Skip to content

Latest commit

 

History

History
81 lines (66 loc) · 4.55 KB

File metadata and controls

81 lines (66 loc) · 4.55 KB

Step-2 result — the pipeline runs, and the LLM-grader fallback does NOT yield a valid κ

disclosed LLM-grader fallback. The honest outcome is a negative about the method, which is more useful than the number.

What ran

  • My pilot labels for Bandit's 31 FN issues, sealed as grader_key_mine.json.
  • An independent LLM grader (fresh general-purpose agent, Sonnet) re-classified the raw issue bodies under the sealed rubric + boundary rule, blind to my labels.
  • Cohen's κ computed: grader_key_mine vs the agent.

The number, and why it is not what it looks like

n=31  agreement=31/31  po=1.000  pe=0.559  Cohen's κ = 1.000

κ = 1.0 is real arithmetic and a near-worthless reliability estimate. Three reasons, each of which I would flag in someone else's paper:

  1. Same-brain graders. The "independent" grader is an LLM of the same model family as the one doing my classification. Two raters that share weights and priors converge by construction; κ between them measures how deterministic the rubric is for one model, not whether the taxonomy carves reality. This is the design's fallback option 2, and step-2's real finding is that option 2 cannot produce a valid κ — you cannot measure inter-rater reliability with raters that share a brain.
  2. Easy-dominated data. Most of the 31 issues state their own root cause in the reporter's words ("only checks extractall, misses extract"; "no plugin exists for SSRF"). Perfect agreement on issues whose text hands you the label says little about the hard cases.
  3. The hard cases agreed by shared rule, not shared truth. The boundary cases the pilot flagged with † (639/605/551 Python-3.8 AST; 119 crash; 88 no-plugins) all agreed — but both raters applied the same pre-registered boundary rule ("AST mismatch → GRADING unless never dispatched"). Agreement there confirms the rule is deterministic, not that a human would call them the same way. A human grader is exactly where 639/551/119 could move to EXCLUSION.

The near-miss (session discipline, recorded)

The first grader run returned κ=1.0 too — but that run was contaminated: I had paraphrased each issue into interpretation-laden text ("a check exists but misses X") in the prompt, so the grader graded my summaries, not the issues. I caught it, threw it out, and re-ran with raw reporter bodies. Both the paraphrase inflation and the same-model correlation are the same failure the parent project is about — a measurement whose instrument cannot produce a counterexample. The raw re-run removes the first; nothing available to a solo run removes the second.

What step-2 actually establishes

  • Pipeline operable: classify → independent re-grade → κ all ran end-to-end. ✓
  • The recruitment blocker is real, not a formality. The design named a human (or cross-model) grader as load-bearing and offered an LLM fallback to de-risk; step-2 shows the LLM fallback does not de-risk the reliability question — it produces a confident number that means nothing. So the full study genuinely needs independent raters; it cannot be shortcut. That tightens the design.
  • Phase A⊥B concordance for Bandit is NOT cleanly testable. I already know Bandit is grading-dominant (the pilot), so any discipline rating I produce now is contaminated. A clean Phase-A-before-Phase-B test requires a fresh tool whose defects I have not seen — that is the next real step, not Bandit.

Revised design deltas (fold into STUDY_DESIGN.md)

  • Drop LLM-graders as a κ source. They may pre-label to reduce human load, but the reliability κ must come from ≥2 independent raters — humans, or at minimum two different model families (and even then, treat cross-model κ as a weak proxy and say so).
  • The two-tool step-2 must use a tool not yet classified, with Phase A sealed before Phase B, to test the concordance claim without the Bandit contamination.
  • Cross-model grader option: a different-family model (e.g. via a peer bridge) would be less correlated — but the workspace's peer-bridge contract forbids treating peer output as evidence, so it is not used here; noted as an option that needs an explicit ruling before it could serve as a grader.

One-line verdict

Step-2 did its job by failing usefully: it shows the study's independent-grader requirement is load-bearing and un-shortcuttable, and it caught two ways I nearly reported an inflated reliability. The pipeline is ready; the missing input is a rater that does not share my brain.