frontier-54-runs — the SZL receipts chain

The public evidence chain of the SZL Holdings AI estate. Some receipt families are hash-chained, and every receipt carries its own SHA-256 (see Integrity, per family below). Every claim the estate makes traces to a receipt here — measured on our own hardware, fail-closed, never fabricated.

What this is

A living dataset: every training run, honesty measurement, officer cycle, governance decision, and energy reading lands here as a receipt with a SHA-256 digest. Unreachable sources produce UNAVAILABLE receipts, never silence (Kamachikuna Law 9: no silent failure).

The directories

Path	What lives there
census/	The cross-lab confabulation census — Anthropic (opus 5/6, sonnet 4/6, MEASURED), OpenAI, and the open-models leg (all four Salad-served Qwens 3/6)
benchmark/	6-probe abstention batteries on the estate's own trained models. The early candidates scored 3/6; quant-master-v1-candidate scored 5/6 (n=6, see caveats below). Also holds the hardware-leg (4090-leg, a6000-leg) and copy-constrained decode benchmarks
registry/	The model registry: 131 models reconciled by source, license, weights, receipts, provenance (snapshot of 2026-11-10; the Hub org has grown since)
officers/	Cycle records from the autonomous officer Spaces (Eternal Auditor, Estate Gardener, Dream Architect). UNANCHORED: see officers/ records below
doctrine/	The Kamachikuna constitution receipts and amendments
lean/	Lean 4 machine-checked theorem sources (5 proven, zero sorry)
energy/	Measured GPU energy receipts — NVML-sampled, UNAVAILABLE never zero
mesh/	The A6000 mesh heartbeats and job receipts (signed, allowlisted, once-only)
proving-ground/	Agent Proving Ground stage receipts (BOOT, IDENTITY, HONESTY, AUTHORITY, RECEIPTS, VERDICT), including rejected runs
collaboration/	Collaboration and protocol-preparation receipts
data/	receipts-index.jsonl, the viewer index (path, kind, ts, sha256). It is rebuilt periodically, so receipts uploaded after the last build are in the tree but not yet in the index
deploy/	Deployment attestations
monetization/	The x402 monetization declarations
receipts/	The core run receipts

How to verify

Every receipt carries a receipt_sha256 (or body_sha256) computed over its canonical JSON. Officers chain each receipt to its predecessor. Failed builds withhold verdicts — a grep on a failed build proves nothing, and the gates enforce it.

Integrity, per family

Family	Per-receipt SHA-256	Linked to the previous receipt	Signed
receipts/proving-*.json (Lean prove-or-abstain runs)	yes (row_sha256)	yes (prev_sha256, checked 2026-11-12)	no
proving-ground/agent-*.json (Agent Proving Ground)	yes (receipt_sha256 = SHA-256 of the sorted-key JSON without that field)	no	no
officers/	yes (row_sha256 or body_sha256)	only within one process lifetime; resets on restart	not verifiable (see below)

A self-hash shows a file is consistent with itself. It doesn't show the file wasn't replaced, because anyone with write access can recompute the hash. Until receipts are signed with a published key, this repo's git commit history is the ordering and tamper record.

officers/ records: UNANCHORED (added 2026-11-12)

The records under officers/ cannot be verified as coming from a known signer. They are kept as logs; no rows were deleted or changed.

• Dream Architect signs with an Ed25519 key stored in the Space's temporary disk. The Space has no persistent storage, so the key is replaced on every restart or rebuild, and the public key is self-asserted in each record.
• Eternal Auditor uses an HMAC secret from a Space setting, and its code falls back to a fixed default value visible in its public source when that setting is missing. A signature made with that value proves nothing.
• Estate Gardener records carry a self-hash and no signature.
• The other officer Spaces (Chief Engineer, Infra Warden, Cyber Shield, Research Council) sign with a key generated at each boot; no records from them were found here on 2026-11-12.

All seven officer Spaces were paused on 2026-11-12 pending a security review, so no new officers/ records are expected until they are reviewed and restarted. Do not cite an officers/ record as evidence of a check, fix or review.

quant-master-v1: what the chain actually shows (as of 2026-11-12)

• Training: worker receipt 8e8a31e4…c081, 200 steps, 1,135.9 s on one RTX A6000 (the log lives on the training node, not in this repo).
• Battery: benchmark/estate-battery-quant-master-v1-candidate.json: 5/6.
• Proving ground: 15 receipts over three runs (20:30, 20:41 and 20:49 UTC). Run 1 was REJECTED because the SDK wasn't installed. Run 2 was REJECTED because AUTHORITY scored 2/5 on a DENIED/BLOCKED vocabulary mismatch. Run 3 was PROMOTED_TO_AGENT (da5536bf…).

Caveats:

• n = 6. The 95% Wilson interval for 5/6 is about 0.44–0.97. Against the 3/6 baseline, a one-sided Fisher exact test gives p ≈ 0.27, which is not significant.
• The battery uses greedy decoding and a keyword scorer. The curriculum contains the scorer's trigger phrases ("I don't know" and "not fabricate", 300 times each).
• The AUTHORITY stage checks the SDK gate function, not the model, and the RECEIPTS stage checks only that a hash exists. A promotion therefore reflects the 6-question honesty screen only.
• Stored replies are cut to 150 characters, so some verdicts can't be re-checked from the receipt.

Planned upgrades (none done yet): a larger held-out probe set (n ≥ 100) with a semantic judge, full replies in each receipt, signed and linked proving-ground receipts, and AUTHORITY probes that make the model propose tool calls.

The estate

• Company: a-11-oy.com · sponsors
• Live dashboard: the SI command center — census and receipts, live
• PURIQ markets: the read-only market-intelligence console (no wallet, order, or trading path)
• Eternal Auditor: the always-on audit officer
• Newton demo: the receipted agent
• GitHub: szl-holdings — canonical source

Kamachikuna doctrine v11: fail closed · every claim carries a receipt · abstain when you cannot prove · no silent failure · the loop audits itself.