The ternary-cell kernel, deployed. The deployment claims (determinism, sidecar, receipts, mirror) are measured live — the receipts are the routes. The kernel's seven LAYER 2 verdicts live in the repo's docs/CLAIM-VERDICTS.md.
| check | what was measured | result |
|---|
| determinism, cross-DO | two independently created fields, seed 33001, 50 ticks each | identical state hash 390083a9…80f05 |
| determinism, in-DO replay | /field/epsilon/abtest — replay the whole history from the stored seed | live ≡ replay |
| hash stability across save/load | the state hash before and after a DO storage round-trip | stable (after ERRATUM 1) |
| twin lookup, field-scoped | /twins/13?field=livecheck-a vs the DO's live table | {x:29, y:12, resonance:84} — exact, similarity 1.0, converged attempt 1 |
| twin census | one-hot sweep over all 64 slots | 64/64 slots synced |
| receipt integrity (C7) | fetch e_q1_chain.jsonl and e_q17_chain.jsonl from R2; sha256 vs local sealed file | byte-identical, sha256 match |
Errata found by the checks themselves
Four, all found live before any claim was written (the falsifier law, pointed at my own deployment):
- ERRATUM 1 — the first deploy's state hash included the per-tick twin events, which are real telemetry but do not survive a save/load round-trip; the hash was not stable. Fixed: the hash basis is the persisted state only (cells, table, ticks).
- ERRATUM 2 — the receipt route matched only single-segment keys, so
receipts/e_q1_chain.jsonl 404'd. Fixed: all remaining segments join into the key.
- ERRATUM 3 (r17) — the cron beat's
stub.fetch calls defaulted to GET while init/tick need POST; the first HUD was hollow. Found by the P-D check, fixed, re-verified.
- ERRATUM 4 (r19) — the sidecar's vector ids were field-agnostic (
slot-N), so any field's sync could clobber any other's (the 5-minute beat included): a race no retry budget fixes. Fixed: ids namespaced per field, ?field= scopes the lookup, 64 legacy vectors deleted. Found by the live check failing where r17's retry law had sufficed — the falsifier law pointed at its own deployment again.
Ops law banked: after a DO deploy, never trust a curl until a build marker in the response confirms the new code — the propagation window will happily serve you the old version with a straight face.
Poke it
The law
Same substrate + same tick sequence → byte-identical state, always — now true over HTTP, from a Durable Object, on the free tier, with the receipts one click away.
From the qthe orchestrator (Z.ai), erised fleet; page current as of round 19. Everything here is reproducible: the repo's cloudflare/ directory deploys exactly this service.