EVIDENCE KERNEL · READ-ONLY SNAPSHOT

Don’t read the claims.
Re-hash the evidence.

A provenance-first console over the SZL Forge snapshot: run manifest, deterministic evaluation receipts, formula registry, governed source ledger, and curriculum. Your browser fetches the packaged artifacts, recomputes their SHA-256 against the manifest declarations, and re-verifies each receipt’s canonical hash — live, on every load.

Source-aware confirmation. When is another observation worth buying? Replay 5,184 recorded runs, compare seven policies and explore fixed-trace costs. Synthetic research: shared-bias comparison passed; clean-sensor guard failed; overall registered gate FAILED.

Evidence boundary. Two published fine-tunes, one GGUF derivative, and one learned kernel contain weight bytes. Eleven additional Hub entries are software kernels, recipes, or cards—not trained models. Signed receipts prove declared-key integrity; they do not establish independent quality, safety, or authorship.

ATELIER lock. This page is SNAPSHOT evidence. Curriculum status is BLUEPRINT_NOT_TRAINED. It is not a trainer, not Unsloth Studio, and not a Jobs launcher. Bounded serve is SZLHOLDINGS/szl-model-inference-lab only. Energy-attested-runs 8/8 are SIMULATED. Ask & Act is not a live control plane.

01 · MODEL / KERNEL PORTFOLIO

What is weighted, what is software, and what is still bounded

02 · INTEGRITY — RECOMPUTED HERE

Manifest-to-Space byte checks

WAITING — run_manifest.json receipt hash (SHA-256 over canonical JSON minus receipt_sha256)

WAITING — eval_receipt.json receipt hash, same rule

WAITING — packaged artifacts re-hashed vs manifest declarations

Declared input	Packaged here as	State	Expected SHA-256	Recomputed in this browser

A matching hash proves byte identity only. Inputs marked NOT PACKAGED live on the owner’s training host and are intentionally not shipped (e.g. training data) — their absence is stated, not hidden.

Documented post-receipt corrections. thesis_formula_index.json was intentionally corrected twice after the July 2026 run receipt. The 2026-10-06 correction means TH10 is no longer marked GREEN (szl-holdings/szl-forge#594). On 2026-10-14, 64 further entries whose cited Lean theorem is not declared in lutar-lean were re-labelled from GREEN to TRACKED. See meta.corrections and each corrected entry’s correction field in that file. The July 2026 run receipt in run_manifest.json is kept byte-identical as a historical record and is not re-minted, so it still pins the pre-correction bytes. The MISMATCH shown for this file therefore reflects documented post-receipt corrections, not tampering. No receipt covers the corrected bytes, and this note does not verify them.

03 · EVALUATION

Deterministic fixture checks — scoped, not a benchmark

loading…

Case	Category	Result	Response SHA-256

04 · FORMULA REGISTRY

Search the packaged formula index

Statuses are declarations from the packaged registry. This page does not execute Lean and never elevates a formula to proven.

05 · SOURCE LEDGER

Default-deny scientific source governance

06 · CURRICULUM

Governed curriculum blueprint — not a training claim

07 · RAW EVIDENCE

Verify from outside this page

# every artifact this page rendered, hash it yourself:
for f in run_manifest.json eval_receipt.json training_summary.json thesis_formula_index.json science_source_ledger.json curriculum.json; do
  curl -sL https://huggingface.co/spaces/SZLHOLDINGS/szl-forge-lab/resolve/main/$f | sha256sum
done

Don’t read the claims.Re-hash the evidence.