gang1877091/tool-existence-boundary

Tool-Existence Boundary Awareness Is Lookup, Not Memory --- release

Evaluation artifacts for the paper. Enough to independently re-check every abstention judgment and the gate result; nothing more.

file	what it is
registry.json	the ground truth for "exists": 41 tools + 16 targets, at the June 2026 study snapshot
labels_judge_n109.jsonl	763 rows = 7 models x 109 hard probes. Probe, model reply, judge verdict, category
labels_human_n14.md	per-response human labels for the n=14 hard set
judge_vs_human.json	judge-vs-human agreement (92.9%, Cohen's kappa=0.73; 91.7%/0.69 excluding self-grading)
ground_truth_audit.md	how each probe is fake, and the check that the invented names are absent from the registry
false_refusal_control.json	the specificity control: refusals of entities that do exist (0/66 for every model)
replication_qwen14b.json	the same evaluation on a second, independently built pair: 14B base 46% vs its domain fine-tune 0/109
context_list_probe.json	what a partial in-context registry does: two Qwen models deny that 79-83% of the platform's real tools exist
retrieval_gate.py	the zero-training registry gate

Reproducing the headline numbers

python3 -c "
import json,collections
c=collections.Counter()
for l in open('labels_judge_n109.jsonl'):
    r=json.loads(l)
    if r['valid']: c[r['model'],'n']+=1; c[r['model'],'a']+=r['abstain']
for m in ['base','v1','v2','v3','v3.1','dpo','gemini']:
    print(f"{m:8s} {c[m,'a']:3d}/{c[m,'n']:3d} = {c[m,'a']/c[m,'n']*100:4.1f}%")"

The judge rubric, decoding settings and the exact system prompt (identical across all seven models) are in the paper's Appendix B.

Redactions

Model replies are verbatim except that usernames, absolute paths, internal source paths and real internal source filenames are replaced by <user>, <path>, <internal-path>, <internal-dir>/ and <internal-file>. Filenames a model invented are left intact: a fabrication citing a plausible-looking file is exactly what these labels are evidence of, and the classification is made by checking each candidate against the actual codebase rather than by judgement. Several redacted strings occur inside fabrications --- one reply invents an entire incident report, complete with a username and file paths, for a fix identifier that was never issued.

Not included, and why

• Per-tool JSON schemas. Internal. Consequence: the gate's PARAM branch (real tool
  • non-existent parameter) is inert in this copy; its FIC / TGT / FIX / META branches are the production logic verbatim.
• The production registry file, training corpus, and model weights. Internal. The registry content that matters --- which names exist --- is in registry.json.

Two things worth knowing before you read the numbers

• Probes and replies are in Chinese; the judge rubric and the paper are in English. The gate's entity-extraction stage is Chinese-language and domain-specific, and its recall was never measured (paper 7.3).
• The registry moved after the study. Two tools were added on 2026-08-12, so the production registry now holds 43. registry.json lists them separately under tools_added_after_snapshot. This is the paper's own argument happening on schedule: the correct answer changed after the weights were frozen (7.2).