Axis 4 · Cohort random access

axis4-threads-1

{
  "affinity": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15
  ],
  "axis": 4,
  "corpus": "d6eb104bff84bec7ee9f5cf23580d4cc80b0fb132570fe0737dbee702a0b1d77",
  "cpu_models": [
    "AMD EPYC 4344P 8-Core Processor"
  ],
  "domain": "uppercase-sequence-without-FASTA-framing",
  "kind": "official",
  "machine": "ace-core",
  "protocol": "4a0027fec59b3252dc07b29458a57a5a451957f1b64b36f3927765cbc6be7a5c",
  "requests": "79638817948218e1b0d8ae3f3ecbe555ace6c33d8c8d3ad098c821e591384b61",
  "schema": "axis4-evidence-v5",
  "scope": "cpu-in-process",
  "threads": 1,
  "timing_boundary": "complete group: native fetch, uppercase normalization, SHA-256 judgement and translation; response file writes outside timer"
}

axis4-threads-16

{
  "affinity": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7,
    8,
    9,
    10,
    11,
    12,
    13,
    14,
    15
  ],
  "axis": 4,
  "corpus": "d6eb104bff84bec7ee9f5cf23580d4cc80b0fb132570fe0737dbee702a0b1d77",
  "cpu_models": [
    "AMD EPYC 4344P 8-Core Processor"
  ],
  "domain": "uppercase-sequence-without-FASTA-framing",
  "kind": "official",
  "machine": "ace-core",
  "protocol": "4a0027fec59b3252dc07b29458a57a5a451957f1b64b36f3927765cbc6be7a5c",
  "requests": "79638817948218e1b0d8ae3f3ecbe555ace6c33d8c8d3ad098c821e591384b61",
  "schema": "axis4-evidence-v5",
  "scope": "cpu-in-process",
  "threads": 16,
  "timing_boundary": "complete group: native fetch, uppercase normalization, SHA-256 judgement and translation; response file writes outside timer"
}

Methodology, definitions and acceptance criteria

Complete source methodology · SHA-256 9b1ce9415146ebd7588791bdff318e58610dfd23afba3d5b40800f3e40b339e7

Canonical bytes and coordinates

Truth is uppercase sequence bytes without FASTA headers or line separators. Preserve the original source objects unchanged. A common request is (assembly_id, contig_id, start0, end0), using zero-based half-open coordinates [start0,end0). It is valid iff 0 <= start0 < end0 <= contig_length; its length is exactly end0-start0. A request never crosses a contig boundary.

Adapters translate only at the library boundary:

Record both canonical and translated coordinates with their conventions in evidence. A native FASTA-aware reader may use a verified FASTA index internally; raw FASTA offsets are never substituted for canonical coordinates. The external request space is per contig even when a codec stores concatenated sequence bytes. Do not clip, substitute another contig, shrink W or repair an invalid/empty request. Such requests produce FAILED. Contigs shorter than W have no valid starts for that W; sampling excludes those starts, not failures after decoding.

Same machine and execution boundary

CPU in-process, GPU in-process and process-per-request CLI are separate scopes; never rank across them. CPU in-process uses one persistent library handle and one decode thread. Open/index/reference setup occurs outside window timing; the reader operation, including its required lookups and decoding, remains inside. Do not silently substitute a helper process, a decoded-data cache or a different machine. Declare residency and use the same access/cache policy for each comparable configuration.

The official host must be ace-core. Before running, retain hostname, uptime/load, one-second process CPU deltas, frequency policy and available storage. The gate requires load <0.5, storage >20,000,000,000 bytes, and no foreign process consuming 10% or more of one CPU over the sample interval. Keep thresholds and raw telemetry. A passing sample is not proof of continuous quiescence. A GitHub-hosted runner is not a substitute for ace-core, and a green PR does not authorize a measurement.

Axis 4: cohort random access (A6)

Axis 4 maps each query group explicitly to every selected assembly. The CLI implements prepare, silence, native run, evidence and independent verify. Prepare binds each original FASTA SHA-256, contig order and length, the complete ordered query groups, protocol and native-reader plan. Coordinates and canonical bytes follow the definitions above. No homology is inferred between contigs of different assemblies.

The timed group includes native fetch, uppercase normalization, SHA judgement and coordinate translation; response-file writes are outside timing. The recorded boundary is not interchangeable with decoder-only Axis 3 timing. Stored bytes charge archives and required indexes/references. Synthetic seconds are not official performance fields. Unsupported/missing values remain absent; the leaderboard never imputes zeros.

verify requires an external prepare-SHA anchor, checks FASTA hashes against prepare, scans FASTA independently of the native reader, and compares the SHA of every answer. It independently reproduces the seed-selected sample IDs and compares retained answer bytes too. The sample limit is 32 responses, established by the A6 native log; hashes cover all responses. A changed FASTA, SHA, duplicate ID, altered coordinate, bad length or retained byte mismatch fails verification. Official evidence additionally requires passing before/after silence telemetry. SHA binds bytes, not authorship or continuous host quiescence.

Schema: axis4-evidence-v3. PASS requires every declared group/response to be represented and verified; reader/data failure cannot become a valid performance row. Native availability alone does not establish correctness, large-input coverage or a throughput win.