Atlas Platform

Document Summariser

complex-agenticconfidence high — a separate axis from tier; low confidence never auto-decides
Atlas PlatformKnowledge Baselabel · app=knowledge-baseDocument Summariserwf-at021c9read throughLLAtlas Gatewayhttps://llm.atlas.internal

Where this came from

Every figure below is only as trustworthy as the account it was read through

Application
Knowledge Base
identified by label · app=knowledge-base
Project
Atlas Platform
Atlas Logistics
Fingerprint
wf-at021c9
inferred from logs

Cost

30-day spend

$12,848

LOGGED

On recommended path

$12,452

MEASURED

Share of attributed spend

81%

Current → recommended (diffed against production)

In production today

gemini-2.5-pro

reasoning high · cap 8,192

Recommended

gemini-2.5-pro (unchanged)

reasoning low · cap 2,048

  • Δ reasoning: high → low
  • Δ response cap: 8,192 → 2,048

Replay evidence

Model

gemini-2.5-pro · reasoning low

Samples

120

Input tokens

1,480,000

Output tokens

61,000

Reasoning tokens

19,000

Latency

1850 ms end-to-end

Judge verdict

HOLDS

Replay spend

$6.40

MEASURED — the only grade replay can produce, and the only stage that spends money
cannot saysession identifiers do not exist in this log source