v1.0.1 · stable
Capability that exists is not capability you can depend on.
DIEEC models an agent as a continuous loop between a finite working field and an unbounded external field — it does not receive a whole world and then answer from a closed context. It compiles the task-relevant world into operable structure, step by step, and has to prove which world it changed.
The v1.0 proposition
capability exists ≠ capability is dependable ≠ capability is migratable ≠ version is releasable
v0.1–v0.9 kept adding capability: dynamic working fields, asynchronous causality, tail risk, record-and-replay, drift, privacy, credit assignment, signed attestation. But a growing pile of modules is not a runtime. If every module emits its own certificate format, if upgrades cannot read old data, if the public API can change freely, and if passing tests is treated as clearance to ship, then what exists is a collection of prototypes — not a layer anything else can build on.
What it actually is
A symbolic runtime plus a benchmark. Everything runs locally and deterministically: no remote language model is called, no production data is touched. The agent groups G0–G7 are ablations — G0 has almost no machinery, G7 has all of it — so the cost and benefit of each governance layer can be read off directly instead of argued about.
Seven layers
Internal state · dynamic working field · operational tokens · tools and external effects · verification and boundaries · unified certificates · release governance. The last one is what v1.0 added, and it is the one that decides whether a candidate version may be committed into the stable authoritative world at all.
What the benchmark shows
G7 costs more than G6. That is the finding, and it is stated in that direction on purpose.
It would have been easy to report only the metrics that improved. Governance is not free: the unified certificate, the compatibility check and the release gate add measurable work to every episode. The honest question is not whether governance costs nothing, but whether its cost is visible, budgetable, and buys a verifiable reduction in risk.
Honest boundaries
These are declared by the release itself, not added afterwards.
- No remote large language model is called anywhere in the benchmark.
- No real personal data and no production data is used.
- No OCI runtime was available, so containment is never claimed to have executed — only the genuine proof that it did not.
- Release attestation uses an ephemeral local Ed25519 key. It is not an external CA, a hardware root, or a third-party timestamp.
- G7's 100% holds inside this controlled symbolic benchmark. It is not a claim about open-world reliability.