One reviewable file from train → patch → gate. Dry-run with no keys; LoRA when you have a GPU; ModelScope to see and patch what's inside.

Run the dry-run (30s) Read the agent contract

The 90-second proof

VoiceCore train_sam_lora.spark (redacted). Histogram floors fail on purpose. Numbers are from the live 2026-09-03 job, not a toy.

# 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

Download train_sam_lora.spark

expect	result	got
gte golden_bucket ≥ 0.8	PASS	0.9929
eq grounding_facts true	PASS	true
eq rachel_bleed 0	PASS	0
histogram_min W2a ≥ 20	FAIL	12
histogram_min W2b ≥ 20	FAIL	0
histogram_min W-SNC ≥ 20	FAIL	17
histogram_min W-TF-RECOVER ≥ 20	PASS	111

class_histogram (live)

{
  "W2a": 12,
  "W2b": 0,
  "W-SNC": 17,
  "W-TF-RECOVER": 111
}

adapter diff heatmap (live)

{
  "layer_count": 2,
  "delta_norms": {
    "base_model.model.layers.0.self_attn.q_proj.lora_A.weight": 12.1372,
    "base_model.model.layers.0.self_attn.q_proj.lora_B.weight": 0.0
  },
  "base": "hf://Qwen/Qwen3-27B",
  "rank": 16,
  "note": "live 2026-09-03 job-lora-sam-live-20260903; stub_weights"
}

What's real today

Same table as spark --status.

verb	reality	since
`model train`	real	0.6.0
`model status`	real	0.6.0
`model inspect`	real	0.8.0
`model diff`	real	0.8.0
`model probe`	real	0.8.0
`model decompile`	experimental	0.8.0
`model patch`	experimental	0.8.0
`model eval`	real	0.8.0
`model stage`	real	0.8.0
`expect contains`	real	0.6.0
`expect gte`	real	0.7.0
`expect lte`	real	0.7.0
`expect eq`	real	0.7.0
`expect histogram_min`	real	0.7.0
`expect score`	real	0.7.0
`http get`	real	0.6.0
`http post`	real	0.6.0
`spark_distill_cpu`	real	0.6.0
`spark_pref_pack`	real	0.6.0
`spark_playbook_fit`	real	0.6.0
`spark_faq_index`	real	0.6.0
`lora_bf16`	real	0.7.0
`hf://`	real	0.7.0
`GitHub Action dry-run`	real	0.7.0

Four CPU methods + lora_bf16 + numeric expects + hf:// + GitHub Action + ModelScope inspect/diff/probe. decompile/patch experimental until probe AUROC ≥ 0.85.

Who it's for

Solo / CPU — Dry-run gate in CI — no keys, no GPU.
GPU teams — LoRA (lora_bf16) + adapter diff heatmap.
Operators — Behavior fixes as reviewable .spark programs with a ledger.

Install

pip install sparklang

spark --dry-run examples/train_eval.spark

uses: Michael20251/sparklang/.github/actions/dry-run@main

docker compose -f tools/spark-train-ref/docker-compose.yml up

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"

curl -sS -X POST "$SPARK_TRAIN_URL/jobs" -H 'Content-Type: application/json' -d '{"dataset":"examples/fixtures/train/dataset.jsonl","base":"fixture-base","out":"out/train/job-dry-001","backend":"http","method":"spark_distill_cpu"}'

Docs · Playground · Source · License

Where we are today

No hosted public trainer yet — SPARK_TRAIN_URL is yours.
GitHub Marketplace Action publish is an owner action.
model decompile / model patch stay experimental until probe AUROC ≥ 0.85.
model stage never sets serving_loaded true.
Voice/PSTN remain gated demos — not a production claim.
Installer kit on this site may lag the language version.
Upstream public mirror may trail this private repo.

version	roadmap
0.6	CPU train methods + dry-run expect + http get/post
0.7	lora_bf16 + numeric expects + hf:// + pip + Action
0.8	ModelScope inspect/diff/probe; decompile/patch experimental
next	Hosted trainer (coming) · LSP · Marketplace Action

Upstream public mirror: https://github.com/sparklang-dev/sparklang. Language-verb grammar PR only — not a hosted trainer.

install	`pip install sparklang`
binary	downloads
dry-run	`spark --dry-run examples/train_eval.spark`
trainer	`POST /jobs` · `GET /jobs/{id}`

exit	meaning
0	success / expect pass
1	expect fail or language fail
2	usage / unknown method / missing program
4	credential unavailable or CAP miss

flag	meaning
`--require-token`	SPARK_REQUIRE_TOKEN=1 — live submit needs token
`--ledger PATH`	SPARK_LEDGER — append live submits as JSONL
`--allowlist`	SPARK_TRAIN_UNIT_ALLOWLIST — optional local-yield

verb	grammar
`model train`	`model train dataset "PATH" base "ID" out "DIR" backend "http" method "METHOD" -> job`
`model status`	`model status "JOB_ID" -> status`
`model inspect`	`model inspect PATH -> summary (CLI: spark-modelscope)`
`model diff`	`model diff A B -> delta (adapter heatmap)`
`model probe`	`model probe base symbols "DIR" -> p (AUROC < 0.85 = unusable)`
`model decompile`	`model decompile contrast→vector — EXPERIMENTAL until probe AUROC ≥ 0.85`
`model patch`	`model patch vector\|rome\|memit + scale 0.4–1.2 — EXPERIMENTAL until probe AUROC ≥ 0.85`
`model eval`	`model eval ARTIFACT suite "PATH" -> e`
`model stage`	`model stage ARTIFACT marker "ID" -> staged (serving_loaded: false always)`
`expect contains`	`expect contains NAME fixture "PATH"`
`expect gte`	`expect gte NAME $.path N`
`expect lte`	`expect lte NAME $.path N`
`expect eq`	`expect eq NAME $.path V`
`expect histogram_min`	`expect histogram_min NAME CLASS N`
`expect score`	`expect score NAME $.path using "URL" >= N`
`http get`	`http get "URL" fixture "PATH" timeout N -> resp`
`http post`	`http post "URL" body "{…}" fixture "PATH" timeout N -> resp`

Agent contract

what-is-real