PXR activation: released-data model benchmark

Checks run in this order. Open a step to inspect its rules, checker or review rubric.

1. Prediction and package checks; retrospective scoreAutomated checkChecker configured

   Mean absolute error across all 513 released pEC50 estimates, in -log10(molar EC50). Confidence intervals and Emax are not part of this metric; uncertain estimates are not treated as hard censoring bounds. This combined evaluation differs from the historical phase-2-only final leaderboard. This stage never executes submitted code.

   A conforming checker is configured. Execution capacity and availability are checked when the Submission runs.

   Script checker

   One python file, checker.py, run by TfP's own runtime tfp-runtime/python:1 (Python 3.12 (the standard library, no third-party packages); the script reads /input/manifest.json and writes /output/result.json). The author built no image; the program's exact bytes are recorded by hash.

   Read checker code

   Program hash

   sha256:bdad5be79ffb559c0efd6a1a3e8722d587c72bccebc3f206b92954fcdcb7ceb7 · 12015 bytes

   Source provenance and runtime metadata

   The program is content-addressed: these bytes are what every run of this implementation executed. The recorded digest pins the actual runtime image; the catalogue name identifies those same bytes on every deployment. Each run also records the resolved platform image digest. Passing this checker does not prove that its criteria establish scientific correctness.

   Recorded image / builtin reference

   tfp-runtime/python:1

   Recorded runtime digest

   sha256:a8752ba652a6650bc9f6b96063a5f4428858ea0e145e65b2d0a9c58ccb182cdd

   Declared entrypoint (argument array)

   ["/tfp/harness","python","/input/checker/checker.py"]

   Declared source

   No usable public HTTPS source repository link is recorded.

   No exact source commit is recorded.

   Declared license: not recorded. TfP has not verified a reproducible build connecting this declared source to the runtime.

   Checker identity and execution limits

   Implementations can be replaced while preserving this Contract’s acceptance rules.

   {
     "verifier_implementation_id": "vfi_qfbxsyvag1",
     "version": 1,
     "content_hash": "sha256:282c80629002329ceddd781257a45f256089ac1855b6a01ba70fa2e1d21f4db7",
     "conformance": {
       "attempts": 1,
       "checked_at": "2026-10-14T17:10:00Z",
       "differential_against_implementation_id": null,
       "failures": [],
       "fixtures_passed": 6,
       "fixtures_total": 6,
       "last_attempted_at": "2026-10-14T17:10:00Z",
       "last_execution_error": null,
       "status": "passed"
     },
     "runtime_image_digest": "sha256:a8752ba652a6650bc9f6b96063a5f4428858ea0e145e65b2d0a9c58ccb182cdd",
     "limits": {
       "disk_mb": 2048,
       "memory_mb": 1024,
       "network": "none",
       "timeout_seconds": 120,
       "vcpu": 0.25
     }
   }

2. Independent reproduction and scientific reviewReviewReview required

   An accountable independent community reviewer reproduces predictions and checks permitted training, reporting and limitations.

   Eligible reviewers must complete the pinned rubric. Seat availability and completion are not guaranteed.

   Read the pinned rubric and reviewer rules

   Independent public-data benchmark reproduction · pinned protocol

   Independent community reviewer · 1 reviewer

   Apply the bound benchmark-method Skill. Independently rerun the submitted source with pinned train/test inputs, compare predictions within max(1e-6, 1e-6*abs(submitted value)) and independently confirm the score. Inspect code/provenance for evaluation-label fitting, external labels/pretrained weights, train-only validation and truthful compute/agent/limitations reporting. Read the source before running it in an isolated environment without credentials. Passing format checks or a low score alone is insufficient. Pass only when reproducibility and all scientific-policy claims are substantiated; fail for a material violation; uncertain when the evidence or available resources do not permit deciding. Cite the inspected source and rerun evidence, and declare any agent assistance. Better-than-baseline performance is not required. No OpenADMET affiliation or endorsement is presumed.

   Rationale requiredAt least 1 citationConfidence required

   Response fields and full schema

   • verdict · required "pass" / "fail" / "uncertain"

   {
     "additionalProperties": false,
     "properties": {
       "verdict": {
         "enum": [
           "pass",
           "fail",
           "uncertain"
         ]
       }
     },
     "required": [
       "verdict"
     ],
     "type": "object"
   }

   Any eligible participant may claim a seat.The submitter is excluded.Responses on this node are withheld until it completes.Declare any agent assistance used; a declaration is optional under this protocol.

   Exact reviewer eligibility, visibility and Skill pins

   {
     "reviewers": {
       "allowed": null,
       "assistance": {
         "allowed": [],
         "required": false
       },
       "count": 1,
       "eligibility": {
         "organization_id": null,
         "rule": "any_participant",
         "user_ids": null
       },
       "independent": true,
       "kind": "human"
     },
     "conflicts": {
       "exclude_submitter": true,
       "exclude_submitter_organization": false
     },
     "visibility": {
       "hidden": [
         "submitter_identity",
         "submitter_organization"
       ],
       "prior_nodes": [],
       "stage_evidence": true
     },
     "skills": [
       {
         "requirement": "required",
         "skill_version_id": "skv_tos7mh7z92"
       }
     ]
   }

   How judgments become the outcome

   Decide

   Accept when at least 1 responses have verdict equals "pass". Reject when at least 1 have verdict equals "fail". Otherwise unresolved.

   Inputs, in order: reviewer · decision

   Final outcome comes from decision. Accept passes this stage; reject refutes it; unresolved is inconclusive. All verification stages must pass.

   Node deadline: 168 hoursClaim hold: 1440 minutes, bounded by the node deadline

   Protocol identity and full rules

   rpt_in8g52sm2l
sha256:3098073acddec678e60fa1de88bc354453f5b09a30a259d596f2650d5bc85f4e

   Read the complete pinned protocol →

   {
     "output": "decision",
     "policy": {
       "assignment_timeout_hours": 168,
       "claim_hold_minutes": 1440
     }
   }