Commit the first real benchmark result + fix a self-invalidating test

Pulled the live result from /api/admin/run-benchmarks (run by the founder
against the bundled BLAT_ECOLX fixture on the actual Space, real ESM-2
small): median Spearman rho 0.5165, top-decile precision 47.8%, across all
4,996 real mutations. This is now the permanent, committed state, so it
survives future redeploys instead of resetting to the empty placeholder.

Also fixes test_benchmarks_route_public_and_honest_empty, which asserted the
bundled file is ALWAYS empty — true when it was written, but wrong the moment
a real run legitimately populates it (which just happened). Replaced with a
shape-only assertion (the route never fabricates, whatever the current file
state is) plus a separate test that the honest-empty behavior still holds
when the file is genuinely absent.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

dee/data/benchmarks.json

@@ -1,12 +1,21 @@
 {
-  "generated_at": null,
-  "model": null,
-  "note": "Validation results are produced by scripts/run_benchmarks.py against published DMS data with the live ESM-2 model (where the weights exist), never hand-entered. Until a real run populates this file, the app shows an honest 'pending' state — no fabricated numbers.",
+  "datasets": [
+    {
+      "n": 4996,
+      "name": "BLAT_ECOLX_Stiffler2015",
+      "protein": "TEM-1 beta-lactamase (BLAT_ECOLX, UniProt P62593, mature residues 24-286)",
+      "source": "Stiffler, Hekstra & Ranganathan 2015, Cell 160:882-892, \"Evolvability as a Function of Purifying Selection in TEM-1 beta-Lactamase\" (DOI 10.1016/j.cell.2015.01.035). Fitness score = the 2500 ug/mL ampicillin selection condition (the most stringent tested), the complete single-mutant scan (all 263 positions x 19 substitutions). Data file redistributed by facebookresearch/esm (examples/variant-prediction/data/BLAT_ECOLX_Ranganathan2015.csv) as the reference DMS example for Meier et al. 2021 'Language models enable zero-shot prediction of the effects of mutations on protein function' -- the exact scoring scheme dee/models/scorer.py implements. Reference sequence reconstructed from the accompanying BLAT_ECOLX_1_b0.5.a3m alignment (header BLAT_ECOLX/24-286); mutation labels renumbered from the paper's absolute residue numbering (offset +23) to 1-indexed local positions matching this sequence. Every one of the 4,996 mutations was cross-validated against this sequence (100% wild-type-residue match) before being committed.",
+      "spearman": 0.5165,
+      "top_decile_precision": 0.478
+    }
+  ],
+  "generated_at": "2026-07-15T08:35:43.806166+00:00",
+  "model": "small",
+  "note": "Computed live on this Space against dee/data/dms_fixtures/ (see manifest.json for provenance/citations). Never hand-entered.",
   "summary": {
-    "n_datasets": 0,
-    "n_variants": 0,
-    "median_spearman": null,
-    "median_top_decile_precision": null
-  },
-  "datasets": []
+    "median_spearman": 0.5165,
+    "median_top_decile_precision": 0.478,
+    "n_datasets": 1,
+    "n_variants": 4996
+  }
 }

tests/test_benchmark.py

@@ -107,13 +107,28 @@ def test_summarize_empty():
                  "median_spearman": None, "median_top_decile_precision": None}
 
 
-def test_benchmarks_route_public_and_honest_empty():
+def test_benchmarks_route_public_and_shaped():
     from dee import server
     app = server.create_app()
     app.config.update(TESTING=True)
     body = app.test_client().get("/api/benchmarks").get_json()   # no auth — public
     assert body["ok"] is True
     assert "summary" in body and "datasets" in body
-    # Ships empty (no fabricated numbers) until a real run populates it.
+    assert set(body["summary"]) == {"n_datasets", "n_variants",
+                                    "median_spearman", "median_top_decile_precision"}
+    # n_datasets/datasets length must agree, whatever the current bundled state is
+    # (empty before the first real run, or real numbers once /api/admin/run-
+    # benchmarks has populated it — this route never fabricates either way).
+    assert body["summary"]["n_datasets"] == len(body["datasets"])
+
+
+def test_benchmarks_route_honest_empty_when_file_absent(monkeypatch, tmp_path):
+    from dee import server
+    monkeypatch.setattr(server, "__file__", str(tmp_path / "server.py"))
+    app = server.create_app()
+    app.config.update(TESTING=True)
+    body = app.test_client().get("/api/benchmarks").get_json()
+    assert body["ok"] is True
     assert body["summary"]["n_datasets"] == 0
     assert body["datasets"] == []
+    assert body["generated_at"] is None