Oct 29 23 days ago 05:17 Trim the artifact-gap note from the drift README 05:15 Document drift-evaluation ground-truth semantics and correctness criterion Clarifies that drift_ips.csv membership (class-level ground truth expanded
to IPs; detection precision 100% by construction, recall 814/1357=60%) and
correct/incorrect_ips.csv (post-drift re-identification vs ground truth,
773/814=95.0%) answer different stages — resolving the reviewer-observed
overlap of incorrect IPs with the ground-truth list. Documents column
semantics and the unretained per-IP prediction labels. 05:01 Replace gate report with annotated .jsonc + add honest vendor-fix report - release_gate_report_v2.jsonc: original numbers verbatim (passed:false,
vendor_novelty_f1 0.0098) with a comment noting the pseudo-positive
root cause (vendors registry-resident, score forced to 0 by construction)
- release_gate_report_v2_vendorfix.json: vendor novelty re-scored with 168
real out-of-registry vendor positives; gate passes honestly
(vendor_novelty_f1 0.994, tp 168/168, note on None-type short-circuit included) Oct 19 07:23 Addendum is dataset-mirror-only: code repository keeps paper code; fix dangling report references 06:47 Bench harness scripts (rebuttal regeneration tooling; kept off the code repository) 06:47 Bench harness scripts (rebuttal regeneration tooling; kept off the code repository) 06:47 Bench harness scripts (rebuttal regeneration tooling; kept off the code repository) 06:47 Bench harness scripts (rebuttal regeneration tooling; kept off the code repository) 06:47 Bench harness scripts (rebuttal regeneration tooling; kept off the code repository) 06:46 Bench harness scripts (rebuttal regeneration tooling; kept off the code repository) 06:24 Addendum 1.2: fullflow validity is enforced by the code schema gates; valid reference run published at evaluation/demo/fullflow/ 06:24 Publish the v3 fullflow reference evaluation report (completion 1.0, zero integrity violations) 06:24 Publish the valid v3 fullflow reference manifest (retrieved from git history fc0e8cf) at the dataset mirror 06:20 Demo resolves evidence from demo/evidence first (fresh-clone --skip-model needs no downloads) 06:20 Ship the ALARM evidence caches + vendor lexicon alongside the unseen-detector demo 06:06 Simplify unseen-detector demo to ALARM-only (drop the CONTROLLER degraded path) 06:06 Simplify unseen-detector demo to ALARM-only (drop the CONTROLLER degraded path) 06:01 Add single-file unseen-detector demo (ALARM/CONTROLLER: non-empty community/reasoning evidence, live 4-bit detector, ranked top-3 vendors) 06:01 Add single-file unseen-detector demo (ALARM/CONTROLLER: non-empty community/reasoning evidence, live 4-bit detector, ranked top-3 vendors) Oct 18 17:52 Publish rebuttal addendum (Table 8/10 matching rules, coverage disclosure, package, identifiers) and pinned manifest 17:52 Publish rebuttal addendum (Table 8/10 matching rules, coverage disclosure, package, identifiers) and pinned manifest 17:50 Restore ground-truth drift reasons 17:50 Restore CADE per-IP detection reasons backing 767/1112 17:50 Restore IoTProber per-IP detection reasons backing 773/814 17:50 Restore CADE Stage-3 classification report 17:50 Restore val_metrics.json (provenance of CADE Stage-3 row 0.8402/0.5430/0.8649) 17:50 Restore IoTProber Stage-3 classification report 17:50 Restore val_metrics.json (provenance of IoTProber Stage-3 row 0.8483/0.5509/0.8711) 17:50 Restore original SVM trainer/scorer (svm.py from initial commit 25f6a47) — source of Table 10 Stage-2 exact-match and Stage-3 normalize_label rules 17:50 score_table8 v1.1.0: full paper-cell comparison; IoTProber vendor cells reproduce under exact match, B2 under fuzzy; B1 documented as not reproducible 17:50 score_table10 v1.1.0: recover the ORIGINAL Stage-3 matching rule from svm.py (commit 25f6a47) — reproduces paper cells bit-exactly; replaces the wrong philipshue alias 17:39 Publish canonical versioned scorers for Tables 8 and 10 (explicit label-matching rules) 17:39 Publish canonical versioned scorers for Tables 8 and 10 (explicit label-matching rules) 17:39 Add missing IoTProber Alarm/Controller type+vendor prediction CSVs (complete the 11-type set) 17:39 Add missing IoTProber Alarm/Controller type+vendor prediction CSVs (complete the 11-type set) 17:39 Add missing IoTProber Alarm/Controller type+vendor prediction CSVs (complete the 11-type set) 17:39 Add missing IoTProber Alarm/Controller type+vendor prediction CSVs (complete the 11-type set) 17:39 Add per-type decision predictions for all 11 known types (incl. Alarm/Controller) used by Tables 2/3 15:25 Add v2 unseen adapter stack (now the agent default path) - fine-tune-v2.py: superset trainer (parallel tokenization, accelerate even_batches fix, --write_model_on_gate_failure/--save_steps/--skip_builtin_calibration)
- dataset/unseen_sft_v2.jsonl: v2 SFT data
- results_v2/final_model: gate-release adapter (type_novelty_f1 0.899, known type_exact 93.4%) + calibrated known_vendors.json — referenced by agent.py DEFAULT_UNSEEN_ADAPTER_PATH
- results_v2/checkpoint-2191: raw trainer checkpoint
- results_v2/release_gate_report_v2.json: gate evidence (vendor_novelty_f1 0.010 failed) Oct 17 17:51 Delete generate_predictions.py Aug 16 10:04 Upload evaluation/unseen/split_and_evaluate.py with huggingface_hub 10:04 Upload evaluation/unseen/metrics.csv with huggingface_hub 10:04 Upload evaluation/unseen/vendor/vendor_predictions_llama3raw_VPN.csv with huggingface_hub 10:04 Upload evaluation/unseen/vendor/vendor_predictions_llama3raw_MEDIA_SERVER.csv with huggingface_hub 10:04 Upload evaluation/unseen/type/type_predictions_llama3raw_VPN.csv with huggingface_hub 10:04 Upload evaluation/unseen/type/type_predictions_llama3raw_MEDIA_SERVER.csv with huggingface_hub 10:04 Upload evaluation/unseen/predict/baseline/llama3raw/all_result_type_VPN.csv with huggingface_hub 10:04 Upload evaluation/unseen/predict/baseline/llama3raw/all_result_type_MEDIA_SERVER.csv with huggingface_hub 10:04 Upload evaluation/unseen/predict/baseline/llama3raw/typeonly_result_VPN.csv with huggingface_hub 10:04 Upload evaluation/unseen/predict/baseline/llama3raw/typeonly_result_MEDIA_SERVER.csv with huggingface_hub