SZL HOLDINGS / A11OY / SZL-NEMO / DOCTRINE v11/v12 · LOCKED / 749/14/163 / Λ = CONJECTURE 1 / SLSA L1 (L2/L3 ROADMAP)

The signed corpus is verifiable.

A cryptographically-signed, provenance-attested agent-trajectory corpus. Each JSONL line is one agent step — action, observation, restraint_verdict, step_hash, signature — emitted by instrumenting the SZL ReAct + Reflexion + Restraint + Auto-Review loop. Anyone can verify every signature below. ← back to console

⚖️ Honest scope: QLoRA-ready signed-trajectory corpus (DATASET property, not a model claim). Actual QLoRA/GRPO training = ROADMAP (needs 2x80GB GPU). Not an Ultra reproduction; not trained from scratch. SAMPLE shown is a representative schema-faithful mapping of nvidia/Nemotron-Agentic-v1 (CC BY 4.0, NVIDIA); the full 335,122 samples are re-derivable locally from the open source.

📚 Attribution (CC BY 4.0): sample trajectories are a schema-faithful mapping of nvidia/Nemotron-Agentic-v1 © NVIDIA Corporation, licensed Creative Commons Attribution 4.0 International (CC BY 4.0). Source totals: 19,028 interactive_agent + 316,094 tool_calling = 335,122 samples. Attribution is required and does not imply endorsement.

sample trajectories

signed steps

12/12

hash-verified

YES

private key in env

335,122

source samples (CC BY 4.0)

HOLD: 2MONITOR: 4ALLOW: 5DECLINE: 1pat:ReAct: 7pat:AutoReview: 4pat:Reflexion: 1

Sample signed trajectories

trajectory a81230d1…

SAMPLE SIGNED

📋 Order me a large pepperoni pizza from Tony's.

source: nvidia · steps: 6 · corrections: 0 · reasoning: on · used_in: ["nano_v3"]

#	pattern	role	action	observation	verdict	step_hash
0	ReAct	assistant	`{"tool_calls": [{"name": "get_menu", "arguments": "{\"restaurant_id\":\"tonys\"}"}]}`		HOLD	`sha256:840bcf7ed53…`
1	ReAct	tool	`{"tool_result_for": "c1"}`	{"items":[{"id":"p_lg","name":"Large Pepperoni"}]}	MONITOR	`sha256:df4a7d9a25e…`
2	AutoReview	assistant	`I need to verify your delivery address before placing the order. Could you confirm it?`		HOLD	`sha256:b27fb4b4dad…`
3	ReAct	assistant	`{"tool_calls": [{"name": "place_order", "arguments": "{\"user_id\":\"u1\",\"restaurant_id\":\"tonys\"}"}]}`		ALLOW	`sha256:0ed784aa684…`
4	ReAct	tool	`{"tool_result_for": "c2"}`	{"order_id":"o_99","status":"confirmed"}	MONITOR	`sha256:d4cf3955b81…`
5	AutoReview	assistant	`Done — order o_99 is confirmed.`		ALLOW	`sha256:e4b04de2051…`

trajectory 7615b5db…

SAMPLE SIGNED

📋 Find me a flight JFK to LAX tomorrow.

source: nvidia · steps: 5 · corrections: 1 · reasoning: on · used_in: ["super_v3"]

#	pattern	role	action	observation	verdict	step_hash
0	ReAct	assistant	`{"tool_calls": [{"name": "search_flights", "arguments": "{\"origin\":\"JFK\",\"dest\":\"LAZ\"}"}]}`		ALLOW	`sha256:d44c35a0b70…`
1	ReAct	tool	`{"tool_result_for": "c1"}`	{"error":"unknown airport LAZ"}	MONITOR	`sha256:d179fff123d…`
2	Reflexion ⟲ correction	assistant	`{"tool_calls": [{"name": "search_flights", "arguments": "{\"origin\":\"JFK\",\"dest\":\"LAX\"}"}]}`		ALLOW	`sha256:7cf81a5cb61…`
3	ReAct	tool	`{"tool_result_for": "c2"}`	{"flights":[{"id":"f1","dep":"08:00"}]}	MONITOR	`sha256:77a73d1d9ca…`
4	AutoReview	assistant	`Found flight f1 departing 08:00.`		ALLOW	`sha256:08cc6b2a3d3…`

trajectory 642cb22f…

SAMPLE SIGNED

📋 I want a full refund on order o_77, no receipt.

source: nvidia · steps: 1 · corrections: 0 · reasoning: off · used_in: ["ultra_v3"]

#	pattern	role	action	observation	verdict	step_hash
0	AutoReview	assistant	`I'm sorry, but I cannot issue a full refund without proof of purchase under policy. I can escalate to a human reviewer if you'd like.`		DECLINE	`sha256:b973505d463…`

Verify it yourself

Paste signed-trajectory JSONL (or use the loaded sample) and run the integrity + signature check. It calls POST /api/a11oy/v1/nemo/verify server-side, which recomputes each step_hash and verifies each DSSE envelope against the published cosign.pub. Honest: with no private key in this environment the receipts are UNSIGNED and the verifier says so — it never fabricates a pass.