S8s/S8t — natural multi-tool admission decision surface

Verdict: decision_surface, not a universal adaptive winner. Natural held-out windows across five evidence classes produced a real quality/cost tradeoff: the deterministic planner was fastest but only 8/10 correct; fixed cascade and both semantic policies were 10/10. At observed tool latency, fixed cascade was fastest among quality-qualified arms. At a modeled 10× tool-cost regime, semantic two-stage became fastest and opened 8 tools rather than 40.

Corpus and independent fold

Twenty naturally phrased operator questions cover:

• packet-only / no tool;
• committed source search;
• mutable GitHub issue state;
• current git commit ancestry;
• named documentation reads.

Opaque alternating IDs freeze 10 tune and 10 test records, with two held-out records per class. Independent gpt-5.6-sol and Groq llama-3.3-70b-versatile judgments are blind to the construction label. The strict fold requires both model-distinct votes and the construction contract to agree.

Result: 19/20 unanimous, model pairwise agreement 0.95. The sole dispute (n01, tune) remains excluded; every held-out record is unanimous. The verifier requires at least two unanimous examples in every class.

Real tool receipts

Every held-out required tool was executed through its actual read seam:

• git grep for source search;
• bounded GitHub issue-state REST/gh fallback;
• git merge-base --is-ancestor for commit state;
• bounded committed-file reads for docs;
• no dispatch for packet-only questions.

Observed receipts and latency are preserved per record. The cheap (0.1×) and expensive (10×) regimes scale those witnessed latencies without changing labels; they are explicitly modeled rather than presented as additional live measurements.

Decision surface

regime	policy	exact	tools	mean wall ms	mean work ms	prompt/output
cheap 0.1×	deterministic	8/10	6	20.1	20.1	0 / 0
cheap 0.1×	fixed cascade	10/10	40	102.8	102.8	0 / 0
cheap 0.1×	semantic two-stage	10/10	8	5,625.2	5,625.2	1,626 / 644
cheap 0.1×	selective parallel	10/10	40	5,604.6	5,707.4	1,626 / 644
observed	deterministic	8/10	6	201.4	201.4	0 / 0
observed	fixed cascade	10/10	40	1,028.5	1,028.5	0 / 0
observed	semantic two-stage	10/10	8	5,810.3	5,810.3	1,626 / 644
observed	selective parallel	10/10	40	5,604.6	6,633.1	1,626 / 644
expensive 10×	deterministic	8/10	6	2,014.0	2,014.0	0 / 0
expensive 10×	fixed cascade	10/10	40	10,284.7	10,284.7	0 / 0
expensive 10×	semantic two-stage	10/10	8	7,661.5	7,661.5	1,626 / 644
expensive 10×	selective parallel	10/10	40	8,217.4	15,889.3	1,626 / 644

Quality is gated before latency: deterministic cannot be called the winner because it misses two evidence classes on held-out records. Among 10/10 arms:

• cheap and observed tools: fixed cascade wins because semantic selection costs several seconds;
• expensive tools: semantic two-stage wins, reducing mean wall 25.5% versus fixed cascade and opening 80% fewer capabilities;
• selective parallel buys lower tail than serial selection only when tool latency approaches selector latency, but spends the most total work and authority.

General micro-window interpretation

The policy should be chosen per bounded window and cost regime, not globally:

1. deterministic structural routing first where contracts are explicit;
2. fixed cascade when tools are cheap and quality requires broad evidence;
3. semantic admission when tool cost/authority exceeds selector cost and evidence need is predictable;
4. selective parallelism when tail latency matters enough to pay duplicated work;
5. typed folds and provenance regardless of admission policy.

This supports the user’s general pattern while rejecting the strongest overclaim. Micro-windows provide a common scheduling unit for large inputs, filters, models, and tools; they do not imply one scheduler dominates every workload.

Reproduction

go run ./cmd/microcontextdemo -verify-natural-multitool-corpus experiments/microcontext/s8s-natural-multitool-corpus-2026-08-10.json -verify-natural-multitool-fold experiments/microcontext/s8s-natural-fold-2026-08-10.json
go run ./cmd/microcontextdemo -verify-natural-multitool-surface experiments/microcontext/s8t-natural-multitool-surface-2026-08-10.json
go test ./cmd/microcontextdemo -run 'TestNatural' -count=1