Plot benchmark performance against inference cost

README.md

@@ -20,7 +20,7 @@ Pairs come from open-source projects in Rust, JavaScript, TypeScript, Python, Go
 | [SemiF](https://github.com/TheoLeeCJ/SemIf) | [Beam endpoint](https://app.beam.cloud/v1/models/jev/semif/invoke) |
 | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | Runpod H100 80 GB pod (US-NE-1) |
 | [Span-01](https://www.respan.ai/docs/documentation/span-01/concept) | [Respan scoring API](https://api.respan.ai/api/v1/scores) (`span-01-pro`) |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | [Liquid decision API](https://api.liquid.ai/decisions/v1/systemone) (`d1:free`) |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | [OpenRouter SystemOne API](https://openrouter.ai/api/v1/systemone) (`liquid/d1`) |
 | [TypeLLM](https://typellm.ai/docs) | [TypeLLM API](https://api.typellm.ai/v1/generate) (`typellm-latest`) |
 | [CodeQL](https://codeql.github.com/) | Local CPU run in Docker; CodeQL CLI 2.27.1 |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | [OpenAI Decisions API](https://api.openai.com/v1/decisions) (`gpt-6-luna`) |
@@ -32,9 +32,9 @@ Pairs come from open-source projects in Rust, JavaScript, TypeScript, Python, Go
 
 ## Reading the results
 
-Each point shows a model at the same cutoff used in its table. X is the false-positive rate: the percentage of clean methods wrongly flagged. Y is recall: the percentage of labelled issues found. Higher and farther left is better.
+Each point shows a model’s score and cost for the same benchmark. Y is Youden’s J: recall minus false-positive rate. X is the run’s inference cost on a logarithmic scale. Higher and farther left is better. The amber Pareto frontier highlights models offering the best score at their price.
 
-The tables use each run's recorded alert settings. A value of 20% under ‘False-positive rate’ means one in five clean methods received an alert. Models are ranked by recall minus that percentage. Bold marks the best full-set value in each column; cost compares priced runs captured here.
+The tables use each run's recorded alert settings. A value of 20% under ‘False-positive rate’ means one in five clean methods received an alert. Models are ranked by Youden’s J. A model that flags every method scores zero. Bold marks the best full-set value in each column; cost compares priced runs captured here.
 
 Failed requests count as incorrect decisions: a failed issue is a miss; a failed clean method is a false positive. Failed cases receive no model score and are included in both the table counts and plotted points.
 
@@ -42,11 +42,11 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 
 [Public pairs](https://huggingface.co/datasets/perchscan/bug-pairs-public)
 
-![Bugs: recall versus false-positive rate at table cutoffs](assets/bug-tradeoff.svg)
+![Bugs: Youden’s J versus inference cost](assets/bug-tradeoff.svg)
 
 ### Public — 2,220 pairs
 
-| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |
+| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |
 | --- | ---: | ---: | ---: | ---: | ---: |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | 48.7% | 38.7% | **10.0%** | **0** | $1.102 |
 | [DiffusionGemma Jev](https://github.com/razorback16/openjev) | 33.4% | 27.1% | 6.3% | **0** | $0.239 |
@@ -56,7 +56,7 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 | [Mercury Decide](https://openrouter.ai/inception/mercury-decide) | 8.8% | 5.6% | 3.2% | 1 | $0.224 |
 | [TypeLLM](https://typellm.ai/docs) | 8.3% | 5.5% | 2.7% | **0** | $0.576 |
 | [Span-01](https://www.respan.ai/docs/documentation/span-01/concept) | 25.3% | 22.6% | 2.7% | **0** | $0.182 |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | 4.1% | 2.5% | 1.6% | **0** | Free tier |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | 5.4% | 3.6% | 1.8% | 1 | $0.875 |
 | [CodeQL](https://codeql.github.com/)* | 5.3% | 4.2% | 1.1% | — | — |
 | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | **81.9%** | 80.9% | 1.0% | **0** | ~$0.210 |
 | [GLiNER-2.5-Decide](https://huggingface.co/fastino/GLiNER2.5-Decide) | 2.3% | **2.1%** | 0.2% | **0** | **$0.175** |
@@ -65,7 +65,7 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 
 ### Private — 498 pairs
 
-| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |
+| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |
 | --- | ---: | ---: | ---: | ---: | ---: |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | 52.4% | 41.3% | **11.1%** | **0** | $0.247 |
 | [DiffusionGemma Jev](https://github.com/razorback16/openjev) | 35.8% | 27.3% | 8.6% | **0** | $0.054 |
@@ -75,7 +75,7 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 | [Mercury Decide](https://openrouter.ai/inception/mercury-decide) | 9.4% | 5.1% | 4.4% | **0** | $0.050 |
 | [Span-01](https://www.respan.ai/docs/documentation/span-01/concept) | 28.5% | 24.3% | 4.2% | **0** | $0.040 |
 | [SemiF](https://github.com/TheoLeeCJ/SemIf) | 44.3% | 41.8% | 2.6% | **0** | $0.054 |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | 5.2% | 3.0% | 2.2% | **0** | Free tier |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | 6.6% | 4.2% | 2.4% | **0** | $0.197 |
 | [GLiNER-2.5-Decide](https://huggingface.co/fastino/GLiNER2.5-Decide) | 1.7% | **2.1%** | -0.4% | **0** | **$0.039** |
 | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | **79.5%** | 80.1% | -0.6% | **0** | ~$0.047 |
 
@@ -83,17 +83,17 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 
 [Public pairs](https://huggingface.co/datasets/perchscan/security-pairs-public)
 
-![Security: recall versus false-positive rate at table cutoffs](assets/security-tradeoff.svg)
+![Security: Youden’s J versus inference cost](assets/security-tradeoff.svg)
 
 ### Public — 1,311 pairs
 
-| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |
+| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |
 | --- | ---: | ---: | ---: | ---: | ---: |
-| [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 34.6% | 23.6% | **11.0%** | **0** | $0.507 |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | 52.3% | 40.7% | **11.5%** | **0** | $6.132 |
+| [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 34.6% | 23.6% | 11.0% | **0** | $0.507 |
 | [TypeLLM](https://typellm.ai/docs) | 27.8% | 17.4% | 10.5% | **0** | $0.566 |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | 54.1% | 44.5% | 9.7% | **0** | $1.469 |
 | [Mercury Decide](https://openrouter.ai/inception/mercury-decide) | 27.1% | 17.8% | 9.3% | **0** | **$0.131** |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | 55.6% | 46.8% | 8.8% | **0** | Free tier |
 | [DiffusionGemma Jev](https://github.com/razorback16/openjev) | 33.1% | 25.8% | 7.3% | **0** | $0.208 |
 | [Microsoft-Decision-1](https://openrouter.ai/microsoft/microsoft-decision-1) | 43.9% | 38.2% | 5.6% | **0** | $0.455 |
 | [Span-01](https://www.respan.ai/docs/documentation/span-01/concept) | 21.0% | 17.2% | 3.9% | **0** | $0.251 |
@@ -106,10 +106,10 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 
 ### Private — 439 pairs
 
-| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |
+| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |
 | --- | ---: | ---: | ---: | ---: | ---: |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | 40.5% | 31.5% | **9.0%** | **0** | $0.530 |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | 34.2% | 29.0% | 5.2% | **0** | Free tier |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | 34.7% | 29.4% | 5.3% | **0** | $2.362 |
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 20.6% | 16.7% | 3.9% | **0** | $0.182 |
 | [DiffusionGemma Jev](https://github.com/razorback16/openjev) | 21.9% | 18.5% | 3.3% | **0** | $0.074 |
 | [TypeLLM](https://typellm.ai/docs) | 8.8% | 6.2% | 2.5% | **0** | $0.203 |
@@ -124,18 +124,18 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 
 [Public pairs](https://huggingface.co/datasets/perchscan/perch-lint-bench-public)
 
-![Perch Lint Bench: recall versus false-positive rate at table cutoffs](assets/lint-tradeoff.svg)
+![Perch Lint Bench: Youden’s J versus inference cost](assets/lint-tradeoff.svg)
 
 ### Public — 214 pairs
 
-| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |
+| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |
 | --- | ---: | ---: | ---: | ---: | ---: |
 | [Mercury Decide](https://openrouter.ai/inception/mercury-decide) | **81.8%** | 6.5% | **75.2%** | **0** | $0.020 |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | 76.6% | 4.7% | 72.0% | **0** | $0.093 |
 | [TypeLLM](https://typellm.ai/docs) | 61.7% | 2.8% | 58.9% | **0** | $0.050 |
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 50.0% | 0.5% | 49.5% | **0** | $0.048 |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | 49.1% | 0.9% | 48.1% | **0** | Free tier |
 | [DiffusionGemma Jev](https://github.com/razorback16/openjev) | 52.8% | 5.1% | 47.7% | **0** | $0.018 |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | 40.7% | 0.5% | 40.2% | **0** | $0.039 |
 | [Span-01](https://www.respan.ai/docs/documentation/span-01/concept) | 50.0% | 11.2% | 38.8% | **0** | $0.014 |
 | [Microsoft-Decision-1](https://openrouter.ai/microsoft/microsoft-decision-1) | 38.3% | 1.9% | 36.4% | **0** | $0.041 |
 | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | 38.8% | 37.4% | 1.4% | **0** | **~$0.010** |
@@ -144,15 +144,15 @@ Failed requests count as incorrect decisions: a failed issue is a miss; a failed
 
 ### Private — 71 pairs
 
-| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |
+| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |
 | --- | ---: | ---: | ---: | ---: | ---: |
 | [Mercury Decide](https://openrouter.ai/inception/mercury-decide) | **87.3%** | 8.5% | **78.9%** | **0** | $0.006 |
 | [GPT-6 Luna Decisions](https://developers.openai.com/api/docs/guides/decisions) | 84.5% | 7.0% | 77.5% | **0** | $0.030 |
 | [TypeLLM](https://typellm.ai/docs) | 59.2% | 5.6% | 53.5% | **0** | $0.016 |
 | [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) | 53.5% | 1.4% | 52.1% | **0** | $0.016 |
-| [Liquid d1](https://docs.liquid.ai/lfm/models/decision-models) | 49.3% | 1.4% | 47.9% | **0** | Free tier |
 | [DiffusionGemma Jev](https://github.com/razorback16/openjev) | 53.5% | 8.5% | 45.1% | **0** | $0.006 |
 | [Microsoft-Decision-1](https://openrouter.ai/microsoft/microsoft-decision-1) | 43.7% | 1.4% | 42.3% | **0** | $0.013 |
+| [Liquid d1](https://openrouter.ai/liquid/d1) | 40.8% | 1.4% | 39.4% | **0** | $0.012 |
 | [Span-01](https://www.respan.ai/docs/documentation/span-01/concept) | 45.1% | 8.5% | 36.6% | **0** | $0.005 |
 | [CLM-v0.1-8B](https://huggingface.co/Contrastive-LM/CLM-v0.1-8B) | 50.7% | 49.3% | 1.4% | **0** | **~$0.003** |
 | [GLiNER-2.5-Decide](https://huggingface.co/fastino/GLiNER2.5-Decide) | 0.0% | **0.0%** | 0.0% | **0** | $0.004 |
@@ -168,23 +168,23 @@ Liquid d1 received the same Perch method state and all applicable questions in o
 
 GLiNER-2.5-Decide scores the answer criteria as native classification labels. Perch sizes each method's context and batches its applicable questions to fit the endpoint. Large methods are read in overlapping source chunks; each question's highest issue score supplies the method decision. Returned label confidences supply the scores. Timeouts and requests that still cannot fit are recorded as failures. Costs use API-reported prompt tokens at the live catalog rate of $0.03 per million input tokens.
 
-Inference costs cover scoring the method sides in each table. Jev costs are estimated from API-reported input tokens at [TypeSafe's published rate](https://typesafe.ai/blog/introducing-system-one-models-and-jev); for semantic lint, input-token usage was measured by replaying the same 570 requests with `jev-1.13.0`. Span-01 costs use API-reported input tokens at [Respan's published rate](https://www.respan.ai/docs/documentation/span-01/concept). Beam costs are API-reported. CLM cost is an estimated share of a 38.4-minute H100 run costing $2.23 total, allocated by embedding count. Costs are not ranked. Liquid d1 was tested on the `d1:free` API tier; a paid per-token price was not published. TypeLLM cost uses API-reported input tokens at its [published rate](https://typellm.ai/pricing); no thinking tokens were used.
+Inference costs cover scoring the method sides in each table. Jev costs are estimated from API-reported input tokens at [TypeSafe's published rate](https://typesafe.ai/blog/introducing-system-one-models-and-jev); for semantic lint, input-token usage was measured by replaying the same 570 requests with `jev-1.13.0`. Span-01 costs use API-reported input tokens at [Respan's published rate](https://www.respan.ai/docs/documentation/span-01/concept). Beam costs are API-reported. CLM cost is an estimated share of a 38.4-minute H100 run costing $2.23 total, allocated by embedding count. Costs are not ranked. TypeLLM cost uses API-reported input tokens at its [published rate](https://typellm.ai/pricing); no thinking tokens were used.
 
 JevLike (192-byte input), Laya (512-token input), and Kev (packed-branch limit) could not score every full-context method. NanoJev was unavailable on Beam. They are excluded rather than ranked on partial results.
 
 GPT-6 Luna Decisions receives each method's saved context with all applicable checks as two-option choice questions in one request. True and false criteria define the options. Semantic lint scores invert the probability that the rule is satisfied. Costs use API-reported input tokens at [$0.10 per million](https://developers.openai.com/api/docs/guides/decisions#pricing-and-availability); output and cache operations carry no extra charge.
 
-Mercury Decide and Microsoft-Decision-1 receive each method's saved context with the applicable checks through OpenRouter's native SystemOne interface. Their returned probabilities use the same question floors and semantic lint cutoff. Costs come from the API's reported `usage.cost`.
+Liquid d1, Mercury Decide and Microsoft-Decision-1 receive each method's saved context with the applicable checks through OpenRouter's native SystemOne interface. Their returned probabilities use the same question floors and semantic lint cutoff. Costs come from the API's reported `usage.cost`.
 
 ## Files and reproduction
 
 - [`results-current.jsonl`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/results-current.jsonl): current-pack aggregate results for public and private splits; no private pair data.
 - [`results.jsonl`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/results.jsonl): all recorded result versions, identified by run and protocol ID.
 - [`predictions/current/`](https://huggingface.co/datasets/perchscan/benchmark-results/tree/main/predictions/current): public method probabilities from this run. Private predictions are not published.
-- [`tradeoff.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/tradeoff.py): rebuilds the public recall/false-alert curves and result tables from saved scores and labels.
+- [`tradeoff.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/tradeoff.py): rebuilds the public Youden’s J versus cost plots, Pareto frontiers and result tables from saved outcomes.
 - [`benchmark.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/benchmark.py) and [`question-pack.json`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/question-pack.json): reproducible current-pack scorer and frozen questions.
 - [`span01.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/span01.py): calls Span-01 with all applicable behaviors in one request per method, saves resumable predictions, and scores the completed dataset.
-- [`liquid.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/liquid.py): calls Liquid d1 with the frozen Perch question pack, saves resumable predictions, and scores each track.
+- [`liquid.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/liquid.py): original Liquid direct-API scorer with the frozen Perch question pack.
 - [`typellm.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/typellm.py): scores each method independently with applicable boolean questions and saves the returned probabilities.
 - [`gliner_decide.py`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/gliner_decide.py): runs GLiNER-2.5-Decide through its native classification endpoint, saves returned confidences and failed requests, and scores each completed track.
 - [`perch-sizing.mjs`](https://huggingface.co/datasets/perchscan/benchmark-results/blob/main/perch-sizing.mjs): uses Perch's source chunking, context sizing, and question batching. Each reading records its source-byte coverage.
@@ -230,10 +230,11 @@ are recorded as failed requests, without a probability score. The six-set local
 run used 24 workers and a request limiter of 1.5 million estimated input tokens
 per minute. Recorded method elapsed times include client throttling and rate-limit backoff. The public manifest contains only the three public sets.
 
-To run Mercury Decide or Microsoft-Decision-1, download the three public files into the same directories listed above, install `httpx`, and set `OPENROUTER_API_KEY`:
+To run Liquid d1, Mercury Decide or Microsoft-Decision-1, download the three public files into the same directories listed above, install `httpx`, and set `OPENROUTER_API_KEY`:
 
 ```sh
 python3 -m pip install httpx
+python3 openrouter_decisions.py --model liquid/d1 --manifest openrouter-decisions-manifest.json --out liquid-results
 python3 openrouter_decisions.py --model inception/mercury-decide --manifest openrouter-decisions-manifest.json --out mercury-results
 python3 openrouter_decisions.py --model microsoft/microsoft-decision-1 --manifest openrouter-decisions-manifest.json --out microsoft-results
 ```

openrouter_decisions.py

@@ -27,6 +27,8 @@ CATALOG = {
         'mercury-decide-perch-systemone-v1', 'mercury-decide-20261009', 'Inception', 65536, 0),
     'inception/mercury-decide': ('mercury-decide', 'inception/mercury-decide-20260930',
         'mercury-decide-perch-systemone-v1', 'mercury-decide-20261009', 'Inception', 65536, .02),
+    'liquid/d1': ('liquid-d1', 'liquid/d1-20260930',
+        'liquid-d1-openrouter-perch-systemone-v1', 'liquid-d1-openrouter-20261012', 'Liquid', 65536, .04),
     'microsoft/microsoft-decision-1': ('microsoft-decision-1', 'microsoft/microsoft-decision-1-20261009',
         'microsoft-decision-1-perch-systemone-v1', 'microsoft-decision-1-20261012', 'Azure', 32768, .042),
 }
@@ -270,12 +272,13 @@ def main():
     iterator, attempted = iter(pending), 0
     reserved = 0.
     budget_paused = False
+    account_error = None
     with ThreadPoolExecutor(max_workers=args.workers) as pool:
         active = {}
 
         def submit():
             nonlocal reserved, budget_paused
-            if (args.budget is not None and spent >= args.budget) or budget_paused:
+            if account_error or (args.budget is not None and spent >= args.budget) or budget_paused:
                 return
             job = next(iterator, None)
             if job:
@@ -299,10 +302,11 @@ def main():
                     identity = pair['pair_id'], role
                     try:
                         row = future.result()
-                    except AccountError:
-                        for f in active:
-                            f.cancel()
-                        raise
+                    except AccountError as exc:
+                        # Stop new calls, but save successful calls already in flight.
+                        # Quota-blocked methods remain pending for resumption.
+                        account_error = account_error or exc
+                        continue
                     except Exception as exc:
                         row = {'pair_id': pair['pair_id'], 'role': role, 'domain': group['domain'],
                                'model_revision': REVISION, 'requested_model': args.model, 'protocol_id': PROTOCOL,
@@ -327,6 +331,8 @@ def main():
                     submit()
         finally:
             save_reports()
+    if account_error:
+        raise account_error
     print('Budget guard paused pending methods.' if budget_paused else 'Finished scheduled methods.', flush=True)
 
 

results-current.jsonl

@@ -28,12 +28,6 @@
 {"alert_rule": "rule violation score > 0.7", "average_precision": 0.8296819071061596, "benchmark_methods": 142, "benchmark_pairs": 71, "confusion": {"fn": 33, "fp": 6, "tn": 65, "tp": 38}, "domain": "lint", "f1": 0.6608695652173913, "inference_cost_basis": "API reported", "inference_cost_usd": 0.005866, "inference_endpoint": "https://app.beam.cloud/v1/models/jev/diffusiongemma/invoke", "inference_host": "Beam", "model_id": "diffusiongemma", "model_revision": "diffusiongemma", "pair_order": 0.8591549295774648, "pair_order_correct": 61, "pair_order_eligible": 71, "precision": 0.8636363636363636, "protocol_id": "perch-semantic-ensure-v1", "question_pack_sha256": null, "recall": 0.5352112676056338, "run_id": "perch-scan-pack-configured-floors-20260928", "threshold": 0.7, "track": "private"}
 {"alert_rule": "rule violation score > 0.7", "average_precision": 0.5141925432023778, "benchmark_methods": 142, "benchmark_pairs": 71, "confusion": {"fn": 35, "fp": 35, "tn": 36, "tp": 36}, "domain": "lint", "f1": 0.5070422535211268, "inference_cost_basis": "allocated GPU runtime by embedding rows", "inference_cost_usd": 0.0033639277067493522, "inference_endpoint": null, "inference_host": "Runpod pod, H100 80 GB, US-NE-1", "model_id": "clm", "model_revision": "Contrastive-LM/CLM-v0.1-8B@head-b2b4a8c9c2d39263eff78a351eb909a342ce9b3bf21a3f07c1d1bf15f1c4eda5", "pair_order": 0.5633802816901409, "pair_order_correct": 40, "pair_order_eligible": 71, "precision": 0.5070422535211268, "protocol_id": "clm-choice-head-v1", "question_pack_sha256": null, "recall": 0.5070422535211268, "run_id": "perch-scan-pack-configured-floors-20260928", "threshold": 0.7, "track": "private"}
 {"alert_rule": "rule violation score > 0.7", "average_precision": 0.8320090658208488, "benchmark_methods": 142, "benchmark_pairs": 71, "confusion": {"fn": 39, "fp": 6, "tn": 65, "tp": 32}, "domain": "lint", "f1": 0.5871559633027523, "inference_cost_basis": "API-reported input tokens at published rate", "inference_cost_usd": 0.00456932, "inference_endpoint": "https://api.respan.ai/api/v1/scores", "inference_host": "Respan API", "model_id": "span01", "model_revision": "span-01-pro (API revision not published)", "pair_order": 0.9154929577464789, "pair_order_correct": 65, "pair_order_eligible": 71, "precision": 0.8421052631578947, "protocol_id": "respan-span01-pro-perch-pack-v1", "question_pack_sha256": null, "recall": 0.4507042253521127, "run_id": "perch-scan-pack-configured-floors-20260928", "threshold": 0.7, "track": "private"}
-{"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.46569203965559486, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "confusion": {"fn": 1785, "fp": 65, "tn": 2514, "tp": 76}, "domain": "bug", "f1": 0.07592407592407592, "inference_cost_basis": "d1:free API tier; no published per-token rate", "inference_cost_usd": null, "inference_endpoint": "https://api.liquid.ai/decisions/v1/systemone", "inference_host": "Liquid AI API", "input_tokens": 44836228, "model_id": "liquid-d1", "model_revision": "d1:free (API revision not published)", "output_tokens": 0, "pair_order": 0.5011223344556678, "pair_order_correct": 893, "pair_order_eligible": 1782, "precision": 0.5390070921985816, "protocol_id": "perch-scan-afe116cc8461", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.04083825900053734, "run_id": "perch-bug-two-questions-20260929", "threshold": 0.5, "track": "public"}
-{"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.48689814906750656, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "confusion": {"fn": 402, "fp": 17, "tn": 555, "tp": 22}, "domain": "bug", "f1": 0.09503239740820735, "inference_cost_basis": "d1:free API tier; no published per-token rate", "inference_cost_usd": null, "inference_endpoint": "https://api.liquid.ai/decisions/v1/systemone", "inference_host": "Liquid AI API", "input_tokens": 10077688, "model_id": "liquid-d1", "model_revision": "d1:free (API revision not published)", "output_tokens": 0, "pair_order": 0.49625935162094764, "pair_order_correct": 199, "pair_order_eligible": 401, "precision": 0.5641025641025641, "protocol_id": "perch-scan-afe116cc8461", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.05188679245283019, "run_id": "perch-bug-two-questions-20260929", "threshold": 0.5, "track": "private"}
-{"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.49359009471673904, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "confusion": {"fn": 507, "fp": 693, "tn": 787, "tp": 635}, "domain": "security", "f1": 0.5141700404858299, "inference_cost_basis": "d1:free API tier; no published per-token rate", "inference_cost_usd": null, "inference_endpoint": "https://api.liquid.ai/decisions/v1/systemone", "inference_host": "Liquid AI API", "input_tokens": 161612859, "model_id": "liquid-d1", "model_revision": "d1:free (API revision not published)", "output_tokens": 0, "pair_order": 0.5222419928825622, "pair_order_correct": 587, "pair_order_eligible": 1124, "precision": 0.47816265060240964, "protocol_id": "liquid-d1-perch-scan-pack-v1", "question_pack_sha256": "bb53c8a6ca30920140e2bf33b2905d159e7b1709b29d1ea896637905dbf9047d", "recall": 0.5560420315236427, "run_id": "liquid-d1-perch-scan-20260929", "threshold": 0.5, "track": "public"}
-{"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.4784904017284298, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "confusion": {"fn": 262, "fp": 139, "tn": 341, "tp": 136}, "domain": "security", "f1": 0.40416047548291234, "inference_cost_basis": "d1:free API tier; no published per-token rate", "inference_cost_usd": null, "inference_endpoint": "https://api.liquid.ai/decisions/v1/systemone", "inference_host": "Liquid AI API", "input_tokens": 62054285, "model_id": "liquid-d1", "model_revision": "d1:free (API revision not published)", "output_tokens": 0, "pair_order": 0.4547738693467337, "pair_order_correct": 181, "pair_order_eligible": 398, "precision": 0.49454545454545457, "protocol_id": "liquid-d1-perch-scan-pack-v1", "question_pack_sha256": "bb53c8a6ca30920140e2bf33b2905d159e7b1709b29d1ea896637905dbf9047d", "recall": 0.3417085427135678, "run_id": "liquid-d1-perch-scan-20260929", "threshold": 0.5, "track": "private"}
-{"alert_rule": "rule violation score > 0.7", "average_precision": 0.9251948043796072, "benchmark_methods": 428, "benchmark_pairs": 214, "benchmark_sha256": "534f6c2abddc507c40f4597607e97e07168c544011d617b83aa39e08abb35d75", "confusion": {"fn": 109, "fp": 2, "tn": 212, "tp": 105}, "domain": "lint", "f1": 0.6542056074766355, "inference_cost_basis": "d1:free API tier; no published per-token rate", "inference_cost_usd": null, "inference_endpoint": "https://api.liquid.ai/decisions/v1/systemone", "inference_host": "Liquid AI API", "input_tokens": 985496, "model_id": "liquid-d1", "model_revision": "d1:free (API revision not published)", "output_tokens": 0, "pair_order": 0.9345794392523364, "pair_order_correct": 200, "pair_order_eligible": 214, "precision": 0.9813084112149533, "protocol_id": "liquid-d1-perch-semantic-ensure-v1", "question_pack_sha256": null, "recall": 0.49065420560747663, "run_id": "liquid-d1-perch-scan-20260929", "threshold": 0.7, "track": "public"}
-{"alert_rule": "rule violation score > 0.7", "average_precision": 0.9239761802875874, "benchmark_methods": 142, "benchmark_pairs": 71, "benchmark_sha256": "a6e9f2e4ee8ae77bdaf23af46920d97d1ba47375dc9ce03b950cdbe95c2777f9", "confusion": {"fn": 36, "fp": 1, "tn": 70, "tp": 35}, "domain": "lint", "f1": 0.6542056074766355, "inference_cost_basis": "d1:free API tier; no published per-token rate", "inference_cost_usd": null, "inference_endpoint": "https://api.liquid.ai/decisions/v1/systemone", "inference_host": "Liquid AI API", "input_tokens": 315606, "model_id": "liquid-d1", "model_revision": "d1:free (API revision not published)", "output_tokens": 0, "pair_order": 0.971830985915493, "pair_order_correct": 69, "pair_order_eligible": 71, "precision": 0.9722222222222222, "protocol_id": "liquid-d1-perch-semantic-ensure-v1", "question_pack_sha256": null, "recall": 0.49295774647887325, "run_id": "liquid-d1-perch-scan-20260929", "threshold": 0.7, "track": "private"}
 {"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.4730101495499027, "benchmark_methods": 4440, "benchmark_pairs": 2220, "confusion": {"fn": 1707, "fp": 143, "tn": 2436, "tp": 154}, "domain": "bug", "f1": 0.14272474513438368, "inference_cost_basis": "API-reported input and thinking tokens at published TypeLLM rates", "inference_cost_usd": 0.57606375, "inference_endpoint": "https://api.typellm.ai/v1/generate", "inference_host": "TypeLLM", "model_id": "typellm", "model_revision": "typellm-latest", "pair_order": 0.5129068462401796, "pair_order_correct": 914, "pair_order_eligible": 1782, "precision": 0.5185185185185185, "protocol_id": "perch-scan-afe116cc8461", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.08275120902740463, "run_id": "typellm-20260930", "threshold": 0.5, "track": "public"}
 {"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.5064587972284842, "benchmark_methods": 996, "benchmark_pairs": 498, "confusion": {"fn": 376, "fp": 33, "tn": 539, "tp": 48}, "domain": "bug", "f1": 0.1900990099009901, "inference_cost_basis": "API-reported input and thinking tokens at published TypeLLM rates", "inference_cost_usd": 0.12900565, "inference_endpoint": "https://api.typellm.ai/v1/generate", "inference_host": "TypeLLM", "model_id": "typellm", "model_revision": "typellm-latest", "pair_order": 0.5511221945137157, "pair_order_correct": 221, "pair_order_eligible": 401, "precision": 0.5925925925925926, "protocol_id": "perch-scan-afe116cc8461", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.11320754716981132, "run_id": "typellm-20260930", "threshold": 0.5, "track": "private"}
 {"alert_rule": "any issue score > max(run threshold, question min)", "average_precision": 0.5078855584544545, "benchmark_methods": 2622, "benchmark_pairs": 1311, "confusion": {"fn": 824, "fp": 257, "tn": 1223, "tp": 318}, "domain": "security", "f1": 0.37041351193942923, "inference_cost_basis": "API-reported input and thinking tokens at published TypeLLM rates", "inference_cost_usd": 0.566494, "inference_endpoint": "https://api.typellm.ai/v1/generate", "inference_host": "TypeLLM", "model_id": "typellm", "model_revision": "typellm-latest", "pair_order": 0.5195729537366548, "pair_order_correct": 584, "pair_order_eligible": 1124, "precision": 0.5530434782608695, "protocol_id": "perch-scan-afe116cc8461", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.27845884413309985, "run_id": "typellm-20260930", "threshold": 0.5, "track": "public"}
@@ -64,3 +58,9 @@
 {"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 996, "average_precision": 0.4664174064351087, "average_precision_answered_methods": 0.4664174064351087, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "complete": true, "confusion": {"fn": 316, "fp": 118, "tn": 454, "tp": 108}, "context_limit_tokens": 32768, "domain": "bug", "f1": 0.3323076923076923, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.2062937062937063, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.107653602, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Azure", "input_tokens": 2563181, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.873144334065728, "latency_p95_seconds": 2.338826625025831, "model_id": "microsoft-decision-1", "model_revision": "microsoft/microsoft-decision-1-20261009", "pair_order": 0.5511221945137157, "pair_order_correct": 221, "pair_order_eligible": 401, "precision": 0.4778761061946903, "protocol_id": "microsoft-decision-1-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.25471698113207547, "refused_methods": 0, "requested_model": "microsoft/microsoft-decision-1", "run_id": "microsoft-decision-1-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 221, "eligible": 401, "ties": 0}}
 {"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 2622, "average_precision": 0.4737128965738396, "average_precision_answered_methods": 0.4737128965738396, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "complete": true, "confusion": {"fn": 641, "fp": 566, "tn": 914, "tp": 501}, "context_limit_tokens": 32768, "domain": "security", "f1": 0.45359891353553644, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.3824324324324324, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.455216076, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Azure", "input_tokens": 10838478, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.028734000050463, "latency_p95_seconds": 2.333685915917158, "model_id": "microsoft-decision-1", "model_revision": "microsoft/microsoft-decision-1-20261009", "pair_order": 0.5266903914590747, "pair_order_correct": 592, "pair_order_eligible": 1124, "precision": 0.46954076850984067, "protocol_id": "microsoft-decision-1-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.43870402802101577, "refused_methods": 0, "requested_model": "microsoft/microsoft-decision-1", "run_id": "microsoft-decision-1-20261012", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 592, "eligible": 1124, "ties": 0}}
 {"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 878, "average_precision": 0.46925395820871746, "average_precision_answered_methods": 0.46925395820871746, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "complete": true, "confusion": {"fn": 326, "fp": 80, "tn": 400, "tp": 72}, "context_limit_tokens": 32768, "domain": "security", "f1": 0.26181818181818184, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.16666666666666666, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.162594978, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Azure", "input_tokens": 3871309, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.0353772909147665, "latency_p95_seconds": 2.391944915987551, "model_id": "microsoft-decision-1", "model_revision": "microsoft/microsoft-decision-1-20261009", "pair_order": 0.4824120603015075, "pair_order_correct": 192, "pair_order_eligible": 398, "precision": 0.47368421052631576, "protocol_id": "microsoft-decision-1-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.18090452261306533, "refused_methods": 0, "requested_model": "microsoft/microsoft-decision-1", "run_id": "microsoft-decision-1-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 192, "eligible": 398, "ties": 0}}
+{"alert_rule": "rule violation probability > 0.7", "answered_methods": 428, "average_precision": 0.9194982168505601, "average_precision_answered_methods": 0.9194982168505601, "benchmark_methods": 428, "benchmark_pairs": 214, "benchmark_sha256": "534f6c2abddc507c40f4597607e97e07168c544011d617b83aa39e08abb35d75", "complete": true, "confusion": {"fn": 127, "fp": 1, "tn": 213, "tp": 87}, "context_limit_tokens": 65536, "domain": "lint", "f1": 0.5761589403973509, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.004672897196261682, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.03890624, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 972656, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.7922159159788862, "latency_p95_seconds": 2.426701958058402, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.9158878504672897, "pair_order_correct": 196, "pair_order_eligible": 214, "precision": 0.9886363636363636, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": null, "recall": 0.40654205607476634, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.7, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 196, "eligible": 214, "ties": 0}}
+{"alert_rule": "rule violation probability > 0.7", "answered_methods": 142, "average_precision": 0.9038197230691274, "average_precision_answered_methods": 0.9038197230691274, "benchmark_methods": 142, "benchmark_pairs": 71, "benchmark_sha256": "a6e9f2e4ee8ae77bdaf23af46920d97d1ba47375dc9ce03b950cdbe95c2777f9", "complete": true, "confusion": {"fn": 42, "fp": 1, "tn": 70, "tp": 29}, "context_limit_tokens": 65536, "domain": "lint", "f1": 0.5742574257425742, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.014084507042253521, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.01245384, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 311346, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.8404922089539468, "latency_p95_seconds": 2.41227475006599, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.9154929577464789, "pair_order_correct": 65, "pair_order_eligible": 71, "precision": 0.9666666666666667, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": null, "recall": 0.4084507042253521, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.7, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 65, "eligible": 71, "ties": 1}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 4439, "average_precision": 0.45826337265301537, "average_precision_answered_methods": 0.45826337265301537, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "complete": true, "confusion": {"fn": 1760, "fp": 93, "tn": 2486, "tp": 101}, "context_limit_tokens": 65536, "domain": "bug", "f1": 0.09829683698296837, "failed_methods": 1, "failure_counts": {"The read operation timed out": 1}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.03606048856145793, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.87490476, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 21872619, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.887577582965605, "latency_p95_seconds": 2.3768632920691743, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.5016835016835017, "pair_order_correct": 894, "pair_order_eligible": 1782, "precision": 0.520618556701031, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.05427189682966147, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 894, "eligible": 1782, "ties": 14}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 996, "average_precision": 0.4758656734035225, "average_precision_answered_methods": 0.4758656734035225, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "complete": true, "confusion": {"fn": 396, "fp": 24, "tn": 548, "tp": 28}, "context_limit_tokens": 65536, "domain": "bug", "f1": 0.11764705882352941, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.04195804195804196, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.19667368, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 4916842, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.7929241659585387, "latency_p95_seconds": 2.340836458024569, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.5062344139650873, "pair_order_correct": 203, "pair_order_eligible": 401, "precision": 0.5384615384615384, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.0660377358490566, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 203, "eligible": 401, "ties": 5}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 2622, "average_precision": 0.49708136950251325, "average_precision_answered_methods": 0.49708136950251325, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "complete": true, "confusion": {"fn": 545, "fp": 603, "tn": 877, "tp": 597}, "context_limit_tokens": 65536, "domain": "security", "f1": 0.5098206660973527, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.40743243243243243, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 6.13173568, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 153293392, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.0774508339818567, "latency_p95_seconds": 2.837762249982916, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.5284697508896797, "pair_order_correct": 594, "pair_order_eligible": 1124, "precision": 0.4975, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.5227670753064798, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 594, "eligible": 1124, "ties": 15}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 878, "average_precision": 0.4793341864639895, "average_precision_answered_methods": 0.4793341864639895, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "complete": true, "confusion": {"fn": 260, "fp": 141, "tn": 339, "tp": 138}, "context_limit_tokens": 65536, "domain": "security", "f1": 0.4076809453471196, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.29375, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 2.36226772, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 59056693, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.1423684580950066, "latency_p95_seconds": 2.9545628749765456, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.4547738693467337, "pair_order_correct": 181, "pair_order_eligible": 398, "precision": 0.4946236559139785, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.34673366834170855, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 181, "eligible": 398, "ties": 4}}

results.jsonl

@@ -110,3 +110,9 @@
 {"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 996, "average_precision": 0.4664174064351087, "average_precision_answered_methods": 0.4664174064351087, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "complete": true, "confusion": {"fn": 316, "fp": 118, "tn": 454, "tp": 108}, "context_limit_tokens": 32768, "domain": "bug", "f1": 0.3323076923076923, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.2062937062937063, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.107653602, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Azure", "input_tokens": 2563181, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.873144334065728, "latency_p95_seconds": 2.338826625025831, "model_id": "microsoft-decision-1", "model_revision": "microsoft/microsoft-decision-1-20261009", "pair_order": 0.5511221945137157, "pair_order_correct": 221, "pair_order_eligible": 401, "precision": 0.4778761061946903, "protocol_id": "microsoft-decision-1-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.25471698113207547, "refused_methods": 0, "requested_model": "microsoft/microsoft-decision-1", "run_id": "microsoft-decision-1-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 221, "eligible": 401, "ties": 0}}
 {"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 2622, "average_precision": 0.4737128965738396, "average_precision_answered_methods": 0.4737128965738396, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "complete": true, "confusion": {"fn": 641, "fp": 566, "tn": 914, "tp": 501}, "context_limit_tokens": 32768, "domain": "security", "f1": 0.45359891353553644, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.3824324324324324, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.455216076, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Azure", "input_tokens": 10838478, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.028734000050463, "latency_p95_seconds": 2.333685915917158, "model_id": "microsoft-decision-1", "model_revision": "microsoft/microsoft-decision-1-20261009", "pair_order": 0.5266903914590747, "pair_order_correct": 592, "pair_order_eligible": 1124, "precision": 0.46954076850984067, "protocol_id": "microsoft-decision-1-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.43870402802101577, "refused_methods": 0, "requested_model": "microsoft/microsoft-decision-1", "run_id": "microsoft-decision-1-20261012", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 592, "eligible": 1124, "ties": 0}}
 {"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 878, "average_precision": 0.46925395820871746, "average_precision_answered_methods": 0.46925395820871746, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "complete": true, "confusion": {"fn": 326, "fp": 80, "tn": 400, "tp": 72}, "context_limit_tokens": 32768, "domain": "security", "f1": 0.26181818181818184, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.16666666666666666, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.162594978, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Azure", "input_tokens": 3871309, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.0353772909147665, "latency_p95_seconds": 2.391944915987551, "model_id": "microsoft-decision-1", "model_revision": "microsoft/microsoft-decision-1-20261009", "pair_order": 0.4824120603015075, "pair_order_correct": 192, "pair_order_eligible": 398, "precision": 0.47368421052631576, "protocol_id": "microsoft-decision-1-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.18090452261306533, "refused_methods": 0, "requested_model": "microsoft/microsoft-decision-1", "run_id": "microsoft-decision-1-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 192, "eligible": 398, "ties": 0}}
+{"alert_rule": "rule violation probability > 0.7", "answered_methods": 428, "average_precision": 0.9194982168505601, "average_precision_answered_methods": 0.9194982168505601, "benchmark_methods": 428, "benchmark_pairs": 214, "benchmark_sha256": "534f6c2abddc507c40f4597607e97e07168c544011d617b83aa39e08abb35d75", "complete": true, "confusion": {"fn": 127, "fp": 1, "tn": 213, "tp": 87}, "context_limit_tokens": 65536, "domain": "lint", "f1": 0.5761589403973509, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.004672897196261682, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.03890624, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 972656, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.7922159159788862, "latency_p95_seconds": 2.426701958058402, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.9158878504672897, "pair_order_correct": 196, "pair_order_eligible": 214, "precision": 0.9886363636363636, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": null, "recall": 0.40654205607476634, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.7, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 196, "eligible": 214, "ties": 0}}
+{"alert_rule": "rule violation probability > 0.7", "answered_methods": 142, "average_precision": 0.9038197230691274, "average_precision_answered_methods": 0.9038197230691274, "benchmark_methods": 142, "benchmark_pairs": 71, "benchmark_sha256": "a6e9f2e4ee8ae77bdaf23af46920d97d1ba47375dc9ce03b950cdbe95c2777f9", "complete": true, "confusion": {"fn": 42, "fp": 1, "tn": 70, "tp": 29}, "context_limit_tokens": 65536, "domain": "lint", "f1": 0.5742574257425742, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.014084507042253521, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.01245384, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 311346, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.8404922089539468, "latency_p95_seconds": 2.41227475006599, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.9154929577464789, "pair_order_correct": 65, "pair_order_eligible": 71, "precision": 0.9666666666666667, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": null, "recall": 0.4084507042253521, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.7, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 65, "eligible": 71, "ties": 1}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 4439, "average_precision": 0.45826337265301537, "average_precision_answered_methods": 0.45826337265301537, "benchmark_methods": 4440, "benchmark_pairs": 2220, "benchmark_sha256": "32bc68bc85d5535ad28e103adbfa0c09c73e09d0f53a53f7ca35ec219a96e6e2", "complete": true, "confusion": {"fn": 1760, "fp": 93, "tn": 2486, "tp": 101}, "context_limit_tokens": 65536, "domain": "bug", "f1": 0.09829683698296837, "failed_methods": 1, "failure_counts": {"The read operation timed out": 1}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.03606048856145793, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.87490476, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 21872619, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.887577582965605, "latency_p95_seconds": 2.3768632920691743, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.5016835016835017, "pair_order_correct": 894, "pair_order_eligible": 1782, "precision": 0.520618556701031, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.05427189682966147, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 894, "eligible": 1782, "ties": 14}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 996, "average_precision": 0.4758656734035225, "average_precision_answered_methods": 0.4758656734035225, "benchmark_methods": 996, "benchmark_pairs": 498, "benchmark_sha256": "3df4ade65fedd1128d30bc17c9c9869c8ab9bd526e666156d0e16685e78f8a73", "complete": true, "confusion": {"fn": 396, "fp": 24, "tn": 548, "tp": 28}, "context_limit_tokens": 65536, "domain": "bug", "f1": 0.11764705882352941, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.04195804195804196, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 0.19667368, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 4916842, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 1.7929241659585387, "latency_p95_seconds": 2.340836458024569, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.5062344139650873, "pair_order_correct": 203, "pair_order_eligible": 401, "precision": 0.5384615384615384, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.0660377358490566, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 203, "eligible": 401, "ties": 5}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 2622, "average_precision": 0.49708136950251325, "average_precision_answered_methods": 0.49708136950251325, "benchmark_methods": 2622, "benchmark_pairs": 1311, "benchmark_sha256": "2407059caa3486371749884138b997d89043eec646e548ee8f44ba3ab08514b8", "complete": true, "confusion": {"fn": 545, "fp": 603, "tn": 877, "tp": 597}, "context_limit_tokens": 65536, "domain": "security", "f1": 0.5098206660973527, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.40743243243243243, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 6.13173568, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 153293392, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.0774508339818567, "latency_p95_seconds": 2.837762249982916, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.5284697508896797, "pair_order_correct": 594, "pair_order_eligible": 1124, "precision": 0.4975, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.5227670753064798, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "public", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 594, "eligible": 1124, "ties": 15}}
+{"alert_rule": "any issue score > max(0.5, question min)", "answered_methods": 878, "average_precision": 0.4793341864639895, "average_precision_answered_methods": 0.4793341864639895, "benchmark_methods": 878, "benchmark_pairs": 439, "benchmark_sha256": "a4b81728e9e520cf1c6f342974bec369d72f5b7d8aabc1c9cf2aceba966f8051", "complete": true, "confusion": {"fn": 260, "fp": 141, "tn": 339, "tp": 138}, "context_limit_tokens": 65536, "domain": "security", "f1": 0.4076809453471196, "failed_methods": 0, "failure_counts": {}, "failure_counts_per_question": {}, "failure_policy": "failed positive counts as FN; failed negative counts as FP; failed pair counts as incorrect ordering", "false_positive_rate": 0.29375, "inference_cost_basis": "OpenRouter API-reported usage.cost", "inference_cost_usd": 2.36226772, "inference_endpoint": "https://openrouter.ai/api/v1/systemone", "inference_host": "OpenRouter / Liquid", "input_tokens": 59056693, "latency_basis": "Per-method elapsed time, including client rate-limit waits and 429 backoff", "latency_median_seconds": 2.1423684580950066, "latency_p95_seconds": 2.9545628749765456, "model_id": "liquid-d1", "model_revision": "liquid/d1-20260930", "pair_order": 0.4547738693467337, "pair_order_correct": 181, "pair_order_eligible": 398, "precision": 0.4946236559139785, "protocol_id": "liquid-d1-openrouter-perch-systemone-v1", "question_pack_sha256": "afe116cc8461552aaca996f8f0b639cdf5e42e9d8291c59995482309b5c330dc", "recall": 0.34673366834170855, "refused_methods": 0, "requested_model": "liquid/d1", "run_id": "liquid-d1-openrouter-20261012", "threshold": 0.5, "track": "private", "vulnerable_or_buggy_before_above_fixed_after": {"correct": 181, "eligible": 398, "ties": 4}}

test_openrouter_decisions.py

@@ -33,6 +33,21 @@ class OpenRouterTests(unittest.TestCase):
         finally:
             runner.select_model(runner.MODELS[0])
 
+    def test_liquid_paid_route_records_native_probability_and_cost(self):
+        model = 'liquid/d1'
+        runner.select_model(model)
+        try:
+            response = self.response({'semantic': {'type': 'noul', 'noul': .1}})
+            response['usage']['cost'] = .00004
+            row = runner.prediction(self.pair, 'before', 'lint', self.pack, response, .2, model)
+            self.assertEqual(row['model_id'], 'liquid-d1')
+            self.assertEqual(row['model_revision'], 'liquid/d1-20260930')
+            self.assertEqual(row['usage']['cost'], .00004)
+            self.assertAlmostEqual(row['score'], .9)
+            self.assertEqual(runner.CONTEXT_LIMIT, 65536)
+        finally:
+            runner.select_model(runner.MODELS[0])
+
     def test_native_schema_and_semantic_polarity(self):
         payload = runner.payload_for(self.pair, 'before', 'lint', self.pack, self.model)
         self.assertEqual(payload['state'], self.pair['before']['state'])

test_tradeoff.py

@@ -7,34 +7,63 @@ import tradeoff
 
 
 class GraphTests(unittest.TestCase):
+    def rows(self):
+        return {model: {'confusion': {'tp': 7, 'fn': 3, 'fp': 2, 'tn': 18},
+                        'inference_cost_usd': .05, 'benchmark_methods': 30}
+                for model in tradeoff.ORDER}
+
     def render(self, current):
         with tempfile.TemporaryDirectory() as temp:
             output = Path(temp) / 'graph.svg'
             tradeoff.svg('bug', current, output)
             return ET.fromstring(output.read_text())
 
-    def test_points_use_recall_and_false_positive_rate_from_table(self):
-        current = {model: {'confusion': {'tp': 7, 'fn': 3, 'fp': 2, 'tn': 18}}
-                   for model in tradeoff.ORDER}
-        root = self.render(current)
-        dots = [e for e in root.findall('{http://www.w3.org/2000/svg}circle') if 'data-model' in e.attrib]
-        self.assertEqual(len(dots), len(tradeoff.ORDER))
-        self.assertAlmostEqual((float(dots[0].attrib['cx']) - 90) / 990, .1)
-        self.assertAlmostEqual((498 - float(dots[0].attrib['cy'])) / 350, .7)
-        self.assertFalse(root.findall('{http://www.w3.org/2000/svg}polyline'))
-        labels = [e.text for e in root.findall('{http://www.w3.org/2000/svg}text')]
-        self.assertIn('False positive rate', labels)
-        self.assertIn('Recall', labels)
-
-    def test_false_positives_do_not_change_vertical_recall_coordinate(self):
-        current = {model: {'confusion': {'tp': 7, 'fn': 3, 'fp': 2, 'tn': 18}}
-                   for model in tradeoff.ORDER}
-        current['typellm'] = {'confusion': {'tp': 7, 'fn': 3, 'fp': 10, 'tn': 10}}
-        root = self.render(current)
-        dots = {e.attrib['data-model']: e for e in root.findall('{http://www.w3.org/2000/svg}circle')
+    def dots(self, root):
+        return {e.attrib['data-model']: e for e in root.findall('{http://www.w3.org/2000/svg}circle')
                 if 'data-model' in e.attrib}
-        self.assertEqual(dots['jev'].attrib['cy'], dots['typellm'].attrib['cy'])
-        self.assertLess(float(dots['jev'].attrib['cx']), float(dots['typellm'].attrib['cx']))
+
+    def test_points_use_recorded_cost_and_youden_j(self):
+        import math
+        current = self.rows();root = self.render(current);dots = self.dots(root)
+        self.assertEqual(len(dots), len(current))
+        dot=dots['jev'];self.assertEqual(float(dot.attrib['data-cost']), .05)
+        self.assertAlmostEqual(float(dot.attrib['data-j']), .6)
+        low,high=float(root.attrib['data-cost-low']),float(root.attrib['data-cost-high'])
+        y_min,y_max=float(root.attrib['data-j-min']),float(root.attrib['data-j-max'])
+        self.assertAlmostEqual((float(dot.attrib['cx'])-100)/1130,
+            (math.log10(.05)-math.log10(low))/(math.log10(high)-math.log10(low)),places=4)
+        self.assertAlmostEqual((float(dot.attrib['cy'])-175)/400,(y_max-60)/(y_max-y_min),places=4)
+        labels=[e.text for e in root.findall('{http://www.w3.org/2000/svg}text')]
+        self.assertIn('Run cost (USD, log scale)', labels)
+        self.assertIn('Youden’s J (recall − false-positive rate)', labels)
+
+    def test_more_false_positives_reduce_j_at_same_cost(self):
+        current=self.rows();current['typellm']['confusion']={'tp':7,'fn':3,'fp':10,'tn':10}
+        dots=self.dots(self.render(current))
+        self.assertEqual(dots['jev'].attrib['cx'],dots['typellm'].attrib['cx'])
+        self.assertLess(float(dots['jev'].attrib['cy']),float(dots['typellm'].attrib['cy']))
+
+    def test_frontier_excludes_strictly_dominated_models_and_keeps_ties(self):
+        def row(cost,tp,fp):
+            return {'inference_cost_usd':cost,'confusion':{'tp':tp,'fn':10-tp,'fp':fp,'tn':10-fp}}
+        current={'mercury-decide':row(.02,8,1),'jev':row(.05,6,1),
+                 'openai-decisions':row(.09,9,1),'typellm':row(.02,8,1)}
+        self.assertEqual(tradeoff.pareto_models(current),{'mercury-decide','openai-decisions','typellm'})
+        current['codeql']=row(None,10,0)
+        self.assertNotIn('codeql',tradeoff.pareto_models(current))
+
+    def test_unknown_cost_is_not_plotted_as_zero(self):
+        current=self.rows();current['codeql']={'confusion':{'tp':10,'fn':0,'fp':0,'tn':20}}
+        dots=self.dots(self.render(current));self.assertNotIn('codeql',dots)
+        current['jev']['inference_cost_usd']=0
+        with self.assertRaises(ValueError):self.render(current)
+
+    def test_axis_shows_negative_scores_and_always_yes_scores_zero(self):
+        current=self.rows();current['jev']['confusion']={'tp':10,'fn':0,'fp':20,'tn':0}
+        current['typellm']['confusion']={'tp':1,'fn':9,'fp':10,'tn':10}
+        root=self.render(current);dots=self.dots(root)
+        self.assertEqual(float(dots['jev'].attrib['data-j']),0)
+        self.assertLess(float(root.attrib['data-j-min']),0)
 
 
 class AlertCurveTests(unittest.TestCase):

tradeoff.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Rebuild public recall/false-positive-rate charts and benchmark result tables.
+"""Rebuild public Youden’s J versus cost charts and benchmark result tables.
 
 Recorded alert counts are validated against each run's question floors and failure policy.
 Private tables use only aggregates in results-current.jsonl.
@@ -11,7 +11,7 @@ import argparse
 from collections import defaultdict
 import html
 import json
-from math import isfinite
+from math import ceil, floor, isfinite, log10
 from pathlib import Path
 from tempfile import TemporaryDirectory
 from urllib.request import urlopen
@@ -26,7 +26,7 @@ DATASET_URLS = {
 NAMES = {
     "jev": ("Jev", "https://typesafe.ai/blog/introducing-system-one-models-and-jev"),
     "typellm": ("TypeLLM", "https://typellm.ai/docs"),
-    "liquid-d1": ("Liquid d1", "https://docs.liquid.ai/lfm/models/decision-models"),
+    "liquid-d1": ("Liquid d1", "https://openrouter.ai/liquid/d1"),
     "diffusiongemma": ("DiffusionGemma Jev", "https://github.com/razorback16/openjev"),
     "span01": ("Span-01", "https://www.respan.ai/docs/documentation/span-01/concept"),
     "clm": ("CLM-v0.1-8B", "https://huggingface.co/Contrastive-LM/CLM-v0.1-8B"),
@@ -193,7 +193,7 @@ def results_table(rows: list[dict]) -> str:
     def mark(value: str, winner: bool) -> str:
         return f"**{value}**" if winner else value
 
-    table = ["| Model | Recall | False-positive rate | Recall − false-positive rate | Failed requests | Inference cost |",
+    table = ["| Model | Recall | False-positive rate | Youden’s J | Failed requests | Inference cost |",
              "| --- | ---: | ---: | ---: | ---: | ---: |"]
     for row in ordered:
         recall, false_positive_rate, j = metrics[row["model_id"]]
@@ -213,51 +213,116 @@ def results_table(rows: list[dict]) -> str:
     return "\n".join(table)
 
 
+def plot_cost(row: dict) -> float | None:
+    cost = row.get('inference_cost_usd')
+    if cost is None:
+        return None
+    if isinstance(cost, bool) or not isinstance(cost, (int, float)) or not isfinite(cost) or cost <= 0:
+        raise ValueError('A cost plot requires a positive recorded inference cost')
+    return cost
+
+
+def pareto_models(current: dict[str, dict]) -> set[str]:
+    values = {model: (plot_cost(row), operating_metrics(row)[2])
+              for model, row in current.items() if plot_cost(row) is not None}
+    return {model for model, (cost, score) in values.items()
+            if not any(other_cost <= cost and other_score >= score and
+                       (other_cost < cost or other_score > score)
+                       for other, (other_cost, other_score) in values.items() if other != model)}
+
+
+def nice_step(value: float) -> float:
+    power = 10 ** floor(log10(max(value, 1e-9)))
+    return next(factor * power for factor in (1, 2, 2.5, 5, 10) if factor * power >= value)
+
+
+def plot_axes(current: dict[str, dict]) -> tuple[float, float, list[float], float, float, float]:
+    costs = [plot_cost(row) for row in current.values() if plot_cost(row) is not None]
+    scores = [operating_metrics(row)[2] * 100 for row in current.values() if plot_cost(row) is not None]
+    if not costs:
+        raise ValueError('No recorded costs to plot')
+    ticks = sorted({factor * 10 ** exponent
+                    for exponent in range(floor(log10(min(costs))) - 1, ceil(log10(max(costs))) + 2)
+                    for factor in (1, 2, 5)})
+    low = max(tick for tick in ticks if tick <= min(costs))
+    high = min(tick for tick in ticks if tick >= max(costs))
+    if low == high:
+        low /= 2
+        high *= 2
+        ticks = sorted(set([*ticks, low, high]))
+    ticks = [tick for tick in ticks if low <= tick <= high]
+    step = nice_step(max(max(scores) - min(0, min(scores)), 1) / 5)
+    bottom = floor(min(0, min(scores)) / step) * step
+    top = max(step, ceil(max(0, max(scores)) * 1.06 / step) * step)
+    return low, high, ticks, bottom, top, step
+
+
 def svg(domain: str, current: dict[str, dict], output: Path) -> None:
-    width, height = 1120, 580
-    left, right, top, bottom = 90, 40, 148, 82
+    width, height = 1280, 690
+    left, right, top, bottom = 100, 50, 175, 115
     plot_w, plot_h = width - left - right, height - top - bottom
-    models = [*ORDER, *(["codeql"] if "codeql" in current else [])]
-    title = {"bug": "Bug detection", "security": "Security detection", "lint": "Perch Lint Bench"}[domain]
-    out = [f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}" role="img" aria-labelledby="title desc">',
-           f'<title id="title">{title}: recall versus false positive rate</title>',
-           '<desc id="desc">Each point is a model at its recorded table cutoff. X is the percentage of clean methods falsely flagged. Y is the percentage of issue-bearing methods found. Higher and farther left is better. CodeQL, where present, uses the subset documented below the table.</desc>',
+    models = [model for model in ORDER if model in current and plot_cost(current[model]) is not None]
+    plotted = {model: current[model] for model in models}
+    frontier = pareto_models(plotted)
+    low, high, cost_ticks, y_min, y_max, y_step = plot_axes(plotted)
+    def x(cost):
+        return left + (log10(cost) - log10(low)) / (log10(high) - log10(low)) * plot_w
+    def y(score):
+        return top + (y_max - score * 100) / (y_max - y_min) * plot_h
+    title = {'bug': 'Bug detection', 'security': 'Security detection', 'lint': 'Perch Lint Bench'}[domain]
+    methods = next(row['benchmark_methods'] for row in current.values() if 'benchmark_methods' in row)
+    out = [f'<svg xmlns="http://www.w3.org/2000/svg" width="{width}" height="{height}" viewBox="0 0 {width} {height}" role="img" aria-labelledby="title desc" data-cost-low="{low}" data-cost-high="{high}" data-j-min="{y_min}" data-j-max="{y_max}">',
+           f'<title id="title">{title}: Youden’s J versus inference cost</title>',
+           '<desc id="desc">Each point uses the table’s recorded alert settings and inference cost. Y is recall minus false-positive rate. X is US dollars on a logarithmic scale. The amber Pareto frontier marks models for which no other recorded run has both equal or lower cost and equal or higher score, with at least one strict improvement. Higher and farther left is better.</desc>',
            f'<rect width="{width}" height="{height}" rx="16" fill="#101318"/>',
-           f'<text x="{left}" y="43" fill="#f2f5fa" font-family="Arial,sans-serif" font-weight="700" font-size="28">{html.escape(title)}</text>',
-           f'<text x="{left}" y="70" fill="#aeb9c7" font-family="Arial,sans-serif" font-size="16">Public methods · one point per model · upper left is better</text>']
+           f'<text x="{left}" y="43" fill="#f2f5fa" font-family="Arial,sans-serif" font-weight="700" font-size="30">{html.escape(title)}</text>',
+           f'<text x="{left}" y="72" fill="#aeb9c7" font-family="Arial,sans-serif" font-size="16">{methods:,} public method evaluations · upper left is better</text>',
+           '<line x1="994" y1="42" x2="1030" y2="42" stroke="#ffb454" stroke-width="2.5" stroke-dasharray="5 4"/>',
+           '<text x="1040" y="47" fill="#ffcc8c" font-family="Arial,sans-serif" font-size="16">Pareto frontier</text>']
     for i, model in enumerate(models):
         col, row = i % 4, i // 4
-        lx, ly = left + col * 245, 91 + row * 22
-        color = COLORS.get(model, "#cfd5df")
-        name = "CodeQL*" if model == "codeql" else NAMES[model][0]
-        out += [f'<circle cx="{lx + 12}" cy="{ly}" r="5" fill="{color}"/>',
-                f'<text x="{lx + 32}" y="{ly + 5}" fill="#e1e7ef" font-family="Arial,sans-serif" font-size="15">{html.escape(name)}</text>']
-    for value in range(0, 101, 20):
-        px = left + value / 100 * plot_w
-        py = top + (1 - value / 100) * plot_h
-        out += [f'<line x1="{px:.1f}" y1="{top}" x2="{px:.1f}" y2="{top + plot_h}" stroke="#29303a"/>',
-                f'<text x="{px:.1f}" y="{top + plot_h + 25}" text-anchor="middle" fill="#aeb9c7" font-family="Arial,sans-serif" font-size="14">{value}%</text>',
-                f'<line x1="{left}" y1="{py:.1f}" x2="{left + plot_w}" y2="{py:.1f}" stroke="#29303a"/>',
-                f'<text x="{left - 12}" y="{py + 5:.1f}" text-anchor="end" fill="#aeb9c7" font-family="Arial,sans-serif" font-size="14">{value}%</text>']
+        lx, ly = left + col * 280, 102 + row * 23
+        out += [f'<circle cx="{lx + 8}" cy="{ly}" r="5" fill="{COLORS[model]}"/>',
+                f'<text x="{lx + 24}" y="{ly + 5}" fill="#e1e7ef" font-family="Arial,sans-serif" font-size="15">{html.escape(NAMES[model][0])}</text>']
+    for value in cost_ticks:
+        px = x(value)
+        text = f'${value:.2f}' if value < 1 else f'${value:g}'
+        out += [f'<line x1="{px:.2f}" y1="{top}" x2="{px:.2f}" y2="{top+plot_h}" stroke="#29303a"/>',
+                f'<text x="{px:.2f}" y="{top+plot_h+27}" text-anchor="middle" fill="#aeb9c7" font-family="Arial,sans-serif" font-size="15">{text}</text>']
+    for index in range(round((y_max-y_min)/y_step)+1):
+        value = y_min + index*y_step
+        py = y(value / 100)
+        label = f'{value:g}%'
+        out += [f'<line x1="{left}" y1="{py:.2f}" x2="{left+plot_w}" y2="{py:.2f}" stroke="#39414f" stroke-width="{1.6 if abs(value)<1e-9 else 1}"/>',
+                f'<text x="{left-13}" y="{py+5:.2f}" text-anchor="end" fill="#aeb9c7" font-family="Arial,sans-serif" font-size="15">{label}</text>']
+    frontier_points = sorted({(plot_cost(current[m]), operating_metrics(current[m])[2]) for m in frontier})
+    if len(frontier_points) > 1:
+        points = ' '.join(f'{x(cost):.2f},{y(score):.2f}' for cost, score in frontier_points)
+        out.append(f'<polyline class="pareto-frontier" points="{points}" fill="none" stroke="#ffb454" stroke-width="2.5" stroke-dasharray="5 4" opacity=".85"/>')
     for model in models:
-        recall, fpr, _ = operating_metrics(current[model])
-        px, py = left + fpr * plot_w, top + (1 - recall) * plot_h
-        color = COLORS.get(model, "#cfd5df")
-        name = "CodeQL*" if model == "codeql" else NAMES[model][0]
-        tooltip = html.escape(f"{name}: recall {recall:.1%}; false positive rate {fpr:.1%}")
-        out.append(f'<circle data-model="{model}" cx="{px:.2f}" cy="{py:.2f}" r="6" fill="{color}" stroke="#fff" stroke-width="1.5"><title>{tooltip}</title></circle>')
-    out += [f'<text x="{left + plot_w / 2:.1f}" y="{height - 23}" text-anchor="middle" fill="#e1e7ef" font-family="Arial,sans-serif" font-size="17">False positive rate</text>',
-            f'<text transform="translate(23 {top + plot_h / 2:.1f}) rotate(-90)" text-anchor="middle" fill="#e1e7ef" font-family="Arial,sans-serif" font-size="17">Recall</text>',
+        row = current[model]; cost = plot_cost(row)
+        recall, fpr, score = operating_metrics(row)
+        px, py = x(cost), y(score)
+        tooltip = html.escape(f'{NAMES[model][0]}: Youden’s J {score:.1%}; recall {recall:.1%}; false-positive rate {fpr:.1%}; run cost ${cost:.4f}' + (' (GPU estimate)' if model == 'clm' else ''))
+        if model in frontier:
+            out.append(f'<circle cx="{px:.2f}" cy="{py:.2f}" r="11" fill="#101318" stroke="#ffb454" stroke-width="2"/>')
+        out.append(f'<circle data-model="{model}" data-cost="{cost}" data-j="{score}" data-frontier="{str(model in frontier).lower()}" cx="{px:.2f}" cy="{py:.2f}" r="6" fill="{COLORS[model]}" stroke="#fff" stroke-width="1.5"><title>{tooltip}</title></circle>')
+    footer = 'CLM cost: GPU runtime estimate'
+    if 'codeql' in current:
+        footer += ' · CodeQL cost: unrecorded'
+    out += [f'<text x="{left+plot_w/2}" y="{height-59}" text-anchor="middle" fill="#e1e7ef" font-family="Arial,sans-serif" font-size="18">Run cost (USD, log scale)</text>',
+            f'<text transform="translate(27 {top+plot_h/2}) rotate(-90)" text-anchor="middle" fill="#e1e7ef" font-family="Arial,sans-serif" font-size="17">Youden’s J (recall − false-positive rate)</text>',
+            f'<text x="{left}" y="{height-23}" fill="#8d9aaf" font-family="Arial,sans-serif" font-size="14">{footer}</text>',
             '</svg>']
     output.parent.mkdir(parents=True, exist_ok=True)
-    output.write_text("\n".join(out) + "\n")
+    output.write_text('\n'.join(out)+'\n')
 
 
 def render_sections(results: list[dict], output_dir: Path,
                     codeql_results: dict) -> str:
     sections = ["## Reading the results", "",
-                "Each point shows a model at the same cutoff used in its table. X is the false-positive rate: the percentage of clean methods wrongly flagged. Y is recall: the percentage of labelled issues found. Higher and farther left is better.", "",
-                "The tables use each run's recorded alert settings. A value of 20% under ‘False-positive rate’ means one in five clean methods received an alert. Models are ranked by recall minus that percentage. Bold marks the best full-set value in each column; cost compares priced runs captured here.", ""]
+                "Each point shows a model’s score and cost for the same benchmark. Y is Youden’s J: recall minus false-positive rate. X is the run’s inference cost on a logarithmic scale. Higher and farther left is better. The amber Pareto frontier highlights models offering the best score at their price.", "",
+                "The tables use each run's recorded alert settings. A value of 20% under ‘False-positive rate’ means one in five clean methods received an alert. Models are ranked by Youden’s J. A model that flags every method scores zero. Bold marks the best full-set value in each column; cost compares priced runs captured here.", ""]
     if any(r.get("failed_methods") for r in results):
         sections += ["Failed requests count as incorrect decisions: a failed issue is a miss; a failed clean method is a false positive. Failed cases receive no model score and are included in both the table counts and plotted points.", ""]
     for domain in ("bug", "security", "lint"):
@@ -279,7 +344,7 @@ def render_sections(results: list[dict], output_dir: Path,
         image = output_dir / f"{domain}-tradeoff.svg"
         svg(domain, {r["model_id"]: r for r in public}, image)
         sections += [f"## {title}", "", f"[Public pairs](https://huggingface.co/datasets/perchscan/{link})", "",
-                     f"![{title}: recall versus false-positive rate at table cutoffs](assets/{domain}-tradeoff.svg)", "",
+                     f"![{title}: Youden’s J versus inference cost](assets/{domain}-tradeoff.svg)", "",
                      f"### Public — {public[0]['benchmark_pairs']:,} pairs", "",
                      results_table(public), ""]
         if domain in ("bug", "security"):