Historical report imported 2026-10-10 from docs/experiments/benchmarks/research-costs-run-01.md. Original SHA-256: 10b98d3b0ab61339d07bb39de0689c4a25ea0a5c16566d2f0637a2fe5324f6db. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Receipt, layout and sidecar costs, run 01

Status: measured on one shared WSL2/ext4 host after the full E3 correctness gate. The separate JSON lifecycle remains the prototype baseline. This experiment compares specific research mechanisms; it does not migrate the application's storage format.

Registered contract and execution

The receipt/resume, layout/postings and collection/sidecar protocols were fixed before formal timing. Harness review clarified persisted metadata and durable index-build cost accounting before this run. Correctness evidence is separate: E3, S2, layout and collection/lifecycle.

python3 -B tools/bench/run_research_costs.py target/research-costs/run-01 --formal

Use a fresh output path on durable local storage. Release builds use locked offline dependencies. One warmup precedes five measured trials in alternating order. S1 mixed and shuffled-Log inputs use seeds 201/202/203 and 2,048 events; the layout-only scale cell uses seed 201 and 8,192 events. There are 128 queries per small snapshot and 512 per scaled snapshot. The E3 fault corpus has deliberately different duplicate-ID and query-offset mutations; it is the correctness gate, not this cost fixture.

The runner exited 0, with 490 child commands, all exit 0. It ran from 2026-09-27T13:40:04.331398+00:00 to 2026-09-27T14:25:59.211819+00:00 on /dev/sdd ext4 /. The recorded CPU is Intel(R) Core(TM) i7-10750H CPU @ 2.60GHz. Every timed answer passed exact positional and full-row-digest equality; sidecar variants also reproduced the same snapshot root. Warmups are retained but excluded from every table below.

The raw results and command ledger (local evidence omitted from stage) preserve individual timings, CPU, peak RSS, source/output hashes, calibration and kernel I/O counters. Descriptive metrics (local evidence omitted from stage) contain per-phase CPU/wall medians, p50/p99 by mode, every query family's paired ratios and resource costs. Phase medians use five trials; query p50/p99 values are medians of each trial's nearest-rank percentile. Gates are evaluated separately per dataset.

Hybrid layout result

Each candidate stores the complete raw Event as well as projected predicate columns. All query paths authenticate the full file; projection saves decoding, not that byte pass. The layout table includes external table anchors and excludes optional postings. The raw summary also records totals with postings and their external digests. Footer bytes are already inside Parquet. Independently retained source JSON is reported separately and is common to all variants; the table below does not include that source or the application's FOL2 journal. These are final retained bytes, not filesystem allocation, replication or compaction write amplification.

In each candidate cell: bytes / projected-time ratio / gate. Bytes are relative to JSON including corresponding trust metadata; time compares median trial p50s. The registered gate requires bytes <=80% and time <=1.10x JSON.

Dataset (shape–seed–rows)	JSON bytes	JSON p50 ms	Plain/64	Zstd/64	Zstd/256
mixed-201-2048	723,381	5.427	153.00% / 0.911x / fail	25.90% / 0.327x / pass	19.02% / 0.204x / pass
mixed-202-2048	723,381	5.475	153.00% / 0.916x / fail	25.90% / 0.336x / pass	19.01% / 0.198x / pass
mixed-203-2048	723,381	5.855	153.00% / 0.866x / fail	25.93% / 0.315x / pass	18.98% / 0.190x / pass
shuffled_logs-201-2048	826,626	5.788	167.14% / 1.067x / fail	23.15% / 0.339x / pass	17.65% / 0.213x / pass
shuffled_logs-202-2048	826,626	5.860	167.14% / 1.048x / fail	23.15% / 0.326x / pass	17.65% / 0.206x / pass
shuffled_logs-203-2048	826,626	5.768	167.14% / 1.065x / fail	23.20% / 0.331x / pass	17.64% / 0.198x / pass
mixed-201-8192	2,898,507	22.184	152.93% / 0.901x / fail	25.74% / 0.325x / pass	18.83% / 0.182x / pass
shuffled_logs-201-8192	3,313,152	39.194	167.07% / 0.955x / fail	23.03% / 0.291x / pass	17.56% / 0.166x / pass

plain64 passes 0/8 dataset gates.

zstd64 passes 8/8 dataset gates.

zstd256 passes 8/8 dataset gates.

The small synthetic datasets, duplicated schema and cached reads limit transfer to production. Log bodies use repeated templates and 128 repeated x bytes, favoring compression; these ratios need a less repetitive workload before generalization. Compression and larger row groups can be promising without selecting a new durable format. Per-family comparisons and full-materialization p99 remain in the descriptive metrics; the aggregate gate does not imply every family improves.

Selective postings

Only rare is indexed. Each cell gives the measured break-even number of rare-token queries, followed by median per-query saving in microseconds. Amortization includes pure construction plus durable publication of postings and their external digest. Index bytes/digest are resident during querying; table reading and authentication remain inside each query. Non-indexed predicates fall back to projection.

Dataset	Plain/64	Zstd/64	Zstd/256
mixed-201-2048	18 / 241.7 µs	22 / 198.3 µs	25 / 191.8 µs
mixed-202-2048	15 / 263.2 µs	19 / 215.2 µs	21 / 202.0 µs
mixed-203-2048	13 / 319.6 µs	21 / 216.4 µs	22 / 200.6 µs
shuffled_logs-201-2048	13 / 404.1 µs	12 / 381.2 µs	13 / 386.7 µs
shuffled_logs-202-2048	20 / 285.4 µs	18 / 351.4 µs	13 / 413.8 µs
shuffled_logs-203-2048	16 / 328.2 µs	16 / 346.7 µs	14 / 358.1 µs
mixed-201-8192	11 / 658.5 µs	8 / 786.0 µs	10 / 722.2 µs
shuffled_logs-201-8192	8 / 1982.0 µs	6 / 1855.0 µs	5 / 1708.5 µs

This is an observed amortization calculation, not a future-query guarantee. It uses the median trial saving and median trial build-plus-publication cost. Repeated rare queries, residency and unchanged snapshot identity are assumptions. Correctness tests cover absent, malformed and wrong-table postings and deterministic rebuild separately.

Receipt and retry overhead

The direct scalar comparison scans the source Vec and hashes matching full Events. The receipt comparison executes E3 and verifies the resulting page. Serialization is outside query timing; wire bytes are reported separately. These are in-memory query costs, not time-to-durable-search or proof of an untrusted executor's work.

Each cell below is the median paired receipt/scalar wall-time ratio for that query family, preserving all favorable and unfavorable predicates.

Dataset	All rows	Narrow time	Tenant	Rare	Absent	Common	Combined	Empty range
mixed-201-2048	1.17x	12.88x	75.93x	4.09x	5.37x	1.26x	91.91x	131.42x
mixed-202-2048	1.17x	12.94x	74.81x	4.15x	5.38x	1.26x	92.01x	164.28x
mixed-203-2048	1.17x	12.94x	67.53x	4.03x	5.25x	1.25x	78.74x	100.27x
shuffled_logs-201-2048	1.15x	6.36x	98.19x	2.93x	3.54x	1.15x	84.61x	151.04x
shuffled_logs-202-2048	1.16x	6.41x	100.07x	2.91x	3.61x	1.16x	86.40x	183.85x
shuffled_logs-203-2048	1.15x	6.53x	98.55x	2.94x	3.61x	1.16x	89.45x	183.12x

Constructor readiness, verification and retries have separate boundaries. The table shows median wall milliseconds. S1 construction includes its conservative summaries; sealed construction includes exact summaries and commitments, so their difference is not hashing alone. Verify checks an already-created all-rows page. Resume executes two complementary pages and merges to closure. Replay merges the same page again.

Dataset	S1 build ms	Sealed build ms	Verify ms	Two-page resume ms	Replay ms	All-rows page B	Two pages B
mixed-201-2048	0.388	3.087	0.275	5.596	0.344	336,959	367,370
mixed-202-2048	0.394	2.845	0.242	4.285	0.251	336,968	367,304
mixed-203-2048	0.421	3.084	0.240	4.342	0.266	336,921	367,115
shuffled_logs-201-2048	0.706	4.529	0.267	5.583	0.340	337,084	367,455
shuffled_logs-202-2048	0.731	4.697	0.237	5.058	0.265	336,965	367,329
shuffled_logs-203-2048	0.705	4.429	0.236	4.649	0.240	337,045	367,458

The full-page byte field in each raw result is the sum over 128 queries; it must not be compared with the two-page all-rows retry as if it were one page. The table uses the all-rows family's single-page size instead. Residual sizes and all family page sizes remain in the metrics. No universal receipt speed gate was registered: the prototype pays this overhead to expose incomplete coverage and reject mismatched resume work. Timing calibration is retained without subtraction; ratios involving very cheap scalar predicates are sensitive to timer overhead.

Conservative sidecar result

The source constructs hints from the same retained Events. The server checks those hints against every raw block before constructing its root. No evidence is dropped. Each value is a median paired phase-CPU ratio. Benefit requires both server CPU <=0.90x baseline and total source+server CPU <=1.10x baseline.

Dataset	Server CPU ratio	Total CPU ratio	Hint bytes	Source bytes	Both gates
mixed-201-2048	0.848x	1.669x	32397	720492	fail
mixed-202-2048	1.028x	1.809x	32397	720492	fail
mixed-203-2048	1.045x	1.815x	32397	720492	fail
shuffled_logs-201-2048	1.040x	1.745x	49118	823737	fail
shuffled_logs-202-2048	0.992x	1.670x	49118	823737	fail
shuffled_logs-203-2048	1.026x	1.721x	49118	823737	fail

Missing, malformed, stale and underinclusive hints each triggered central rebuild and preserved the exact answers. Their phase/resource costs are retained alongside the valid and baseline runs. These are logical file-transfer payload sizes, not network measurements. Validating raw data centrally can erase the proposed source work saving; changing that trust boundary would require a different contract.

Review, limitations and decision impact

Before formal execution, independent GPT and Claude reviewers exercised the harness and its evidence validator. Review found that a mutated stored percentile could reverse a gate, and that external metadata and durable index publication were omitted from cost accounting. The repaired harness recomputes percentiles, enforces the sample grid and matched counts, validates retained file sizes/hashes, and includes the full publication boundary. Injected percentile, missing/duplicate sample, negative/bool timing, wrong count, source/answer mismatch and fsync-error defects were rejected. The review archive (local evidence omitted from stage) retains failures and source-pinned approvals; smoke trials are not mixed into formal results.

The independent formal audit (local evidence omitted from stage) exited 0, rechecking all eight cells, 490 successful commands and 4,436 original file hashes. It recomputed every gate and every query family's observations; a forged gate with an updated hash manifest exited 1 as intended. Twenty-eight descriptive metric spot-checks matched exactly. This audit was independent of the harness summarizer; Claude reviewed the harness and smoke behavior before formal execution. All 29 measured source hashes (local evidence omitted from stage) matched integration before a later import-only rustfmt adjustment (local evidence omitted from stage). The selected layout body is unchanged, and its post-format tests are recorded in the final check ledger.

The retention manifest (local evidence omitted from stage) identifies regenerable data files omitted from Git. The full original was audited before omission; the retained samples and metadata remain sufficient to recalculate the reported costs.

This host is shared: start/end process lists and load averages are evidence of load, not isolation. E3 had finished before timing; no cache drop was performed. Most reads can hit the OS cache. logical_data_bytes, process read syscalls, kernel device bytes and retained storage bytes are distinct metrics. Layout execution order alternates between trials, but each query runs full, projected and then postings in fixed order; this can favor later modes through cache effects. Correctness prechecks also read the files before timed queries. CLOCK_PROCESS_CPUTIME_ID measures phase CPU; whole-process CPU/RSS include setup and oracle checks and are reported separately. A fresh process does not make the OS page cache cold. The resource record does not establish device IOPS or real-network cost.

The measurements justify further compressed-layout experiments, subject to their per-dataset gates. They do not justify migrating FOL2, weakening verification, automatically creating indexes, or dropping raw source data. The local lifecycle continues with inspectable S2 JSON snapshots. Hardware offload, external-system comparisons, network OTLP, clustering, event-time partition redesign and adaptive indexes remain explicitly conditional future work in the agenda.