# SparkLang llms-full
version: 0.8.0
sha: df8ec5b5a0f8
verified: 2026-09-04
site: https://sparklang.dev

## index
https://sparklang.dev/llms.txt
https://sparklang.dev/llms-full.txt
https://sparklang.dev/agent.md
https://sparklang.dev/robots.txt

## core
# Core
Lexical rules, let/print, expect gates, pipelines, and include.
This is the smallest useful .spark surface.
Does not: Does not train models or call external AI.
Verbs: let, print, expect, pipeline, include
Example: examples/dry_run.spark
```
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "fixture-base" out "out/train/job-dry-001" backend "http" method "spark_distill_cpu" -> job
model status "job-dry-001" -> status
expect contains job fixture "examples/fixtures/train/want_accepted.txt"
expect contains status fixture "examples/fixtures/train/want_succeeded.txt"
```

## models
# Models
model train/status/plan, LoRA methods, trainer contract, and
ModelScope inspect/diff/probe/decompile/patch.
Does not: Does not load staged adapters into serving without an owner token.
Verbs: model train, model status, model inspect, model diff, model probe, model decompile, model patch, model eval, model stage
Example: examples/train_sam_lora.spark
```
# VoiceCore reference — Sam store persona LoRA (redacted for public)
# Dataset: #321 mini-freeze 140 rows
# sha256: 272b64eb1d40aaf8a186df6a2f6c8e7090a1d69d03ac4928ac75a246f8e439f0
# Live trainer: SPARK_TRAIN_URL (host redacted)
# Histogram floors of 20 fail on a thin freeze — that is the gate.
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "hf://Qwen/Qwen3-27B" out "out/train/job-lora-001" backend "http" method "lora_bf16" -> job
model status "job-lora-001" -> status
expect gte status $.metrics.golden_bucket 0.8
expect eq status $.metrics.grounding_facts true
expect eq status $.metrics.rachel_bleed 0
expect histogram_min status W2a 20
expect histogram_min status W2b 20
expect histogram_min status W-SNC 20
expect histogram_min status W-TF-RECOVER 20
```

## ai-ops
# AI ops
ask, classify, extract, embed/retrieve, tools, and abstain heads.
Does not: Does not invent Bifrost task aliases from prompt text.
Verbs: ask, classify, extract, embed, retrieve
Example: examples/dry_run.spark
```
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "fixture-base" out "out/train/job-dry-001" backend "http" method "spark_distill_cpu" -> job
model status "job-dry-001" -> status
expect contains job fixture "examples/fixtures/train/want_accepted.txt"
expect contains status fixture "examples/fixtures/train/want_succeeded.txt"
```

## integration
# Integration
http, shell, and host embed for Python, JS, and C.
Does not: Does not open network in dry-run without fixtures.
Verbs: http get, http post, shell
Example: examples/dry_run.spark
```
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "fixture-base" out "out/train/job-dry-001" backend "http" method "spark_distill_cpu" -> job
model status "job-dry-001" -> status
expect contains job fixture "examples/fixtures/train/want_accepted.txt"
expect contains status fixture "examples/fixtures/train/want_succeeded.txt"
```

## voice
# Voice
Voice verbs for speak and related store-persona work.
CallsBack workers are a separate product surface.
Does not: Does not ship CallsBack.ai workers from this site.
Verbs: speak
Example: examples/train_sam_lora.spark
```
# VoiceCore reference — Sam store persona LoRA (redacted for public)
# Dataset: #321 mini-freeze 140 rows
# sha256: 272b64eb1d40aaf8a186df6a2f6c8e7090a1d69d03ac4928ac75a246f8e439f0
# Live trainer: SPARK_TRAIN_URL (host redacted)
# Histogram floors of 20 fail on a thin freeze — that is the gate.
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "hf://Qwen/Qwen3-27B" out "out/train/job-lora-001" backend "http" method "lora_bf16" -> job
model status "job-lora-001" -> status
expect gte status $.metrics.golden_bucket 0.8
expect eq status $.metrics.grounding_facts true
expect eq status $.metrics.rachel_bleed 0
expect histogram_min status W2a 20
expect histogram_min status W2b 20
expect histogram_min status W-SNC 20
expect histogram_min status W-TF-RECOVER 20
```

## ide
# IDE
IDE verbs that are verified: new, open, save, run, buffer, ask, show.
Does not: Paint is PPM wire only — not a language op.
Verbs: ide
Example: examples/dry_run.spark
```
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "fixture-base" out "out/train/job-dry-001" backend "http" method "spark_distill_cpu" -> job
model status "job-dry-001" -> status
expect contains job fixture "examples/fixtures/train/want_accepted.txt"
expect contains status fixture "examples/fixtures/train/want_succeeded.txt"
```

## experimental
# Experimental
Browser/MITM, engine, network, binary, CUDA/PCIe, OS blueprints,
and encrypt gateway — each with a status badge.
Does not: Experimental pages do not claim production readiness. decompile and
patch stay experimental until probe AUROC is at least 0.85.
Verbs: model decompile, model patch
Example: examples/dry_run.spark
```
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "fixture-base" out "out/train/job-dry-001" backend "http" method "spark_distill_cpu" -> job
model status "job-dry-001" -> status
expect contains job fixture "examples/fixtures/train/want_accepted.txt"
expect contains status fixture "examples/fixtures/train/want_succeeded.txt"
```

## conformance
# Conformance
Statement to example file to last CI run. Proof narration lives here,
not in product prose.
Does not: Does not replace dry-run CI — it reports it.
Example: examples/dry_run.spark
```
model code
model train dataset "examples/fixtures/train/dataset.jsonl" base "fixture-base" out "out/train/job-dry-001" backend "http" method "spark_distill_cpu" -> job
model status "job-dry-001" -> status
expect contains job fixture "examples/fixtures/train/want_accepted.txt"
expect contains status fixture "examples/fixtures/train/want_succeeded.txt"
```

## conformance
statement	reality	since	grammar
model train	real	0.6.0	model train dataset "PATH" base "ID" out "DIR" backend "http" method "METHOD" -> job
model status	real	0.6.0	model status "JOB_ID" -> status
model inspect	real	0.8.0	model inspect PATH -> summary (CLI: spark-modelscope)
model diff	real	0.8.0	model diff A B -> delta (adapter heatmap)
model probe	real	0.8.0	model probe base symbols "DIR" -> p (AUROC < 0.85 = unusable)
model decompile	experimental	0.8.0	model decompile contrast→vector — EXPERIMENTAL until probe AUROC ≥ 0.85
model patch	experimental	0.8.0	model patch vector|rome|memit + scale 0.4–1.2 — EXPERIMENTAL until probe AUROC ≥ 0.85
model eval	real	0.8.0	model eval ARTIFACT suite "PATH" -> e
model stage	real	0.8.0	model stage ARTIFACT marker "ID" -> staged (serving_loaded: false always)
expect contains	real	0.6.0	expect contains NAME fixture "PATH"
expect gte	real	0.7.0	expect gte NAME $.path N
expect lte	real	0.7.0	expect lte NAME $.path N
expect eq	real	0.7.0	expect eq NAME $.path V
expect histogram_min	real	0.7.0	expect histogram_min NAME CLASS N
expect score	real	0.7.0	expect score NAME $.path using "URL" >= N
http get	real	0.6.0	http get "URL" fixture "PATH" timeout N -> resp
http post	real	0.6.0	http post "URL" body "{…}" fixture "PATH" timeout N -> resp
spark_distill_cpu	real	0.6.0	Trainer method — reply-class student → weights.pt
spark_pref_pack	real	0.6.0	Trainer method — preference pairs + ranker
spark_playbook_fit	real	0.6.0	Trainer method — intent→playbook router
spark_faq_index	real	0.6.0	Trainer method — FAQ dual-encoder
lora_bf16	real	0.7.0	Trainer method — Unsloth/PEFT LoRA (GPU) or fixture
hf://	real	0.7.0	hf://org/name on dataset/base/out (hf_resolve.py)
GitHub Action dry-run	real	0.7.0	uses: Michael20251/sparklang/.github/actions/dry-run

## grammar.ebnf
(* SparkLang homepage verbs — generated from status.py *)
program     = { statement } ;
statement   = model_stmt | expect_stmt | http_stmt ;
model_stmt  = train | status | inspect | diff | probe
            | decompile | patch | eval | stage ;
train       = "model", "train", "dataset", string, "base", string,
              "out", string, "backend", string,
              [ "method", string ], [ "->", ident ] ;
status      = "model", "status", string, [ "->", ident ] ;
inspect     = "model", "inspect", path, [ "->", ident ] ;
diff        = "model", "diff", path, path, [ "->", ident ] ;
probe       = "model", "probe", "base", "symbols", string,
              [ "->", ident ] ;
decompile   = "model", "decompile", { token } ; (* experimental *)
patch       = "model", "patch", { token } ;     (* experimental *)
eval        = "model", "eval", ident, "suite", string,
              [ "->", ident ] ;
stage       = "model", "stage", ident, "marker", string,
              [ "->", ident ] ;
expect_stmt = contains | gte | lte | eq | histogram_min | score ;
contains    = "expect", "contains", ident, "fixture", string ;
gte         = "expect", "gte", ident, jsonpath, number ;
lte         = "expect", "lte", ident, jsonpath, number ;
eq          = "expect", "eq", ident, jsonpath, value ;
histogram_min = "expect", "histogram_min", ident, ident, number ;
score       = "expect", "score", ident, jsonpath, "using", string,
              ">=", number ;
http_stmt   = http_get | http_post ;
http_get    = "http", "get", string, [ "fixture", string ],
              [ "timeout", number ], [ "->", ident ] ;
http_post   = "http", "post", string, [ "body", string ],
              [ "fixture", string ], [ "timeout", number ],
              [ "->", ident ] ;
string      = '"', { char }, '"' ;
ident       = letter, { letter | digit | "_" | "-" } ;
jsonpath    = "$", { ".", ident | "[" , digit, "]" } ;
path        = string | ident ;
value       = string | number | "true" | "false" ;