{
  "schema_version": "1.0.0",
  "id": "U02",
  "record_type": "static_evidence_access_guide",
  "path": "/for-ai/",
  "title": "Query Deployment Evidence with Its Conditions",
  "modified_on": "2026-10-18",
  "source_of_truth": [
    "/catalog.json",
    "/schema/community-catalog-3.1.0.json",
    "/articles.json"
  ],
  "evidence_boundary": {
    "site_reproduced_records": 0,
    "querying_is_model_benchmark": false,
    "server_side_query_API": false,
    "unknown_values": "null; preserve known false and zero; absent optional fields do not mean false or zero"
  },
  "field_groups": [
    "catalog",
    "articles"
  ],
  "fields": [
    {
      "collection": "catalog",
      "path": "schema_version / dataset_version",
      "type": "string / string",
      "meaning": "Contract version and editorial dataset revision are separate. Catalog schema is 3.1.0; a new explainer does not silently change the source dataset."
    },
    {
      "collection": "catalog",
      "path": "source_count / records",
      "type": "integer / array",
      "meaning": "Number of original reference records and the complete collection. This collection has 59 entries."
    },
    {
      "collection": "catalog",
      "path": "records[].id / kind",
      "type": "string / model | deployment | benchmark",
      "meaning": "Stable reference identity and source role. A benchmark role does not mean the site ran it."
    },
    {
      "collection": "catalog",
      "path": "records[].model_family / runtime_topics / hardware_topics",
      "type": "string / string[] / string[]",
      "meaning": "Discovery facets. Topic membership is not proof that a particular artifact works on a device."
    },
    {
      "collection": "catalog",
      "path": "records[].source.publisher / pinned_url",
      "type": "string / URL",
      "meaning": "Original publisher and revision-pinned document. Preserve attribution when extracting a result."
    },
    {
      "collection": "catalog",
      "path": "records[].source.revision / document_sha256",
      "type": "string / SHA-256",
      "meaning": "Document-repository revision and digest of the document bytes. Neither is a weight hash or proof of factual correctness."
    },
    {
      "collection": "catalog",
      "path": "records[].source.checked_on / last_commit_at",
      "type": "date / timestamp or null",
      "meaning": "Editorial check date and source commit timestamp. Neither substitutes for the experiment date."
    },
    {
      "collection": "catalog",
      "path": "records[].model_artifact",
      "type": "object or null",
      "meaning": "The exact model/weight artifact when stated. Its id and revision identify weights, not an internal content-generation model."
    },
    {
      "collection": "catalog",
      "path": "records[].conditions.runtime_execution_revision",
      "type": "string or null",
      "meaning": "Executed inference-engine revision, only when the source supplies it. The report repository revision is a different identity."
    },
    {
      "collection": "catalog",
      "path": "records[].reported_setup / conditions / test_profiles",
      "type": "object / object / array",
      "meaning": "Report-wide scope, shared declarations and distinct test shapes. Do not merge profiles into an invented common run."
    },
    {
      "collection": "catalog",
      "path": "records[].recipe_parameters / deployment_options",
      "type": "object or null / array",
      "meaning": "Recipe declarations and selectable configurations. Planning memory and context settings are not measured peak memory or achieved context."
    },
    {
      "collection": "catalog",
      "path": "records[].editorial.reported_results[]",
      "type": "object[]",
      "meaning": "Limited attributed source results. Carry each complete point and its own conditions instead of just a value."
    },
    {
      "collection": "catalog",
      "path": "...reported_results[].metric / value / unit",
      "type": "string / number / string",
      "meaning": "The reported quantity, numeric value and unit. Served rate and decode rate have different timing denominators; GB and GiB differ."
    },
    {
      "collection": "catalog",
      "path": "...reported_results[].statistic / uncertainty",
      "type": "string / object or null",
      "meaning": "Reported statistic and supplied dispersion. SD, a spread ratio and an observed range are different quantities. Null is not zero uncertainty."
    },
    {
      "collection": "catalog",
      "path": "...reported_results[].conditions",
      "type": "object",
      "meaning": "The result-specific model, hardware, input/output shape, concurrency, cache, speculation and measurement definition. Retain missing fields."
    },
    {
      "collection": "catalog",
      "path": "...reported_results[].evidence_url / source_document_sha256",
      "type": "URL / SHA-256",
      "meaning": "Exact original location and the matching document digest. Use the digest of that document, including supplemental documents where applicable."
    },
    {
      "collection": "catalog",
      "path": "records[].evidence_state / site_reproduced / metrics",
      "type": "source_reported / false / null",
      "meaning": "Current evidence is attributed to sources. The legacy site-measurement metrics field is empty; this does not say the original author measured zero."
    },
    {
      "collection": "catalog",
      "path": "records[].comparison_status / performance_ranking_allowed",
      "type": "string / false",
      "meaning": "Discovery is not a cross-condition ranking. Unknown identity or mismatched workloads prevent a direct performance verdict."
    },
    {
      "collection": "catalog",
      "path": "records[].license",
      "type": "object",
      "meaning": "Recorded repository terms with their stated scope. Model, conversion, dataset and container terms need their own reading."
    },
    {
      "collection": "articles",
      "path": "schema_version / source_count / articles",
      "type": "1.0.0 / integer / array",
      "meaning": "Separate explainer envelope. Its legacy source_count counts article entries, not unique original reports; do not add it to catalog source_count."
    },
    {
      "collection": "articles",
      "path": "articles[].id / path / body_markdown",
      "type": "string / path / Markdown",
      "meaning": "Stable note identity, reading URL and the same reviewed body used by HTML and Markdown exports."
    },
    {
      "collection": "articles",
      "path": "articles[].sources / supplemental_sources",
      "type": "array / optional array",
      "meaning": "Primary catalog references and distinct supporting documents. Each supporting document has its own document_id, revision, digest and parent_record_id."
    },
    {
      "collection": "articles",
      "path": "articles[].facts / declared_variants",
      "type": "array / optional array",
      "meaning": "Source declarations or recipe selections with their classifications. A declared planning minimum is not a measured run."
    },
    {
      "collection": "articles",
      "path": "articles[].reported_results / conditions",
      "type": "array / object",
      "meaning": "Per-result source observations and the note’s common scope. A note can select values that overlap the catalog; repeated citations are not new experiments."
    },
    {
      "collection": "articles",
      "path": "articles[].deployment_profiles",
      "type": "optional array",
      "meaning": "Detailed reported launches: artifact revision, weight format, image tag, allocated KV, speculation, context cap and measured users. An image tag does not pin an immutable image digest."
    },
    {
      "collection": "articles",
      "path": "articles[].comparison_groups",
      "type": "optional array",
      "meaning": "Derived summaries reference their raw source point IDs. Keep derivation and interval_kind; A06 min/max is the observed range across three cells, not a confidence interval."
    },
    {
      "collection": "articles",
      "path": "articles[].generation / image",
      "type": "object / object",
      "meaning": "Writing and illustration provenance. AI-assisted writing and a conceptual image do not create independent model evidence."
    },
    {
      "collection": "articles",
      "path": "articles[].runtime_options[]",
      "type": "object[] (optional)",
      "meaning": "Documentation-informed runtime options: keep maintainer capability, editorial shortlist, operator checks and source identity together. tested_engine_version=null means no installed version was tested here."
    },
    {
      "collection": "articles",
      "path": "articles[].case_reports[]",
      "type": "object[] (optional)",
      "meaning": "Historical incident records with evidence_status, per-case conditions, observations, remediation state, follow_up and unresolved fields. A source_warning_unperformed is not a completed test; a proposed change is not a verified fix."
    },
    {
      "collection": "articles",
      "path": "articles[].quality_comparisons[]",
      "type": "object[] (optional)",
      "meaning": "Paired task-score differences with sample counts, reported confidence intervals and distinct quality gates. Preserve custom passes=true together with non_inferior=false; no equivalence conclusion. New community catalog records can carry the same optional object."
    },
    {
      "collection": "articles",
      "path": "articles[].memory_observations[]",
      "type": "object[] (optional)",
      "meaning": "Optional source-backed memory records. Preserve memory kind and original units: file bytes, host RSS including reclaimable mmap pages, anonymous residence, minimum MemAvailable and KV-token allocation are different quantities. A configured floor or allocation is not measured peak usage.",
      "example": {
        "kind": "file_inventory_not_RAM",
        "actual_peak_RAM": null
      },
      "missing_semantics": "Absent means this optional record was not curated; null remains unreported, never zero."
    },
    {
      "collection": "articles",
      "path": "articles[].long_context_cases[]",
      "type": "object[] (optional)",
      "meaning": "Optional narrow case records with actual input/output tokens, context limit, exact task and success criterion, cache hits, runtime identity and unknown sampling. One exact-string success at one depth does not establish general reasoning, long-response or concurrency quality.",
      "example": {
        "classification": "source_reported_single_needle_success",
        "actual_output_tokens": 7,
        "sampling": null
      },
      "missing_semantics": "Absent means no such curated case; null conditions remain unknown."
    },
    {
      "collection": "catalog",
      "path": "records[].editorial.reported_results[].conditions.repetitions",
      "type": "integer | string | null (optional)",
      "meaning": "Preserve the original summary unit: a C1 median of three trials, a single higher-concurrency trial, three runs in one boot, or four boots with three cold prompts are different. SD does not imply normality; a boot-level SD is not a confidence interval.",
      "example": "mean of 4 boots x 3 cold prompts, sd between boots",
      "missing_semantics": "Null is unreported repetition count; do not infer one or treat selected result points as trials."
    },
    {
      "collection": "catalog",
      "path": "records[].editorial.reported_results[].conditions.ngram_speculation",
      "type": "boolean (optional)",
      "meaning": "Ngram speculative decoding is separate from the model n-gram embedding table. False indicates speculation off for this gate; a NOT_RUN_BY_GATE follow-up is not a measured zero benefit. Preserve ngram_embedding_table_present separately.",
      "example": false,
      "missing_semantics": "Absent or null means that feature was not specified; it never proves the embedding table is disabled."
    },
    {
      "collection": "articles",
      "path": "articles[].case_reports[].memory",
      "type": "object (optional)",
      "meaning": "A peak summed over a full GPU campaign is not a per-cell peak. Retain memory kind, unit, sample count, time window and whether memory is paired with the same workload. Free VRAM, used VRAM, host RSS and KV token occupancy cannot be merged.",
      "example": {
        "kind": "sampled_peak_summed_GPU_used_over_entire_matrix",
        "paired_per_cell_peak": null
      },
      "missing_semantics": "Null per-cell peak and host peak remain unknown; do not infer them from a campaign maximum."
    },
    {
      "collection": "articles",
      "path": "articles[].quality_comparisons[]",
      "type": "object[] (optional)",
      "meaning": "Keep benchmark subset size, first-pass scores, timeouts, retry selection and composite scores together. Store false gates and skipped checks. Unequal sampling or output caps prevent a quantization equivalence or global ranking claim.",
      "example": {
        "initial_local_correct": 14,
        "initial_timeouts": 4,
        "retry_correct": 4,
        "final_local_correct": 18,
        "subset_size": 20
      },
      "missing_semantics": "An absent score is unreported. A skipped check is not a pass or a measured algorithm failure."
    },
    {
      "collection": "articles",
      "path": "articles[].case_reports[].ship_tree",
      "type": "object (optional)",
      "meaning": "Preserve invalid initial runs, their harness errors and zero injections separately from a valid rerun. A development-branch recovery battery with host KV off is not the ship-tree sample.",
      "example": {
        "initial_valid": false,
        "initial_injects": 0,
        "rerun_injections": 2,
        "rerun_recoveries": 2
      },
      "missing_semantics": "Unknown rerun scope stays null. An invalid run is not a zero model-accuracy score."
    },
    {
      "collection": "articles",
      "path": "articles[].quality_comparisons[].old_task_scores_withdrawn",
      "type": "boolean (optional)",
      "meaning": "A source correction may withdraw old task scores while publishing a new probability-health check. Retain build, reference precision, panel, source scope and the missing task rerun. Do not transfer scores between NVFP4 and EXL3.",
      "example": {
        "old_task_scores_withdrawn": true,
        "current_task_accuracy": null,
        "production_KL": 0.00828,
        "production_PPL": null
      },
      "missing_semantics": "True means the source withdrew that line's earlier task results; null current accuracy is not a score of zero and KL is not accuracy."
    }
  ],
  "query_examples": [
    {
      "id": "exact-1017",
      "title": "Find the exact 1017 deployment artifact",
      "needs": [
        "catalog"
      ],
      "purpose": "Discovery IDs only; read the complete record before evaluating compatibility.",
      "expression": "catalog.records.filter(r => r.kind === 'deployment' && r.model_artifact?.id === 'deepseek-ai/DeepSeek-V4.1-Flash-1017').map(r => r.id)"
    },
    {
      "id": "spark-sglang",
      "title": "Find Spark reports tagged with SGLang",
      "needs": [
        "catalog"
      ],
      "purpose": "Source discovery; the tags do not certify support for every listed model.",
      "expression": "catalog.records.filter(r => r.kind === 'benchmark' && r.hardware_topics.includes('spark') && r.runtime_topics.includes('SGLang')).map(r => r.id)"
    },
    {
      "id": "mac-ttft",
      "title": "Keep Mac TTFT points with their full conditions",
      "needs": [
        "catalog"
      ],
      "purpose": "Preserve each original point, including uncertainty and unknown output length.",
      "expression": "catalog.records.filter(r => r.id === 'qwen35-m4pro-20260306').flatMap(r => r.editorial.reported_results.filter(p => p.metric === 'ttft').map(p => ({record_id:r.id, source:r.source, ...p})))"
    },
    {
      "id": "mac-bf16-memory",
      "title": "Read the two bf16 memory selections in A04",
      "needs": [
        "articles"
      ],
      "purpose": "Source-reported selections, potentially overlapping catalog points; not additional experiments.",
      "expression": "articles.articles.filter(a => a.id === 'A04').flatMap(a => a.reported_results.filter(p => p.metric === 'peak_memory' && p.conditions.weight_precision === 'bf16').map(p => ({article_id:a.id, sources:a.sources, ...p})))"
    },
    {
      "id": "qwen-spark-profile",
      "title": "Return the exact Qwen launch profile and source documents",
      "needs": [
        "articles"
      ],
      "purpose": "Artifact and configuration lookup; missing executed versions and image digests remain null.",
      "expression": "articles.articles.filter(a => a.id === 'A05').map(a => ({article_id:a.id, conditions:a.conditions, sources:a.sources, supplemental_sources:a.supplemental_sources, deployment_profiles:a.deployment_profiles.filter(p => p.model_artifact === 'RadixArk/Qwen3.8-Flash-Next-NVFP4')}))"
    },
    {
      "id": "kv-pool-tradeoff",
      "title": "Read the A06 raw cells together with their derived summaries",
      "needs": [
        "articles"
      ],
      "purpose": "Keep the model, conditions, raw point references and observed-range interpretation together.",
      "expression": "articles.articles.filter(a => a.id === 'A06').map(a => ({article_id:a.id, conditions:a.conditions, sources:a.sources, supplemental_sources:a.supplemental_sources, reported_results:a.reported_results, comparison_groups:a.comparison_groups}))"
    },
    {
      "id": "deepseek-1017-community",
      "title": "Read the exact 1017 TensorFold community report",
      "needs": [
        "catalog"
      ],
      "purpose": "Keep the converted artifact, separate Engram tables, engine configuration and per-result cohort.",
      "expression": "catalog.records.filter(r => r.id === 'deepseek-1017-tensorfold-spark')"
    },
    {
      "id": "qwen38-chisel",
      "title": "Read Qwen3.8 Chisel timing and negative quality gates",
      "needs": [
        "catalog"
      ],
      "purpose": "Return the complete source report; preserve thinking modes, paired intervals and false non-inferiority flags.",
      "expression": "catalog.records.filter(r => r.id === 'qwen38-chisel-dflash-replay')"
    },
    {
      "id": "qwen-flash-next-memory",
      "title": "Read Qwen Flash-Next memory types and cold-context conditions",
      "needs": [
        "catalog"
      ],
      "purpose": "Return both complete reports; keep Spark and RTX PRO hardware, mmap/anonymous memory and cold-prefix cohorts separate.",
      "expression": "catalog.records.filter(r => ['qwen38-flash-next-single-spark-context','qwen38-flash-next-ple-memory'].includes(r.id))"
    },
    {
      "id": "deepseek-1017-long-context",
      "title": "Read near-1M 1017 needle cases with output and headroom",
      "needs": [
        "catalog"
      ],
      "purpose": "Retain both saved prompts, seven-token completions, same-depth retrieval scope, unknown sampling and minimum MemAvailable.",
      "expression": "catalog.records.filter(r => r.id === 'deepseek-1017-million-token-needle')"
    },
    {
      "id": "deepseek-1017-quantization",
      "title": "Separate 1017 EXL3 runtime evidence from VQ artifact inspection",
      "needs": [
        "catalog"
      ],
      "purpose": "Return complete identities and conditions; VQ serving speed and runtime RAM remain unmeasured, not borrowed from older V4.1.",
      "expression": "catalog.records.filter(r => ['deepseek-1017-dual-spark-quick','deepseek-1017-vq-artifact-inspection'].includes(r.id))"
    },
    {
      "id": "qwen-flash-next-spark-concurrency",
      "title": "Read both Spark concurrency and cold-prefix reports",
      "needs": [
        "catalog"
      ],
      "purpose": "Keep distinct hardware, checkpoint releases, exact request mix, trial units, SD and configured capacity together.",
      "expression": "catalog.records.filter(r => ['qwen38-flash-next-sglang-concurrency','qwen38-flash-next-tp2-concurrency-context'].includes(r.id))"
    },
    {
      "id": "qwen-flash-next-5090-gates",
      "title": "Read actual 5090 context gates and unpaired memory telemetry",
      "needs": [
        "catalog"
      ],
      "purpose": "Retain the 104805-token retrieval, 32K-only peaks, unknown sampling scope and separate ngram speculation/embedding states.",
      "expression": "catalog.records.filter(r => r.id === 'qwen38-flash-next-5090-offload-gates')"
    },
    {
      "id": "deepseek-1017-vq-followup",
      "title": "Read the dated VQ follow-up with its missing runtime results",
      "needs": [
        "articles"
      ],
      "purpose": "Return the complete A12 interpretation including the October 18 bounded source recheck. The absent 1017 VQ serving measurements remain unknown.",
      "expression": "articles.articles.filter(a => a.id === 'A12')"
    },
    {
      "id": "qwen-long-concurrency-failures",
      "title": "Read Qwen long concurrent batches, memory scopes and failures",
      "needs": [
        "articles"
      ],
      "purpose": "Return A15 complete, preserving original source identities, concurrent batch conditions, campaign-only peak and open-PR failure scope.",
      "expression": "articles.articles.filter(a => a.id === 'A15')"
    },
    {
      "id": "spark-deployment-score-audit",
      "title": "Read GLM, MiniMax and Kimi deployments with score limitations",
      "needs": [
        "articles"
      ],
      "purpose": "Keep complete versions, raw medians, false quality gates, July/August dates, missing peak RAM, subset sizes and timeout retries.",
      "expression": "articles.articles.filter(a => ['A16','A17','A18'].includes(a.id))"
    },
    {
      "id": "qwen-paired-memory-recovery",
      "title": "Read paired Qwen memory and actual recovery reruns",
      "needs": [
        "articles"
      ],
      "purpose": "Keep the unchanged concurrent peak, different 500K peak, production/validation hardware, invalid first ship run and two-event rerun.",
      "expression": "articles.articles.filter(a => a.id === 'A19')"
    },
    {
      "id": "new-spark-score-retractions",
      "title": "Read Spark deployments, score retractions and build-specific retries",
      "needs": [
        "articles"
      ],
      "purpose": "Return full GLM and MiMo guides, preserving historical scores, EXL3 withdrawal, missing current task rerun, exact versions and C1/C8 denominators.",
      "expression": "articles.articles.filter(a => ['A20','A21','A22'].includes(a.id))"
    }
  ],
  "example_download": "/downloads/evidence-query-examples.json",
  "executable_download": "/downloads/query-evidence.mjs",
  "generation": {
    "text_draft": "local Spark Qwen",
    "review": "Run AI Here with Codex",
    "code": "Codex; executed against the actual candidate JSON; static read/filter only"
  },
  "body_markdown": "# Query Deployment Evidence with Its Conditions\n\nFetch a static collection, filter it locally, and carry the original source and test conditions into the answer. The examples below read public JSON only. They do not launch models or establish compatibility.\n\n## Choose the right collection\n\n| Resource | What it contains | How to use it |\n|---|---|---|\n| [catalog.json](/catalog.json) | 59 original references; catalog schema 3.1.0 | Discover sources by stable ID, source role and topics. |\n| [articles.json](/articles.json) | Reviewed explanatory notes; envelope 1.0.0 | Read the same body, selected evidence and interpretation as the articles. |\n| [Catalog JSON Schema](/schema/community-catalog-3.1.0.json) | Machine-readable catalog contract | Validate the catalog before extracting fields. |\n| [This field guide as JSON](/for-ai/index.json) | Field meanings and query definitions | Keep the collection paths and evidence rules with your integration. |\n| [catalog.md](/catalog.md), [articles.md](/articles.md) | Readable versions of the same material | Use Markdown when a full JSON extraction is unnecessary. |\n\nIndividual source records use `/community/<id>/index.json` and `index.md`; explanatory notes use their `/guides/<slug>/index.json` and `index.md`. [llms.txt](/llms.txt) lists entry points, and [changes.json](/changes.json) describes the source dataset revision. A06 and other notes may select an existing author’s values again: note counts, source counts and experiment counts are separate.\n\n## A small reading workflow\n\n1. **Fetch:** get the required collection once and check its contract version.\n2. **Filter:** choose a stable ID or exact artifact, then inspect the complete record.\n3. **Carry evidence:** retain source identity, metric definition, units, uncertainty and per-result conditions with every extracted value.\n\n`/catalog.json?model=Qwen` still returns the complete static file. Filters belong in your client. No authentication, database, WordPress or query API is required.\n\n## Read the fields without changing their meaning\n\n### Original-source catalog\n\n| Field | Type | Meaning |\n|---|---|---|\n| `schema_version / dataset_version` | string / string | Contract version and editorial dataset revision are separate. Catalog schema is 3.1.0; a new explainer does not silently change the source dataset. |\n| `source_count / records` | integer / array | Number of original reference records and the complete collection. This collection has 59 entries. |\n| `records[].id / kind` | string / model  /  deployment  /  benchmark | Stable reference identity and source role. A benchmark role does not mean the site ran it. |\n| `records[].model_family / runtime_topics / hardware_topics` | string / string[] / string[] | Discovery facets. Topic membership is not proof that a particular artifact works on a device. |\n| `records[].source.publisher / pinned_url` | string / URL | Original publisher and revision-pinned document. Preserve attribution when extracting a result. |\n| `records[].source.revision / document_sha256` | string / SHA-256 | Document-repository revision and digest of the document bytes. Neither is a weight hash or proof of factual correctness. |\n| `records[].source.checked_on / last_commit_at` | date / timestamp or null | Editorial check date and source commit timestamp. Neither substitutes for the experiment date. |\n| `records[].model_artifact` | object or null | The exact model/weight artifact when stated. Its id and revision identify weights, not an internal content-generation model. |\n| `records[].conditions.runtime_execution_revision` | string or null | Executed inference-engine revision, only when the source supplies it. The report repository revision is a different identity. |\n| `records[].reported_setup / conditions / test_profiles` | object / object / array | Report-wide scope, shared declarations and distinct test shapes. Do not merge profiles into an invented common run. |\n| `records[].recipe_parameters / deployment_options` | object or null / array | Recipe declarations and selectable configurations. Planning memory and context settings are not measured peak memory or achieved context. |\n| `records[].editorial.reported_results[]` | object[] | Limited attributed source results. Carry each complete point and its own conditions instead of just a value. |\n| `...reported_results[].metric / value / unit` | string / number / string | The reported quantity, numeric value and unit. Served rate and decode rate have different timing denominators; GB and GiB differ. |\n| `...reported_results[].statistic / uncertainty` | string / object or null | Reported statistic and supplied dispersion. SD, a spread ratio and an observed range are different quantities. Null is not zero uncertainty. |\n| `...reported_results[].conditions` | object | The result-specific model, hardware, input/output shape, concurrency, cache, speculation and measurement definition. Retain missing fields. |\n| `...reported_results[].evidence_url / source_document_sha256` | URL / SHA-256 | Exact original location and the matching document digest. Use the digest of that document, including supplemental documents where applicable. |\n| `records[].evidence_state / site_reproduced / metrics` | source_reported / false / null | Current evidence is attributed to sources. The legacy site-measurement metrics field is empty; this does not say the original author measured zero. |\n| `records[].comparison_status / performance_ranking_allowed` | string / false | Discovery is not a cross-condition ranking. Unknown identity or mismatched workloads prevent a direct performance verdict. |\n| `records[].license` | object | Recorded repository terms with their stated scope. Model, conversion, dataset and container terms need their own reading. |\n\n### Explanatory articles\n\n| Field | Type | Meaning |\n|---|---|---|\n| `schema_version / source_count / articles` | 1.0.0 / integer / array | Separate explainer envelope. Its legacy source_count counts article entries, not unique original reports; do not add it to catalog source_count. |\n| `articles[].id / path / body_markdown` | string / path / Markdown | Stable note identity, reading URL and the same reviewed body used by HTML and Markdown exports. |\n| `articles[].sources / supplemental_sources` | array / optional array | Primary catalog references and distinct supporting documents. Each supporting document has its own document_id, revision, digest and parent_record_id. |\n| `articles[].facts / declared_variants` | array / optional array | Source declarations or recipe selections with their classifications. A declared planning minimum is not a measured run. |\n| `articles[].reported_results / conditions` | array / object | Per-result source observations and the note’s common scope. A note can select values that overlap the catalog; repeated citations are not new experiments. |\n| `articles[].deployment_profiles` | optional array | Detailed reported launches: artifact revision, weight format, image tag, allocated KV, speculation, context cap and measured users. An image tag does not pin an immutable image digest. |\n| `articles[].comparison_groups` | optional array | Derived summaries reference their raw source point IDs. Keep derivation and interval_kind; A06 min/max is the observed range across three cells, not a confidence interval. |\n| `articles[].generation / image` | object / object | Writing and illustration provenance. AI-assisted writing and a conceptual image do not create independent model evidence. |\n\n### Unknown, false, zero and absent\n\n`null` means unknown in this extraction. `false` is a known boolean and `0` is a known number. An absent optional property is not a measured zero or a negative finding. Avoid `value || fallback` when preserving zero or false; check property presence and null explicitly. An image digest of null does not become known because an image tag is present.\n\n### Preserve the measurement boundary\n\nA report can contain several models or workloads. Keep result-specific conditions and the selected test profile alongside broader setup notes; do not fill an unknown from an unrelated row. An input length differs from a configured context limit, an allocated KV pool differs from per-request context, and aggregate output tokens differ from a per-request budget.\n\nRecipe declarations, publisher claims, raw campaign cells and computed summaries remain separate. An observed minimum/maximum across a few campaigns is not a confidence interval. Preserve `classification`, `statistic`, `uncertainty`, `derivation` and `interval_kind` when present. The original authors supply the evidence; `site_reproduced: false` remains attached to it.\n\n## Run the JSON query examples\n\nDownload the [tested query definitions and real expected outputs](/downloads/evidence-query-examples.json), or the [small executable JavaScript examples](/downloads/query-evidence.mjs). The module needs Node.js with built-in `fetch`. It fetches each required file once, filters in memory and prints JSON.\n\n```sh\nnode query-evidence.mjs exact-1017\nnode query-evidence.mjs kv-pool-tradeoff\n```\n\nTo use the expressions in a browser console on this site, load the needed inputs first. These requests read the same public files linked above:\n\n```js\nconst catalog = await fetch('/catalog.json').then(r => {\n  if (!r.ok) throw new Error(`HTTP ${r.status}`);\n  return r.json();\n});\nconst articles = await fetch('/articles.json').then(r => {\n  if (!r.ok) throw new Error(`HTTP ${r.status}`);\n  return r.json();\n});\n```\n\n### Find the exact 1017 deployment artifact\n\nDiscovery IDs only; read the complete record before evaluating compatibility.\n\n```js\ncatalog.records.filter(r => r.kind === 'deployment' && r.model_artifact?.id === 'deepseek-ai/DeepSeek-V4.1-Flash-1017').map(r => r.id)\n```\n\n### Find Spark reports tagged with SGLang\n\nSource discovery; the tags do not certify support for every listed model.\n\n```js\ncatalog.records.filter(r => r.kind === 'benchmark' && r.hardware_topics.includes('spark') && r.runtime_topics.includes('SGLang')).map(r => r.id)\n```\n\n### Keep Mac TTFT points with their full conditions\n\nPreserve each original point, including uncertainty and unknown output length.\n\n```js\ncatalog.records.filter(r => r.id === 'qwen35-m4pro-20260306').flatMap(r => r.editorial.reported_results.filter(p => p.metric === 'ttft').map(p => ({record_id:r.id, source:r.source, ...p})))\n```\n\n### Read the two bf16 memory selections in A04\n\nSource-reported selections, potentially overlapping catalog points; not additional experiments.\n\n```js\narticles.articles.filter(a => a.id === 'A04').flatMap(a => a.reported_results.filter(p => p.metric === 'peak_memory' && p.conditions.weight_precision === 'bf16').map(p => ({article_id:a.id, sources:a.sources, ...p})))\n```\n\n### Return the exact Qwen launch profile and source documents\n\nArtifact and configuration lookup; missing executed versions and image digests remain null.\n\n```js\narticles.articles.filter(a => a.id === 'A05').map(a => ({article_id:a.id, conditions:a.conditions, sources:a.sources, supplemental_sources:a.supplemental_sources, deployment_profiles:a.deployment_profiles.filter(p => p.model_artifact === 'RadixArk/Qwen3.8-Flash-Next-NVFP4')}))\n```\n\n### Read the A06 raw cells together with their derived summaries\n\nKeep the model, conditions, raw point references and observed-range interpretation together.\n\n```js\narticles.articles.filter(a => a.id === 'A06').map(a => ({article_id:a.id, conditions:a.conditions, sources:a.sources, supplemental_sources:a.supplemental_sources, reported_results:a.reported_results, comparison_groups:a.comparison_groups}))\n```\n\n### Read the exact 1017 TensorFold community report\n\nKeep the converted artifact, separate Engram tables, engine configuration and per-result cohort.\n\n```js\ncatalog.records.filter(r => r.id === 'deepseek-1017-tensorfold-spark')\n```\n\n### Read Qwen3.8 Chisel timing and negative quality gates\n\nReturn the complete source report; preserve thinking modes, paired intervals and false non-inferiority flags.\n\n```js\ncatalog.records.filter(r => r.id === 'qwen38-chisel-dflash-replay')\n```\n\n### Read Qwen Flash-Next memory types and cold-context conditions\n\nReturn both complete reports; keep Spark and RTX PRO hardware, mmap/anonymous memory and cold-prefix cohorts separate.\n\n```js\ncatalog.records.filter(r => ['qwen38-flash-next-single-spark-context','qwen38-flash-next-ple-memory'].includes(r.id))\n```\n\n### Read near-1M 1017 needle cases with output and headroom\n\nRetain both saved prompts, seven-token completions, same-depth retrieval scope, unknown sampling and minimum MemAvailable.\n\n```js\ncatalog.records.filter(r => r.id === 'deepseek-1017-million-token-needle')\n```\n\n### Separate 1017 EXL3 runtime evidence from VQ artifact inspection\n\nReturn complete identities and conditions; VQ serving speed and runtime RAM remain unmeasured, not borrowed from older V4.1.\n\n```js\ncatalog.records.filter(r => ['deepseek-1017-dual-spark-quick','deepseek-1017-vq-artifact-inspection'].includes(r.id))\n```\n\n## Read runtime options and historical incidents\n\nThe [runtime choice guide](/guides/inference-runtime-choice/) and [GB10 incident guide](/guides/gb10-long-context-failures/) add two optional article fields. They do not add independent performance experiments.\n\n| Field | Type | Meaning |\n|---|---|---|\n| `articles[].runtime_options[]` | object[] | Keep documented capability, editorial shortlist, operator checks and source pointers together. `tested_engine_version: null` means no installed version was tested here. |\n| `articles[].case_reports[]` | object[] | Read `evidence_status`, per-case `conditions`, observations, remediation state, any follow-up and unresolved fields together. `source_warning_unperformed` is a warning with no completed test in that snapshot. A proposed change is not a verified fix. |\n\nThe new Qwen source and article also retain `quality_comparisons`: paired sample counts, score differences, reported intervals and separate `passes` / `non_inferior` flags. A custom bar can pass while non-inferiority remains unproved. Return the whole comparison rather than reducing it to one boolean.\n\nFor incidents, the article's overall `evidence_state: source_reported` identifies who supplied the evidence. The nested status explains what kind of evidence it is. A source-reproduced failure remains `site_reproduced: false` here. Document revision, executed engine version and image digest remain separate identities, with unknown values preserved.\n\n## Cite and verify what you reuse\n\nLink to the pinned original location, name its publisher and preserve the matching document revision and digest. A hash confirms which bytes were read; it does not independently validate a claim. Supplemental sources have their own identities and hashes. Keep the experiment date separate from the source check date.\n\nThe current collections contain no independently reproduced site model runs and no cross-source leaderboard. Repository terms are recorded with their scope; copying a link or finding a source through a topic does not grant additional artifact rights. Use the [collection method](/benchmarks/methodology/) and the original report when a condition is missing or contradictory.\n\nUpdated 18 October 2026. Local Spark Qwen drafted the explanations; Run AI Here reviewed field meanings and executed the query examples against the actual published data shape. The JSON and Markdown versions preserve the same guidance.\n\n## Memory and long-context observations\n\nKeep `memory_observations` and `long_context_cases` beside their source report. File bytes, reclaimable mapped RSS, anonymous memory, minimum host MemAvailable and allocated KV tokens describe different quantities. Configured context and request slots are limits; actual input/output tokens and task scope describe the performed case. A seven-token exact-needle completion is not broad long-context quality. Unknown image digests, sampling and runtime RAM remain null.\n\n### Read both Spark concurrency and cold-prefix reports\n\nKeep distinct hardware, checkpoint releases, exact request mix, trial units, SD and configured capacity together.\n\n```js\ncatalog.records.filter(r => ['qwen38-flash-next-sglang-concurrency','qwen38-flash-next-tp2-concurrency-context'].includes(r.id))\n```\n\n### Read actual 5090 context gates and unpaired memory telemetry\n\nRetain the 104805-token retrieval, 32K-only peaks, unknown sampling scope and separate ngram speculation/embedding states.\n\n```js\ncatalog.records.filter(r => r.id === 'qwen38-flash-next-5090-offload-gates')\n```\n\n### Read the dated VQ follow-up with its missing runtime results\n\nReturn the complete A12 interpretation including the October 18 bounded source recheck. The absent 1017 VQ serving measurements remain unknown.\n\n```js\narticles.articles.filter(a => a.id === 'A12')\n```\n\n## Concurrency, repetition units and model components\n\nKeep aggregate output, per-request decode, TTFT and time per output token as separate metrics. A scheduler ceiling is not a completed full-window load. Preserve the original repetition unit and the SD definition. Memory telemetry must remain paired with its actual gate; an unspecified VM/host sampling boundary remains unknown. `ngram_speculation=false` does not disable the model n-gram embedding table. Discourse post version 3 is a mutable public report with a saved dated hash, not an immutable Git commit. The A12 VQ recheck covers named sources only and does not invent a global absence of community tests.\n\n### Read Qwen long concurrent batches, memory scopes and failures\n\nReturn A15 complete, preserving original source identities, concurrent batch conditions, campaign-only peak and open-PR failure scope.\n\n```js\narticles.articles.filter(a => a.id === 'A15')\n```\n\n### Read GLM, MiniMax and Kimi deployments with score limitations\n\nKeep complete versions, raw medians, false quality gates, July/August dates, missing peak RAM, subset sizes and timeout retries.\n\n```js\narticles.articles.filter(a => ['A16','A17','A18'].includes(a.id))\n```\n\n## Memory scope, quality scores and retries\n\nA full-campaign summed GPU peak does not give a peak for each context/concurrency cell. Preserve per-cell peak and host peak as null when absent. A post-graph free-memory observation is not a measured peak used-memory observation. Submitted requests may queue; retain observed running concurrency separately. Quality summaries must retain subset size, initial results, timeouts and any selected retries. The Kimi source has 14/20 initial passes, four timeout-only retry passes and an 18/20 composite. The GLM source has `quality_passed=false` in every arm. No full-suite or hardware-independent score is assigned. A mutable website or GitHub comment has a dated saved SHA256; its snapshot label is not a Git commit.\n\n### Read paired Qwen memory and actual recovery reruns\n\nKeep the unchanged concurrent peak, different 500K peak, production/validation hardware, invalid first ship run and two-event rerun.\n\n```js\narticles.articles.filter(a => a.id === 'A19')\n```\n\n### Read Spark deployments, score retractions and build-specific retries\n\nReturn full GLM and MiMo guides, preserving historical scores, EXL3 withdrawal, missing current task rerun, exact versions and C1/C8 denominators.\n\n```js\narticles.articles.filter(a => ['A20','A21','A22'].includes(a.id))\n```\n\n## Paired memory, withdrawn scores and invalid attempts\n\nMatch hardware edition, memory kind, workload and run window before comparing peaks. Qwen's concurrent host-KV gather A/B is +1248 MiB over idle in both arms, while the separate 500K single-request device-used peak falls. Invalid recovery attempts with zero injections remain visible beside the smaller valid ship-tree rerun. An EXL3 task-score retraction is scoped to that source line: the separate historical NVFP4 panel is not a current corrected-EXL3 score. Current task accuracy remains null. MiMo's long extraction is a long-output fixture; its C1 log reports1498 input tokens, and the C8 input count is absent. Retain the original build for every needle or score.\n",
  "collections": {
    "catalog": {
      "url": "/catalog.json",
      "sha256": "169d3662ce74cf538adf13e375682dd472a2659dcb1e6ca61428dcf0d7e85449"
    },
    "articles": {
      "url": "/articles.json",
      "sha256": "c0a50d7242e4a653f13cc48e0576be22d41a7465845a97787c7fd80fdff2895d"
    }
  },
  "query_verification": {
    "state": "executed_against_same_build",
    "queries": 18,
    "model_executions": 0,
    "output_counts": {
      "exact-1017": 1,
      "spark-sglang": 4,
      "mac-ttft": 4,
      "mac-bf16-memory": 2,
      "qwen-spark-profile": 1,
      "kv-pool-tradeoff": 1,
      "deepseek-1017-community": 1,
      "qwen38-chisel": 1,
      "qwen-flash-next-memory": 2,
      "deepseek-1017-long-context": 1,
      "deepseek-1017-quantization": 2,
      "qwen-flash-next-spark-concurrency": 2,
      "qwen-flash-next-5090-gates": 1,
      "deepseek-1017-vq-followup": 1,
      "qwen-long-concurrency-failures": 1,
      "spark-deployment-score-audit": 3,
      "qwen-paired-memory-recovery": 1,
      "new-spark-score-retractions": 3
    }
  }
}