DARE — End-to-End Field Test

⚠️ This is NOT a real research project. It is an end-to-end (e2e) field test of an AI research-skill orchestration system (the De-Anthropocentric Research Engine, "DARE"). The scientific artifacts in this repo — including the proposed method SPB-Net — were produced by an automated skill pipeline as a capability test, and have NOT been experimentally validated. Do not cite, build on, or treat any claim here as an established research result.

What this repo actually is

A single complete run of the DARE pipeline, captured as artifacts, to test whether the skill system can take a vague research interest all the way to an executable, self-critiqued research plan. The topic (few-shot medical image segmentation robustness) is just the test subject — the real subject under test is the pipeline itself.

The run exercised the full pipeline:

North Star crystallization
  → Research Spec generation
    → 12-stage spiral execution (breadth Round 1 → GATE-1 → depth Round 2 → GATE-2)
      → adversarial stress-test (incl. simulated MICCAI Reviewer-2)
        → experiment design (design only — no experiments were run)

Repo contents

Folder	What's inside
context/	The full reasoning trail — one timestamped file per stage (knowledge acquisition, gap analysis, hypotheses, ideation, two convergence gates, stress test, experiment design) + INDEX.md.
docs/	The generated Research Spec (the machine-executable plan the pipeline produced).
presentations/	A self-contained HTML slide deck (presentations/spb-net/index.html) explaining the produced method in plain language. Open it in any browser.

The produced method (for context only — unvalidated)

SPB-Net targets an observation that few-shot medical segmentation results swing widely depending on which support samples are drawn, while papers typically report only the mean. It proposes (1) RQA — robust quality-weighted aggregation of support, and (2) BSI — bounded-sensitivity support injection, plus a statistical protocol to quantify support-set-composition variance. Its honest self-assessed contribution is problem-framing + method + evaluation protocol, not a new architectural primitive.

Honest limitations of this field test

• No experiments were executed; all numbers referenced are from the literature, not from running SPB-Net.
• The literature retrieval covered the field but missed several near-neighbor works (e.g. TraNFS, DETA++, LRFS); these were caught only by external review and recorded in context/ as a correction — an honest artifact of the test, not hidden.
• The reasoning in most stages is single-context synthesis built on a few deep-retrieval steps; depth is concentrated in the knowledge-acquisition and stress-test stages.

How to read it

• Plain-language overview → open presentations/spb-net/index.html.
• Final method definition → context/2026-06-05-21-13-gate2-final-approach.md.
• Experiment design → context/2026-06-05-21-20-experiment-design-final.md.
• Full stage-by-stage trail → context/INDEX.md.