosted application.

The owner-approved artifact is Qwen/Qwen2.5-1.5B-Instruct revision 8ee93c2720648b8ffbb4942ceed829eb72978c29, mechanically converted and quantized as Q5_K_M. Its exact GGUF SHA-256 is 52ba7f9a754e675f19b4dc9a70ff0f3396fd43ffa7052304f596235b3e7efa37. The public release page is https://github.com/MaPeL-LAB/MethodBridge/releases/tag/v0.1.0-adtc-2026 and the credential-free stable asset endpoint is https://github.com/MaPeL-LAB/MethodBridge/releases/latest/download/methodbridge-local-qwen2.5-1.5b-instruct-q5_k_m.gguf.

Why this matters

MethodBridge treats offline availability, privacy, cost, and 8 GB-class laptop constraints as core design requirements. It provides a focused learning tool without claiming to replace supervisors, statisticians, ethics committees, clinicians, legal advisers, or institutional decision-makers.

How AI is used

  • Qwen2.5-1.5B-Instruct supplies the local language-model capability.
  • The MethodBridge contract structures answers around reasoning, uncertainty, limitations, and human review.
  • A frozen, training-excluded 60-case public benchmark covers study design, quantitative reasoning, causal interpretation, bias, uncertainty, reproducibility, citation integrity, and safe refusal.
  • Qualified semantic review of the 60 contract-mode public cases recorded 54 pass, 4 fail, and 2 inconclusive Q5_K_M judgments, with 33 case preferences. These are bounded reviewer counts—not accuracy, hidden-prompt results, or an official ADTC score.
  • The protected 20-case private challenger and raw evaluation outputs remain outside the public repository.

How Codex was used

Codex helped turn the product theory of change into versioned requirements, architecture decisions, evidence schemas, privacy boundaries, and executable tests. It also helped implement the non-interactive llama-completion path, identify earlier canned and interactive outputs as invalid empirical evidence, and keep model weights and private captures out of Git. The accountable owner, not Codex, made the finalist, public-prompt, release, and final submission decisions.

Key features

  • Fully offline, CPU-first GGUF inference through pinned llama.cpp.
  • Research-methods and scientific-reasoning support paired with education.
  • Structured explanations of assumptions, missing information, limitations, and appropriate human review.
  • Explicit safety, academic-integrity, and human-authority boundaries.
  • Credential-free HTTPS model download with SHA-256 verification.
  • Reproducible source, runtime, benchmark, and model provenance.

Performance and evidence scope

The organizer-prescribed participant_laptop profiler record was collected on the participant's Apple Silicon Mac. It reports 152.17 generation tokens/s, 871.64 ms first-token latency, and 2300.44 MB peak RSS. Those are participant submission evidence only. They are not ADTC Standard Laptop results and do not establish Standard Laptop throughput, memory, efficiency, TTFT, or thermal behaviour. The organizer-controlled Standard Laptop audit remains separate and controlling.

Simulation output is a plumbing-only test double. Its automated_keyword_proxy_pass_rate is not accuracy or expert judgment.

Testing instructions

git clone https://github.com/MaPeL-LAB/MethodBridge.git
cd MethodBridge
python3.12 -m venv .venv
.venv/bin/python -m pip install -e '.[dev]'

./download_model.sh
PYTHONPATH=src .venv/bin/python scripts/validate_gguf.py model/methodbridge-local-final.gguf
PYTHONPATH=src .venv/bin/python scripts/validate_repository.py
PYTHONPATH=src .venv/bin/python scripts/validate_public_claims.py
PYTHONPATH=src .venv/bin/python -m pytest -q

Build the pinned llama.cpp revision and run the learner command according to USAGE.md. The model is Apache-2.0; the repository includes attribution, licence, mechanical-change, and no-endorsement notices.

Limits and safe use

MethodBridge does not approve research protocols, analysis plans, ethics, clinical care, legal or regulatory compliance, or institutional decisions. Do not put participant data, confidential materials, credentials, or private benchmark cases into prompts. Verify citations independently and seek qualified human support for consequential decisions.

Official form values

  • Project report public URL: https://github.com/MaPeL-LAB/MethodBridge/blob/codex/ops-001-assurance-hardening/REPORT.md
  • Test prompt 1: At a university, postgraduate students who voluntarily attended a research-methods workshop were more likely to submit their proposals on time than students who did not attend. A researcher concludes that the workshop caused the improvement. Critique this causal conclusion, identify important sources of bias or confounding, state what additional information is needed, and propose a stronger design or analysis.
  • Test prompt 2: A cohort study reports a risk ratio of 1.40 with a 95% confidence interval of 0.98 to 2.00 and p=0.064. Explain what can and cannot be concluded. Discuss effect size, uncertainty, the danger of a binary significant/not-significant interpretation, and what additional context is needed before judging practical importance.
  • Problem domain: Math & Scientific Reasoning.
  • Participant throughput: 152.17 generation tokens/s — participant-laptop evidence only; not a Standard Laptop result.
  • Participant peak RSS: 2300.44 MB — participant-laptop evidence only; not a Standard Laptop result.
  • Required video URL: https://youtu.be/pwILKBdyXaA

The public release and independent clean-download/load verification are complete. The final Devpost action still requires the participant's separate explicit confirmation.

Built With

Share this project:

Updates