4,047,104 recorded tokens including cached input; 84.9% of analysis/coordinator input token volume was cache reads. Cache reads are still tokens—not token elimination. The separate verifier logs contain 499,967 “tokens used” across 86 logs without an input/output/cache split; these are not folded into the provider buckets above.

Reprice the recorded analysis/coordinator token mix at DeepSeek’s published USD rates checked 27 September 2026. The original run records the alias “sonnet” and a $4.766 subtotal for these transcripts. This is a rate-card calculation, not a DeepSeek execution, a whole-run bill, or a quality comparison.

Formula: cache-read tokens × cache-hit rate + (uncached input + cache-write tokens) × cache-miss rate + output tokens × output rate, divided by one million. Verifier charges, hardware, ingestion, and service costs are excluded. Tokenizers, reasoning length, cache behavior, quality, and latency can differ in an actual run.

DeepSeek’s official API also supports tools and structured output. A fair comparison is Jianmu’s orchestration against a specified tool-enabled baseline on identical data—not “agents versus a model that cannot use tools.”

The read-only audit checks the event sequence, final verdict corrections, dataset-manifest hash, and unique session IDs. Usage totals use only top-level session summaries; nested iterations are not added again. 38 proposals finish with 13 verified labels, 2 hypotheses, and 23 rejected labels. These are review states, not independently scored accuracy. The 34 seats are distinct roles across the run, not a simultaneous concurrency measurement.

The manual baseline explicitly assumes a person dispatches and reviews each of 500 packages: 500 × 2 + 5 project gates = 1,005 workflow touches, plus one context handoff. These are configured human actions, not 1,005 mandatory security approvals or a requirement imposed by every competing product. Automated MAS can implement comparable control policies.

Context arithmetic: 500 compact results × 2,000 tokens + 400,000 tokens for instructions, decisions, and working state = 1.4M unique controller-facing tokens. With 20% headroom in a 1M window, usable capacity is 800k; ceil(1.4M / 800k) = 2 segments. This simplified budget excludes repeated transmission and assumes handoff summaries fit the reserved capacity. Worker model traffic is separate; the scenario’s 1.74B billed/processed tokens cannot be divided by 1M to infer session count.

1. Authorize the objective, data scope, and boundaries.
2. Approve the analysis plan and acceptance criteria.
3. Review a representative sample before expanding work.
4. Resolve material contradictions and exceptional decisions.
5. Accept the final evidence package and authorize delivery.

Guided mode reserves these decisions for a person. The illustrative automatic policy predefines in-scope decisions and acceptance rules. It does not bypass human authority, erase verification, or guarantee no interruptions.

No drift probability is assigned here: measure constraint retention, citation correctness, missed dependencies, and recovery over long tasks and repeated handoffs. More agents and a larger context window alone establish none of those rates.

Token reduction = 1 − 1,740 / 42,000 = 95.857%. Wall time = token total / effective tokens per second / 3,600 + non-model critical-path allowance. Raw pipeline allowance: 10 h; MAS: 12 h; Jianmu: 10.4 h. No extra multiplier is applied for 32 slots: their effect is already represented by aggregate throughput.

The raw baseline re-reads extracted material with additional analysis/review passes. The illustrative MAS adds repeated shared context and review messages. Jianmu’s 1.74B budget assumes selective evidence routing and script execution. These allocations are chosen scenario assumptions—not tokenizer logs or measured extraction ratios. Optimized retrieval-based or tool-using systems can narrow the difference. No model vendor or MAS framework is represented by these baselines.