More useful work.
Less wasted effort.
THE LAB NOTEBOOK / LAVETT
Cost. Context. Tool calls. Time. Quality.
Useful efficiency has to account for all of them.
These historical internal tests compare the same tasks with and without Lavett. They show promising reductions and real tradeoffs. They do not certify the current release or predict your own results.
EVERYDAY TASKS / TWO REPEATS
8–11%
$0.922 → $0.821 in repeat 1; $0.916 → $0.845 in repeat 2.
16 synthetic tasks repeated twice. Cost-equivalent 95% intervals: 7.8–14.6% and 3.3–12.6% lower. The first repeat’s answer-time interval crosses zero.
01 / THE WHOLE PICTURE
96 native runs: 48 paired comparisons over four test groups. One run per arm per case. Synthetic tasks, one model and effort, quality reviewed by the harness author.
| Workload | Pairs | Cost equivalent without → with | Largest request | Tool calls | Answer time |
|---|---|---|---|---|---|
| Everyday, repeat 1 | 16 | $0.922 → $0.821 11.0% lower | 2.4% smaller | 29 → 20 | 7.0% less |
| Everyday, repeat 2 | 16 | $0.916 → $0.845 7.7% lower | 2.4% smaller | 30 → 20 | 10.7% less |
| History ~41 KB | 8 | $1.364 → $0.666 51.2% lower | 45.4% smaller | 19 → 14 | 3.1% more |
| History ~240 KB | 8 | $1.938 → $0.858 55.7% lower | 46.3% smaller | 44 → 18 | 37.4% less |
Time above excludes history ingestion. “Cost equivalent” is Claude’s reported list-price calculation, including its token accounting. Actual subscription cost and allowance consumption are separate. No universal model or client claim follows.
Median history ingestion was 0.07 seconds for everyday tasks, 1.48 seconds at 41 KB, and 30.7 seconds at 240 KB. Including ingestion for each of eight first questions, the 240 KB group took 202.6 seconds without Lavett and 373.8 seconds with it. First visible answer text was not faster in the pooled study.
Account meters were recorded, but other activity used the same account. We cannot isolate the test’s effect on those meters. Actual bill or allowance savings are not established. This is the outcome Lavett is working toward; token or price-equivalent reductions alone do not prove it.
Paired, seeded percentile bootstrap intervals over tasks (95%). Cost-equivalent reductions: everyday r1 7.8–14.6%, r2 3.3–12.6%, 41 KB 38.1–60.2%, 240 KB 40.9–65.0%. Answer-time reductions: r1 −0.9–13.9%, r2 3.0–18.6%, 41 KB −23.9–12.9%, 240 KB 28.9–43.6%. Negative time reductions mean slower answers. These intervals do not cover generalization to other workloads or current software.
02 / QUALITY IS PART OF EFFICIENCY
No measurable answer-quality difference was found in the reviewed study. Automated checklist differences were phrasing misses in both arms; they are not an accuracy score. The author’s review found no superseded value presented as current in either arm.
A general hallucination-reduction rate has not been measured. A source receipt proves where quoted words occurred, not that the source or the model’s conclusion is true.
All eight 41 KB packets contained the required notes. One of eight 240 KB packets missed a required note; the model found it through three tool calls and answered correctly. Near-duplicate filler had crowded the note out.
Lexical retrieval can miss paraphrases or relevant evidence. The ×6 test history also included future dates that caused hedging in both arms. These are limitations of the test and retrieval path, not hidden successes.
03 / KEEP THE LOSSES IN VIEW
An earlier full v2 pilot reduced aggregate input and time but increased price equivalents by 0.14% for Codex and 4.3% for Claude. Earlier integration also added substantial overhead. The lesson: smaller context alone is not the same as lower total cost.
These dated studies are directional evidence. A current candidate needs its own verification. No comparison here establishes superiority over another memory product, raw context or BM25 across general tasks.
Source records: docs/evidence-ledger.md, LED-032 and LED-036; docs/native-client-benchmark.md; the 23 September 2026 quality and accounting review. Public export of raw runs is pending privacy review. Ask about the evidence ↗
UNDER THE NUMBERS