S8n: filter/tool micro-window scheduler with witnessed sufficiency

Date: 2026-08-10
Issue: #6167
Inputs: S8m tool fold (bce85d6330)

Verdict

The controlled matrix supports a narrower general claim: bounded records make task-shaped stopping, queue-slot release, and selective stage admission executable and auditable. They do not make a model controller universally best. At identical fixture quality, the deterministic planner wins the existence query; adaptive routing wins top-k and exhaustive work under the declared calibrated stage model; and selective hedging improves the two expensive adaptive slices while universal hedging spends extra work. These are controlled scheduler milliseconds, not live provider latency, tokens, dollars, or billing.

Workload and authority

The frozen workload joins:

• 32 S8m records: 21 majority read_only, 11 majority current_state, preserving the source fold digest;
• 16 declared exact-filter controls, split 8 tune / 8 test, because S8m contains no none labels.

Six arms share 48 records, four workers, a 900 ms per-attempt deadline, and one quality oracle:

1. tuned run-all;
2. tuned fixed cascade;
3. deterministic planner;
4. adaptive three-seam micro-window scheduler;
5. adaptive scheduler with selective read-only hedging;
6. adaptive scheduler with universal hedging.

The stage catalog is closed: exact filter, control window, semantic filter window, repository read, and live read tool. No effect/write stage exists. A selector changes neither capabilities nor arguments. Receipts distinguish unopened, completed, timed-out, sufficiency-cancelled, and hedge-loser work.

Task-specific stop contracts

• Existence: stop after a witnessed positive; otherwise inspect all records.
• Top-k: stop only when the witnessed kth score is at least every unopened record’s structural upper bound. Merely collecting k candidates is insufficient.
• Exhaustive: never cancel primary work; all records must complete. Hedge losers may be cancelled.

Tests compare every early fold with the all-record oracle. Timeout receipts explicitly say deadline_released_slot; sufficiency cancellations say witnessed_sufficiency_slot_released.

Results

All 18 task/policy cells meet the controlled quality floor (quality=1.0). Values are deterministic calibrated milliseconds over five identical replay trials.

Task	Policy	Wall ms	Work ms	Opened	Cancelled	Hedges
existence	run-all	481.62	1,926.48	9	3	0
existence	fixed cascade	2.69	10.78	8	3	0
existence	planner	1.17	4.68	4	3	0
existence	adaptive	5.23	20.90	4	3	0
existence	adaptive selective hedge	6.34	25.38	4	3	0
existence	adaptive universal hedge	5.40	21.62	4	3	0
top-k	run-all	3,391.81	13,567.24	48	3	0
top-k	fixed cascade	1,426.21	5,628.62	48	1	0
top-k	planner	528.54	2,114.15	11	3	0
top-k	adaptive	323.24	1,292.95	12	3	0
top-k	adaptive selective hedge	267.25	1,069.02	12	3	0
top-k	adaptive universal hedge	288.08	1,152.33	14	3	0
exhaustive	run-all	3,528.44	13,735.35	48	0	0
exhaustive	fixed cascade	1,543.98	5,746.39	48	0	0
exhaustive	planner	2,329.48	8,884.92	48	0	0
exhaustive	adaptive	1,359.88	4,917.08	48	0	0
exhaustive	adaptive selective hedge	1,218.14	4,669.38	48	0	0
exhaustive	adaptive universal hedge	1,363.17	5,095.95	50	2	2

Interpretation:

• Existence has a structural answer and ordering signal, so a deterministic planner dominates the controller. This is an intended static-policy win, not a failure to hide.
• Top-k repays adaptive stage admission because it avoids irrelevant expensive stages and has a sound upper-bound stop. Selective hedging reduces both observed wall and total work in this deterministic schedule because faster winners release slots before expensive queued primaries open.
• Exhaustive work cannot use early stopping. Adaptive still avoids semantically inapplicable stages; selective hedging improves the modeled tail, while universal hedging opens two duplicates and loses.
• Run-all is a correctness reference but not a tuned economic winner for this heterogeneous catalog.

Artifact and exact rerun

Artifact: experiments/microcontext/s8n-filter-tool-scheduler-2026-08-10.json.

go run ./cmd/microcontextdemo `
  -filter-tool-scheduler-fold experiments/microcontext/s8m-semantic-tool-fold-2026-08-10.json `
  -filter-tool-scheduler-output experiments/microcontext/s8n-filter-tool-scheduler-2026-08-10.json `
  -filter-tool-scheduler-trials 5

go run ./cmd/microcontextdemo `
  -verify-filter-tool-scheduler experiments/microcontext/s8n-filter-tool-scheduler-2026-08-10.json

The verifier requires all 18 cells, quality 1.0, positive work/wall metrics, receipt digests, no exhaustive primary cancellation, source/gold provenance, and a complete claim boundary.

Steelman and falsification boundary

• Planner: wins when query structure exposes the right stage and stop rule, as existence does here. Adding a model controller would be overhead.
• Fixed cascade: remains easier to test and likely dominates under stable stage distributions without strong per-record heterogeneity.
• Run-all: is operationally simple and can be rational when catalogs are tiny, stages cheap, or audit costs dominate compute.
• Live endpoint: may reverse every calibrated ranking through network variance, provider batching, prefix caching, rate limits, cancellation billing, or model quality. S8n does not answer that question.
• Gold: S8m is majority-labeled and low-unanimity, with zero naturally adjudicated none records. Exact controls are explicit fixtures, not disguised adjudicator output.
• Hedging: selective’s controlled win depends on the tail predictor and free-slot rule. A poor predictor or billed cancellation can erase it; universal hedging already loses here.

Therefore S8n validates the scheduling mechanism and a static/adaptive decision boundary. It does not close #6033 or #6111; the quality-qualified live comparison remains not-yet.