A real evidence pack.
Check it without trusting us.
This is the pack we sealed on 30 September 2026 when we ran the 258-scenario agentic screen against kimi-k2.6 (Moonshot AI) for the industry-standard tests comparison published on 2 October. Nothing in it has changed since. The manifest, the file hashes and the Bitcoin stamp are how you know that — not our word.
It is a preliminary record, not a sealed finding: three of the 258 turns still await a human ruling, and the pack says so about itself. We publish it anyway, because a buyer should see what the deliverable looks like before paying, flaws included.
Download both. Run one command.
Potestas_Specimen_Evidence_Pack_2026-09-30.zip · 1.7 MB
SHA-256 45491c0797ec94fa42d4e02b95bb9d558d15384516ee73f2512e9c3639b896e7
katana_verify_pack.py · one Python file, standard library only. The Bitcoin check asks a public block explorer for one block header and needs pip install opentimestamps-client.
SHA-256 33689692b0435eb64b44b8b42274641f8b482a8287e7c62425399735f84d8e61
python katana_verify_pack.py Potestas_Specimen_Evidence_Pack_2026-09-30.zip
You do not need us for any of it. The verifier recomputes every hash, every count and every index from the raw files in the pack and compares them with what the manifest claims. It was written to read a pack without the code that produced it.
What it prints on this pack. All fifteen checks, 14 October 2026.
[PASS] file integrity (recomputed sha256 vs manifest) — 8 file(s) hashed and matched
[PASS] no unlisted files in pack
[PASS] safety index recomputed independently — manifest 88.3, this tool 88.3
[PASS] security-critical count recomputed — manifest 21, this tool 21
[PASS] unresolved-turn count recomputed — manifest 3, this tool 3
[PASS] clean categories carry an upper bound (informational)
[PASS] probe coverage (distinct vs delivered) — 258 row(s) in the trace
[PASS] CSV cross-checked against the primary record — 258 turn(s)
[PASS] primary record present (re-gradeable)
[PASS] seal verdict consistent with its own blockers — 3 turn(s) await a human ruling, and the pack says so
[PASS] manifest records random_seed
[PASS] manifest records battery_hash
[PASS] manifest records instrument
[PASS] manifest records temperature_effective
[PASS] externally anchored in time — Bitcoin block 971410, mined 2026-10-14 14:00:10 UTC; merkle root matches the proof
LATE STAMP: this pack was sealed before stamping existed (30 September 2026); the block proves it existed by 14 October 2026
------------------------------------------------------------------------------
RESULT: 15 checks passed.
The last line is the one that was red on every pack we had ever produced until 14 October 2026. The stamp was built that day; this pack, sealed two weeks earlier, carries a late stamp — it proves the pack existed by 14 October, not on 30 September. Every pack we seal from now on is stamped the moment it is sealed, so its date is its run date.
What is inside.
- The findings report (PDF) — every scenario that broke and exactly how.
- The tool-call log (JSON lines, one record per turn) — what the model was handed, what it said, what it actually invoked, with the arguments. The primary record. Anyone can re-grade every turn from it.
- The per-turn table (CSV), the summary (text), the replay (HTML), the dashboard (JSON) and the charts (PNG) — all derived from the log, and the verifier checks that they agree with it.
- The encrypted transcripts — the full conversations, sealed under a per-engagement key the buyer receives separately. The hash of the key is in the manifest; the key is not in the pack.
- The manifest — the SHA-256 of every file above, the seed, the battery hash, the instrument identity, the run conditions — and its proof, a 700-byte OpenTimestamps file that ties the manifest's hash to a Bitcoin block.
What a pass does and does not mean.
It does mean the files are unaltered since sealing, the numbers in the manifest are the numbers the raw data produces, and the pack's claims about itself are consistent with its contents.
It does not mean any verdict is correct. Grading correctness is a separate question, answered by reading the log, not by recomputing hashes. Custody is not truth — which is why the log is in the pack.
The published specimen report is a different engagement (a text battery, August 2026) and shows what the written report looks like. This page shows what the sealed evidence looks like. Your own screen produces both, for your model, within 72 hours: $3,000, bought with a card.