GenoLeWM-FX feasibility and kill report¶
Generated by tools.research.fx_feasibility at 2026-06-11T10:00:00Z.
Parent epic: #257. Contract: #258. Probe: #259.
Decision: kill.
The FX pivot has no model-quality claim, no clinical utility claim, no broad variant-effect prediction claim, and no useful-planning claim. The public output is a kill report unless a future source/teacher cache passes this contract.
Reproduce¶
uv run python -m tools.research.fx_feasibility \
--manifest configs/fx/feasibility_sources.json \
--run-public-source-probe \
--output-json docs/research/fx-feasibility-report.json \
--output-md docs/research/fx-feasibility-report.md
Machine-readable report: fx-feasibility-report.json.
Blockers¶
no_public_teacher_delta_cache: No candidate source provides a public, checksum-addressed ref/alt teacher-delta cache for the required 10k-50k probe.no_bulk_admissible_teacher: No teacher is currently admissible for bulk training-cache use under the contract and local artifact state.local_teacher_runtime_missing: The local runtime has no AlphaGenome credential/package and no TensorFlow or AlphaGenome Research teacher stack.
Public Source Probe¶
| Baseline | Rows | AUROC | AUPRC | Balanced accuracy |
|---|---|---|---|---|
label_prior_no_teacher |
3390 | 0.5000 [0.5000, 0.5000] | 0.1000 [0.1000, 0.1000] | 0.5000 [0.5000, 0.5000] |
maf_source_only |
3390 | 0.5057 [0.4785, 0.5324] | 0.1024 [0.0892, 0.1184] | 0.4987 [0.4853, 0.5153] |
ld_score_source_only |
3390 | 0.4923 [0.4646, 0.5249] | 0.0997 [0.0877, 0.1122] | 0.5018 [0.4855, 0.5168] |
tss_distance_source_only |
3390 | 0.5089 [0.4780, 0.5392] | 0.1091 [0.0911, 0.1249] | 0.5188 [0.4970, 0.5440] |
These rows test whether the public TraitGym label slice is trivially saturated by simple metadata. They do not provide the required teacher residual target.
Candidate Teacher Status¶
| Teacher | Allowed role | Bulk training admissible | Local runtime | Blocking reason |
|---|---|---|---|---|
alphagenome_api |
sampled calibration/API baseline | False | False | The contract permits AlphaGenome only as a bounded sampled calibration/API baseline unless terms and rates support larger use.; The local run has no AlphaGenome API key or package, and the public API is not a checksum-addressed 10k-50k teacher cache. |
alphagenome_weights |
future heavyweight local teacher if terms, hardware, and artifacts are locked | False | False | Weights require accepting model terms and the upstream docs recommend H100-class hardware for model inference.; No accepted-terms receipt, model revision, hardware receipt, or ref/alt teacher-cache artifact is available in this repo. |
borzoi |
candidate local teacher | False | False | The model is public, but the local run does not have the TensorFlow teacher stack or downloaded model weights.; The upstream large training data is requester-pays; no compact public ref/alt cache is locked for GenoLeWM-FX. |
enformer |
candidate local teacher | False | False | No TensorFlow teacher stack, model revision, input window manifest, or ref/alt output cache is present in this repo.; Using Enformer now would make the probe depend on ad hoc setup rather than a public GenoLeWM-FX cache manifest. |
chrombpnet |
candidate task-specific teacher | False | False | The repository provides model/training tooling, but no selected cell-type model, peaks, genome, or ref/alt cache is locked here.; No GPU/Docker teacher runtime receipt or checksum-addressed GenoLeWM-FX cache is present. |
Recommended Issue Resolution¶
| Issue | Action | Reason |
|---|---|---|
| #258 | close-completed |
The source-controlled contract is locked in docs/research/fx-experiment-contract.md. |
| #259 | close-kill |
The feasibility gate found no reproducible 10k-50k public teacher-delta slice under the contract. |
| #260 | close-not-planned |
Teacher adapters and caches should not be built until a source/teacher pair passes the feasibility gate. |
| #261 | close-not-planned |
The residual model path should not be implemented without a selected target and teacher cache. |
| #262 | close-not-planned |
No staged Hugging Face jobs should launch without the #259 go decision. |
| #263 | close-not-planned |
The locked benchmark gate is unreachable because the experiment stops at feasibility. |
| #264 | close-completed |
The public output is the kill report, not a model artifact release. |
| #265 | close-completed |
The final paper/demo decision is no FX paper update and no demo because evidence is insufficient. |
Final Interpretation¶
The FX pivot is stopped before teacher-cache implementation, residual model work, Hugging Face sweeps, or a public demo. The public TraitGym complex-trait split is large enough for a cheap source-only probe, but this repo does not have a locked, checksum-addressed 10k-50k ref/alt teacher-delta cache from AlphaGenome, Borzoi, Enformer, ChromBPNet, or another functional teacher. Continuing would require private credentials, heavyweight teacher setup, or newly curated artifacts before the actual GenoLeWM-FX hypothesis could be tested.