Benchmarks
Geltre claims are scoped to measured artifacts. Positive results and failed gates are both part of the public technical story.
ServiceOps v0.3
| Selector | Required-evidence recall @ 5 | JEV accuracy |
|---|---|---|
| BM25 | 0.6255 | 0.8235 |
| Geltre linear | 0.9961 | 1.0000 |
| Geltre neural | 0.9647 | 1.0000 |
The linear ServiceOps ranker is the current reference because it matched the downstream neural result while being smaller and faster.
CodeOps: keep the failure
| v0.8 metric | Actual | Required |
|---|---|---|
| Required-evidence recall | 0.808219 | ≥ 0.98 |
| Context reduction | 0.927891 | ≥ 0.50 |
The v0.8 cohort is consumed development evidence and is not reused as fresh proof.
CodeOps H256-R4 successors
Geltre then froze new cohorts around the H256-R4 terminal-verification selector instead of repairing v0.8 after the fact.
| Frozen cohort | Checkpoints | Required-evidence recall | Exact coverage | Context reduction |
|---|---|---|---|---|
| v0.12 | 215 | 1.000000 | 1.000000 | 0.714299 |
| v0.13 | 197 | 1.000000 | 1.000000 | 0.680607 |
v0.13 also froze the fail-closed reliability policy before scoring. It compressed 171/197 records, passed 26/197 through raw state, omitted 0 required evidence overall, and retained 0.716822 total context reduction after fallbacks.
Live v0.13 Geltre → JEV
| Lane | Decision accuracy | JEV input tokens | Recorded cost | Mean latency |
|---|---|---|---|---|
| Raw → JEV | 0.969543 | 1,685,006 | $0.07077025 | 259.852 ms |
| H256-R4 Geltre → JEV | 0.974619 | 597,196 | $0.02508223 | 218.378 ms |
Across all 197 paired records, Geltre caused 0 damage, rescued 1 Raw → JEV error, reduced JEV input/cost by 64.5582%, and reduced mean remote latency by 15.9606%. These are frozen-protocol results, not a universal cross-domain guarantee.
Release principle
Geltre must reduce context, cost, or distraction while preserving downstream task success better than ordinary retrieval. Context reduction alone is not sufficient. Negative cohorts remain part of the record even after later selectors improve.