# Goals 1--5: finalized CPU and corpus results The values below are transcribed from the finalized manifests, summaries, plot data, and reports in `GEML_artifacts`. They are not estimates and do not include raw shards. Each number is mapped to its source in `PROVENANCE.md`. ## Goal 1 -- source-expression corpus The final corpus is `geml-goal1-final`, schema `geml-corpus-v1`, generated with seed 20260721. It contains 250,000 accepted rows split 175,000/25,000/25,000/25,000 (train/validation/test-IID/test-OOD), with zero storage, AST, parse, LaTeX, or round-trip failures. The generator attempted 286,413 rows: 35,768 were duplicate attempts and 645 were rejected by the explicit triviality policy; 250,000 were accepted, giving an acceptance rate of 0.8728654076456027. Unsupported rows and identity conflicts were both zero. Generation took 620.2928234 s. Accepted-row throughput was 403.0354544966034 rows/s, attempted-row throughput was 461.7383745149426 rows/s, and peak RSS was 2,193,186,816 bytes. Domain counts were 68,345 nonzero-real, 84,843 positive-real, and 96,812 safe-real. Family counts were 70,000 algebraic-core, 40,000 exp/log, 35,000 mixed-elementary, 25,000 OOD-stress, 40,000 powers/division/rationals, and 40,000 trig/hyperbolic. Variable counts for 1 through 6 variables were 68,934, 61,244, 46,743, 34,511, 24,095, and 14,473. The policy audit reports all trig operators covered, approved-domain checks true, blanket `log(exp(...))` wrapping false, 196,656 certified log arguments, 28,648 certified tan arguments, 257,698 lowered reciprocal candidates, and 292,783 negative power arguments. A LaTeX parser was unavailable for 64 checks, but no round-trip failure was recorded; this is a tooling limitation, not a silent success. QA found 250,000 unique authoritative s-expressions, expression IDs, and structural identities; cross-split ID collisions and duplicate authoritative occurrences were zero. Triviality feature counts were constant-only 159,469, exp/log 34,685, log/exp 42,093, log(1) 6,400, and multiplication-by-one 50,000, subject to the configured caps. ## Goal 2 -- pure EML The official-v4 compiler processed all 250,000 rows with zero conversion failures. Elapsed time was 2,557.261734 s, processing throughput 99.16671338859994 rows/s, aggregation throughput 97.76081840827327 rows/s, and peak RSS 1,617,088,512 bytes. The semantic audit selected 280 rows: 273 were materialized and audited (203 passed, 3 mismatch, 45 nonfinite, and 22 overflow), while 7 exceeded the node limit before materialization. A further 249,720 rows were not selected for that audit. These audit statuses are not corpus conversion failures. Raw pure-EML expansion had median alpha 40.6602, mean 952.1371, and approximate p99 10,448.6. Zero of 250,000 rows fell below the preregistered 1.29--1.50 thresholds. ## Goal 3 -- AST/EML DAGs and costs Direct hash-consing processed 250,000 rows with zero failures, audit gate ready, in 1,568.0315 s (159.4356 rows/s; cumulative 162.7917 rows/s) and peak RSS 1,444,044,800 bytes. AST-DAG maximum indegree was 63; EML-DAG maximum indegree was 1,939. Mean reused nodes were 3.430812 (AST) and 18.130736 (EML); mean reused references were 13.759432 and 276.193548; mean excess references were 10.32862 and 258.062812. Mean reuse depth was 4.6421511875 and 27.7799859862, with sharing concentration 0.5698556037 and 0.7978003017. The mean raw-tree alpha was 952.1371252900141. Reported means were 8.334401271758448 for EML-DAG versus AST-tree, 10.474953890182578 for EML-DAG versus AST-DAG, 1.361707663394114 AST compression, and 39.37500771693145 EML compression. No row was structurally competitive under the published criterion; the best remaining ratio was 8/7. ## Goal 4 -- verifier-gated e-graph study The study selected 30,000 expressions and evaluated two 30,000-row modes. Each mode costed 18,210 rows, had 11,790 failures, and zero timeouts. For `safe_real`, 4,349 costed rows improved (23.88248215% of costed; 14.49666667% of processed), with mean signed improvement 5.2496430533 nodes and mean relative improvement 0.0236117313. For `positive_real_formal`, 5,026 improved (27.60021966% of costed; 16.75333333% of processed), with mean signed improvement 5.7658429434 nodes and relative improvement 0.02848704193. Failures were principally unsupported trig/hyperbolic operators or independent-validation rejection; no timeout occurred. ## Goal 5 -- motifs, ranker, and export The selected frequent vocabulary has 1,024 motifs of sizes 2--4 with minimum support 32. On test-IID it used 317,678,264 bits versus a 572,524,716-bit baseline, saving 254,846,452 bits (44.512742398%); on test-OOD it used 818,099,441 versus 1,453,062,331 bits, saving 634,962,890 bits (43.69825550%). Reconstruction and source failures were zero. The learned vocabulary also has size 1,024. Its test-IID total was 324,485,346 bits, versus 317,678,264 for equal-budget frequent, and random median 368,943,015 bits. This is a null result against equal-budget frequent, positive against random and uncompressed macro baselines. Test-OOD totals were 821,843,999 learned and 818,099,441 equal-budget frequent bits. The neural ranker used 30,000 expressions, 60,000 groups, and 299,645 candidates (298,211 valid, 1,434 failed), with 23,132 empty groups and zero replay mismatches. On test-IID, 10,752 groups were evaluable and 7,298 unevaluable: neural exact-best was 8,349/10,752 (77.65067%), mean regret 2.9454055, total regret 31,669, and 15.0162x exact-scoring speedup. EML-tree and AST-DAG heuristics were exact-best on 8,796 and 8,661 groups, respectively, and therefore outperform the neural ranker on the preregistered comparison. On test-OOD, neural was 1,499/2,034 (73.69715%), mean regret 3.09636185, total regret 6,298, and 5.15466x speedup; AST-DAG was 1,565/2,034 and again outperformed neural. These comparisons are null, not a claim of neural superiority. The production export contains exactly 250,000 expressions, 1,250,000 graph views, 250,000 hierarchies, and 2,500 batches, with zero validation or reconstruction failures. The published learned-motif MDL denominator is 75,000 rows for its MDL aggregate; it must not be presented as a 250,000-row denominator.