Controlled result
In a frozen matched benchmark using gpt-5.6-sol at xhigh, across 36 fresh sessions and six task classes, context-applicable tasks used 31.20% fewer provider-native worker tokens in aggregate. The median reduction was 33.03%, with zero quality regressions.
Applicable-task proxies also reduced broad searches from 18 to 0 and repository read commands from 25 to 7.
What the mechanism does
- Selects relevant reviewed memory and repository evidence within a bounded budget.
- Requires exact-revision, current evidence before substituting it for redundant discovery.
- Keeps provenance, validation and contradiction fallback mandatory.
- Escalates progressively when the initial evidence does not answer a safe question.
Limitations
- The claim applies only to tasks where relevant exact-revision context existed.
- Unknown/new tasks received no ReasonTrail context and do not support a savings claim.
- The result is not proven across every model or repository.
- It measures provider-native worker usage, not whole-Mission billing or planning/review overhead.
- A one-pair Terra check is directional only.