Applicable tasks
31.20%
lower aggregate provider-native worker tokens
Median reduction: 33.03%. Quality regressions: 0.
Evaluation · controlled benchmark
Trust should be measurable.
ReasonTrail publishes the bounded evidence behind its context-efficiency claim and the limits on what that evidence means.
Discovery proxies
- Broad searches
- 18 → 0
- Repository reads
- 25 → 7
These are applicable-task counts, not a universal promise about every session.
Method
Frozen, matched and provider-native.
- gpt-5.6-sol, reasoning xhigh
- 36 fresh matched sessions across six task classes, three repetitions per arm/task
- Provider-native counters with deterministic quality gates
- Exact-revision context, repository reads and broad-search counts retained as evidence
Limitations
What this does not prove.
- • Only context-applicable tasks are included in the supported percentage claim.
- • Unknown/new tasks received no ReasonTrail context and do not support a savings claim.
- • The result is not proven across every model, repository or task shape.
- • The benchmark measures provider-native worker usage, not whole-Mission billing.
- • Planning and review calls were outside the worker benchmark.
- • The single Terra confirmation pair is directional evidence only.
Interpretation
Less rediscovery, not a blanket guarantee.
The supported claim is: “In controlled gpt-5.6-sol xhigh benchmarks on tasks where ReasonTrail had relevant exact-revision context, ReasonTrail reduced provider-native worker token usage by 31.20% in aggregate while preserving the frozen quality gates.”