Skip to content

Commit fd218fe

Browse files
AbdelStarkAbdelStark
committed
Document GenoLeWM-FX kill gate
1 parent e9ce347 commit fd218fe

12 files changed

Lines changed: 1743 additions & 0 deletions

‎README.md‎

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -90,6 +90,10 @@ findings. The strongest current claim is that the repository can train,
9090
package, evaluate, benchmark, and replay a genomic-edit world-model
9191
pipeline with content-addressed evidence.
9292

93+
The GenoLeWM-FX pivot is documented as a killed feasibility experiment:
94+
public TraitGym labels are available, but no reproducible 10k-50k
95+
teacher-delta cache is locked under the experiment contract.
96+
9397
## Quickstart
9498

9599
```python
@@ -316,6 +320,8 @@ jobs. Fixture outputs are test evidence, not model results.
316320
- The K=20 AR rollout speed target remains open.
317321
- The released planning demo exercises the model path but does not show
318322
useful planning behavior.
323+
- No GenoLeWM-FX model or demo ships; the FX pivot is stopped at the
324+
feasibility gate.
319325
- No clinical utility claim; the public model evidence is not clinical
320326
evidence.
321327
- Personal-genome workflows are local-first, but local execution is not

‎configs/fx/feasibility_sources.json‎

Lines changed: 204 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,204 @@
1+
{
2+
"schema_version": "1.0.0",
3+
"epic_issue": 257,
4+
"contract_issue": 258,
5+
"probe_issue": 259,
6+
"sample_window": {
7+
"min_examples": 10000,
8+
"max_examples": 50000
9+
},
10+
"source_probe": {
11+
"dataset": "songlab/TraitGym",
12+
"config": "complex_traits",
13+
"split": "test",
14+
"holdout_chromosomes": [
15+
"3",
16+
"11",
17+
"14",
18+
"16",
19+
"17",
20+
"20"
21+
],
22+
"bootstrap_samples": 200,
23+
"bootstrap_seed": 257,
24+
"score_columns": [
25+
{
26+
"id": "maf_source_only",
27+
"column": "maf",
28+
"direction": "positive",
29+
"description": "Minor-allele-frequency metadata score."
30+
},
31+
{
32+
"id": "ld_score_source_only",
33+
"column": "ld_score",
34+
"direction": "positive",
35+
"description": "LD-score metadata score."
36+
},
37+
{
38+
"id": "tss_distance_source_only",
39+
"column": "tss_dist",
40+
"direction": "negative_abs",
41+
"description": "Distance-to-TSS metadata score, closer variants ranked higher."
42+
}
43+
]
44+
},
45+
"candidate_sources": [
46+
{
47+
"id": "traitgym_complex_traits",
48+
"name": "TraitGym complex_traits",
49+
"url": "https://github.com/songlab-cal/TraitGym",
50+
"dataset_ref": "songlab/TraitGym:complex_traits:test",
51+
"public_labels": true,
52+
"observed_examples": 11400,
53+
"within_probe_scale": true,
54+
"teacher_delta_cache": "missing",
55+
"admissible_for_fx_probe": false,
56+
"blocking_reasons": [
57+
"Variant labels are public and load through Hugging Face datasets, but the source does not provide a GenoLeWM-FX ref/alt teacher-delta cache.",
58+
"The loaded columns expose variant metadata and binary labels, not signed teacher residuals or teacher feature tensors."
59+
]
60+
},
61+
{
62+
"id": "traitgym_mendelian_traits",
63+
"name": "TraitGym mendelian_traits",
64+
"url": "https://github.com/songlab-cal/TraitGym",
65+
"dataset_ref": "songlab/TraitGym:mendelian_traits:test",
66+
"public_labels": true,
67+
"observed_examples": 3380,
68+
"within_probe_scale": false,
69+
"teacher_delta_cache": "missing",
70+
"admissible_for_fx_probe": false,
71+
"blocking_reasons": [
72+
"The public split is below the 10k cheap-probe floor.",
73+
"The source does not provide ref/alt teacher deltas or teacher feature tensors."
74+
]
75+
},
76+
{
77+
"id": "mave_mpra_qtl_candidate_slices",
78+
"name": "MAVE, MPRA, and QTL candidate slices",
79+
"url": "https://www.mavedb.org/",
80+
"dataset_ref": "not_locked",
81+
"public_labels": "mixed",
82+
"observed_examples": null,
83+
"within_probe_scale": false,
84+
"teacher_delta_cache": "missing",
85+
"admissible_for_fx_probe": false,
86+
"blocking_reasons": [
87+
"No exact assay, genome build, windowing rule, split manifest, and teacher-cache manifest is locked in this repo.",
88+
"Using these sources now would require new curation before the probe can be reproduced."
89+
]
90+
}
91+
],
92+
"candidate_teachers": [
93+
{
94+
"id": "alphagenome_api",
95+
"name": "AlphaGenome API",
96+
"url": "https://github.com/google-deepmind/alphagenome",
97+
"role_allowed_by_contract": "sampled calibration/API baseline",
98+
"bulk_training_admissible": false,
99+
"local_runtime_available": false,
100+
"blocking_reasons": [
101+
"The contract permits AlphaGenome only as a bounded sampled calibration/API baseline unless terms and rates support larger use.",
102+
"The local run has no AlphaGenome API key or package, and the public API is not a checksum-addressed 10k-50k teacher cache."
103+
]
104+
},
105+
{
106+
"id": "alphagenome_weights",
107+
"name": "AlphaGenome released weights",
108+
"url": "https://github.com/google-deepmind/alphagenome_research",
109+
"role_allowed_by_contract": "future heavyweight local teacher if terms, hardware, and artifacts are locked",
110+
"bulk_training_admissible": false,
111+
"local_runtime_available": false,
112+
"blocking_reasons": [
113+
"Weights require accepting model terms and the upstream docs recommend H100-class hardware for model inference.",
114+
"No accepted-terms receipt, model revision, hardware receipt, or ref/alt teacher-cache artifact is available in this repo."
115+
]
116+
},
117+
{
118+
"id": "borzoi",
119+
"name": "Borzoi",
120+
"url": "https://github.com/calico/borzoi",
121+
"role_allowed_by_contract": "candidate local teacher",
122+
"bulk_training_admissible": false,
123+
"local_runtime_available": false,
124+
"blocking_reasons": [
125+
"The model is public, but the local run does not have the TensorFlow teacher stack or downloaded model weights.",
126+
"The upstream large training data is requester-pays; no compact public ref/alt cache is locked for GenoLeWM-FX."
127+
]
128+
},
129+
{
130+
"id": "enformer",
131+
"name": "Enformer",
132+
"url": "https://www.nature.com/articles/s41592-021-01252-x",
133+
"role_allowed_by_contract": "candidate local teacher",
134+
"bulk_training_admissible": false,
135+
"local_runtime_available": false,
136+
"blocking_reasons": [
137+
"No TensorFlow teacher stack, model revision, input window manifest, or ref/alt output cache is present in this repo.",
138+
"Using Enformer now would make the probe depend on ad hoc setup rather than a public GenoLeWM-FX cache manifest."
139+
]
140+
},
141+
{
142+
"id": "chrombpnet",
143+
"name": "ChromBPNet",
144+
"url": "https://github.com/kundajelab/chrombpnet",
145+
"role_allowed_by_contract": "candidate task-specific teacher",
146+
"bulk_training_admissible": false,
147+
"local_runtime_available": false,
148+
"blocking_reasons": [
149+
"The repository provides model/training tooling, but no selected cell-type model, peaks, genome, or ref/alt cache is locked here.",
150+
"No GPU/Docker teacher runtime receipt or checksum-addressed GenoLeWM-FX cache is present."
151+
]
152+
}
153+
],
154+
"required_baselines": [
155+
"zero-delta/no-edit",
156+
"source-only/label-prior",
157+
"Carbon likelihood/log-odds where applicable",
158+
"direct teacher ref-alt score",
159+
"linear/logistic probe on teacher features",
160+
"available public model predictions when feasible"
161+
],
162+
"recommended_issue_actions": [
163+
{
164+
"issue": 258,
165+
"action": "close-completed",
166+
"reason": "The source-controlled contract is locked in docs/research/fx-experiment-contract.md."
167+
},
168+
{
169+
"issue": 259,
170+
"action": "close-kill",
171+
"reason": "The feasibility gate found no reproducible 10k-50k public teacher-delta slice under the contract."
172+
},
173+
{
174+
"issue": 260,
175+
"action": "close-not-planned",
176+
"reason": "Teacher adapters and caches should not be built until a source/teacher pair passes the feasibility gate."
177+
},
178+
{
179+
"issue": 261,
180+
"action": "close-not-planned",
181+
"reason": "The residual model path should not be implemented without a selected target and teacher cache."
182+
},
183+
{
184+
"issue": 262,
185+
"action": "close-not-planned",
186+
"reason": "No staged Hugging Face jobs should launch without the #259 go decision."
187+
},
188+
{
189+
"issue": 263,
190+
"action": "close-not-planned",
191+
"reason": "The locked benchmark gate is unreachable because the experiment stops at feasibility."
192+
},
193+
{
194+
"issue": 264,
195+
"action": "close-completed",
196+
"reason": "The public output is the kill report, not a model artifact release."
197+
},
198+
{
199+
"issue": 265,
200+
"action": "close-completed",
201+
"reason": "The final paper/demo decision is no FX paper update and no demo because evidence is insufficient."
202+
}
203+
]
204+
}

‎docs/index.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -16,6 +16,8 @@ model-quality evidence is mixed or negative versus Carbon.
1616
- [Architecture](architecture.md)
1717
- [Public API contract](api/public-surface.md)
1818
- [API reference](reference/index.md)
19+
- [GenoLeWM-FX contract](research/fx-experiment-contract.md)
20+
- [GenoLeWM-FX feasibility report](research/fx-feasibility-report.md)
1921

2022
## Public Artifacts
2123

@@ -46,5 +48,7 @@ for Carbon-backed training paths.
4648
- No broad model-quality claim.
4749
- No K=20 rollout-speed closure.
4850
- No useful-planning claim from the current planning demo.
51+
- No GenoLeWM-FX model or demo ships; the FX pivot is stopped at the
52+
feasibility gate.
4953
- No runtime or privacy assurance beyond local execution contracts and
5054
checksum provenance.

‎docs/research/fx-decision-package.md‎

Lines changed: 54 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,54 @@
1+
# GenoLeWM-FX decision package
2+
3+
Status: final decision package for issues #257 and #265.
4+
5+
The GenoLeWM-FX pivot is stopped at the feasibility gate. No FX teacher
6+
cache, residual JEPA model, Hugging Face sweep, locked benchmark, model
7+
artifact release, paper update, or public demo ships from this
8+
experiment.
9+
10+
Authoritative artifacts:
11+
12+
- [FX experiment contract](fx-experiment-contract.md)
13+
- [FX feasibility and kill report](fx-feasibility-report.md)
14+
- [Machine-readable feasibility report](fx-feasibility-report.json)
15+
16+
## Decision
17+
18+
The parent epic takes the kill path. The checked-in feasibility report
19+
found a public 11,400-row TraitGym source slice suitable for a cheap
20+
source-only probe, but no reproducible 10k-50k ref/alt teacher-delta
21+
cache from AlphaGenome, Borzoi, Enformer, ChromBPNet, or another
22+
functional teacher.
23+
24+
Continuing would require private credentials, heavyweight teacher setup,
25+
new assay curation, or newly published teacher-cache artifacts before
26+
the actual GenoLeWM-FX hypothesis could be tested. That does not justify
27+
medium or expensive training jobs.
28+
29+
## Demo Decision
30+
31+
No FX demo ships. A demo would imply an available FX model or teacher
32+
residual score that this experiment did not produce. Existing GenoLeWM
33+
demos remain manifest-backed execution examples only and should not be
34+
described as evidence of useful planning behavior.
35+
36+
## Paper Decision
37+
38+
No FX positive-result paper section is added. If referenced in future
39+
paper-facing text, this path should be described as a killed feasibility
40+
probe: public labels were available, but the required teacher-delta
41+
target was not reproducible under the contract.
42+
43+
## Child Issue Resolution
44+
45+
| Issue | Resolution |
46+
| ---: | --- |
47+
| #258 | Completed by the source-controlled contract. |
48+
| #259 | Completed as a kill decision at feasibility. |
49+
| #260 | Not planned after the kill gate; no teacher cache should be built now. |
50+
| #261 | Not planned after the kill gate; no residual model target is selected. |
51+
| #262 | Not planned after the kill gate; no Stage B or Stage C job should launch. |
52+
| #263 | Not planned because the locked benchmark depends on the killed path. |
53+
| #264 | Completed by publishing the kill report instead of model artifacts. |
54+
| #265 | Completed by this no-demo/no-paper-update decision package. |

0 commit comments

Comments
 (0)