# Query Deployment Evidence with Its Conditions

Fetch 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.

## Choose the right collection

| Resource | What it contains | How to use it |
|---|---|---|
| [catalog.json](/catalog.json) | 59 original references; catalog schema 3.1.0 | Discover sources by stable ID, source role and topics. |
| [articles.json](/articles.json) | Reviewed explanatory notes; envelope 1.0.0 | Read the same body, selected evidence and interpretation as the articles. |
| [Catalog JSON Schema](/schema/community-catalog-3.1.0.json) | Machine-readable catalog contract | Validate the catalog before extracting fields. |
| [This field guide as JSON](/for-ai/index.json) | Field meanings and query definitions | Keep the collection paths and evidence rules with your integration. |
| [catalog.md](/catalog.md), [articles.md](/articles.md) | Readable versions of the same material | Use Markdown when a full JSON extraction is unnecessary. |

Individual 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.

## A small reading workflow

1. **Fetch:** get the required collection once and check its contract version.
2. **Filter:** choose a stable ID or exact artifact, then inspect the complete record.
3. **Carry evidence:** retain source identity, metric definition, units, uncertainty and per-result conditions with every extracted value.

`/catalog.json?model=Qwen` still returns the complete static file. Filters belong in your client. No authentication, database, WordPress or query API is required.

## Read the fields without changing their meaning

### Original-source catalog

| Field | Type | Meaning |
|---|---|---|
| `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. |
| `source_count / records` | integer / array | Number of original reference records and the complete collection. This collection has 59 entries. |
| `records[].id / kind` | string / model  /  deployment  /  benchmark | Stable reference identity and source role. A benchmark role does not mean the site ran it. |
| `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. |
| `records[].source.publisher / pinned_url` | string / URL | Original publisher and revision-pinned document. Preserve attribution when extracting a result. |
| `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. |
| `records[].source.checked_on / last_commit_at` | date / timestamp or null | Editorial check date and source commit timestamp. Neither substitutes for the experiment date. |
| `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. |
| `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. |
| `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. |
| `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. |
| `records[].editorial.reported_results[]` | object[] | Limited attributed source results. Carry each complete point and its own conditions instead of just a value. |
| `...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. |
| `...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. |
| `...reported_results[].conditions` | object | The result-specific model, hardware, input/output shape, concurrency, cache, speculation and measurement definition. Retain missing fields. |
| `...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. |
| `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. |
| `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. |
| `records[].license` | object | Recorded repository terms with their stated scope. Model, conversion, dataset and container terms need their own reading. |

### Explanatory articles

| Field | Type | Meaning |
|---|---|---|
| `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. |
| `articles[].id / path / body_markdown` | string / path / Markdown | Stable note identity, reading URL and the same reviewed body used by HTML and Markdown exports. |
| `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. |
| `articles[].facts / declared_variants` | array / optional array | Source declarations or recipe selections with their classifications. A declared planning minimum is not a measured run. |
| `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. |
| `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. |
| `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. |
| `articles[].generation / image` | object / object | Writing and illustration provenance. AI-assisted writing and a conceptual image do not create independent model evidence. |

### Unknown, false, zero and absent

`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.

### Preserve the measurement boundary

A 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.

Recipe 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.

## Run the JSON query examples

Download 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.

```sh
node query-evidence.mjs exact-1017
node query-evidence.mjs kv-pool-tradeoff
```

To use the expressions in a browser console on this site, load the needed inputs first. These requests read the same public files linked above:

```js
const catalog = await fetch('/catalog.json').then(r => {
  if (!r.ok) throw new Error(`HTTP ${r.status}`);
  return r.json();
});
const articles = await fetch('/articles.json').then(r => {
  if (!r.ok) throw new Error(`HTTP ${r.status}`);
  return r.json();
});
```

### Find the exact 1017 deployment artifact

Discovery IDs only; read the complete record before evaluating compatibility.

```js
catalog.records.filter(r => r.kind === 'deployment' && r.model_artifact?.id === 'deepseek-ai/DeepSeek-V4.1-Flash-1017').map(r => r.id)
```

### Find Spark reports tagged with SGLang

Source discovery; the tags do not certify support for every listed model.

```js
catalog.records.filter(r => r.kind === 'benchmark' && r.hardware_topics.includes('spark') && r.runtime_topics.includes('SGLang')).map(r => r.id)
```

### Keep Mac TTFT points with their full conditions

Preserve each original point, including uncertainty and unknown output length.

```js
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})))
```

### Read the two bf16 memory selections in A04

Source-reported selections, potentially overlapping catalog points; not additional experiments.

```js
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})))
```

### Return the exact Qwen launch profile and source documents

Artifact and configuration lookup; missing executed versions and image digests remain null.

```js
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')}))
```

### Read the A06 raw cells together with their derived summaries

Keep the model, conditions, raw point references and observed-range interpretation together.

```js
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}))
```

### Read the exact 1017 TensorFold community report

Keep the converted artifact, separate Engram tables, engine configuration and per-result cohort.

```js
catalog.records.filter(r => r.id === 'deepseek-1017-tensorfold-spark')
```

### Read Qwen3.8 Chisel timing and negative quality gates

Return the complete source report; preserve thinking modes, paired intervals and false non-inferiority flags.

```js
catalog.records.filter(r => r.id === 'qwen38-chisel-dflash-replay')
```

### Read Qwen Flash-Next memory types and cold-context conditions

Return both complete reports; keep Spark and RTX PRO hardware, mmap/anonymous memory and cold-prefix cohorts separate.

```js
catalog.records.filter(r => ['qwen38-flash-next-single-spark-context','qwen38-flash-next-ple-memory'].includes(r.id))
```

### Read near-1M 1017 needle cases with output and headroom

Retain both saved prompts, seven-token completions, same-depth retrieval scope, unknown sampling and minimum MemAvailable.

```js
catalog.records.filter(r => r.id === 'deepseek-1017-million-token-needle')
```

### Separate 1017 EXL3 runtime evidence from VQ artifact inspection

Return complete identities and conditions; VQ serving speed and runtime RAM remain unmeasured, not borrowed from older V4.1.

```js
catalog.records.filter(r => ['deepseek-1017-dual-spark-quick','deepseek-1017-vq-artifact-inspection'].includes(r.id))
```

## Read runtime options and historical incidents

The [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.

| Field | Type | Meaning |
|---|---|---|
| `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. |
| `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. |

The 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.

For 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.

## Cite and verify what you reuse

Link 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.

The 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.

Updated 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.

## Memory and long-context observations

Keep `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.

### Read both Spark concurrency and cold-prefix reports

Keep distinct hardware, checkpoint releases, exact request mix, trial units, SD and configured capacity together.

```js
catalog.records.filter(r => ['qwen38-flash-next-sglang-concurrency','qwen38-flash-next-tp2-concurrency-context'].includes(r.id))
```

### Read actual 5090 context gates and unpaired memory telemetry

Retain the 104805-token retrieval, 32K-only peaks, unknown sampling scope and separate ngram speculation/embedding states.

```js
catalog.records.filter(r => r.id === 'qwen38-flash-next-5090-offload-gates')
```

### Read the dated VQ follow-up with its missing runtime results

Return the complete A12 interpretation including the October 18 bounded source recheck. The absent 1017 VQ serving measurements remain unknown.

```js
articles.articles.filter(a => a.id === 'A12')
```

## Concurrency, repetition units and model components

Keep 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.

### Read Qwen long concurrent batches, memory scopes and failures

Return A15 complete, preserving original source identities, concurrent batch conditions, campaign-only peak and open-PR failure scope.

```js
articles.articles.filter(a => a.id === 'A15')
```

### Read GLM, MiniMax and Kimi deployments with score limitations

Keep complete versions, raw medians, false quality gates, July/August dates, missing peak RAM, subset sizes and timeout retries.

```js
articles.articles.filter(a => ['A16','A17','A18'].includes(a.id))
```

## Memory scope, quality scores and retries

A 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.

### Read paired Qwen memory and actual recovery reruns

Keep the unchanged concurrent peak, different 500K peak, production/validation hardware, invalid first ship run and two-event rerun.

```js
articles.articles.filter(a => a.id === 'A19')
```

### Read Spark deployments, score retractions and build-specific retries

Return full GLM and MiMo guides, preserving historical scores, EXL3 withdrawal, missing current task rerun, exact versions and C1/C8 denominators.

```js
articles.articles.filter(a => ['A20','A21','A22'].includes(a.id))
```

## Paired memory, withdrawn scores and invalid attempts

Match 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.