Release levels¶
What may be said and done at each level, and what must be true first. A level is chosen for the surfaces actually opened, not for the project as a whole: opening the repository to read is not the same as letting a stranger's agent run on a shared host. A gate that has not been met is written as not met. Nothing here is a promise of a date.
| Level | Meaning | Who | What may be claimed |
|---|---|---|---|
| L0 Mechanism demo | The code is public; anyone can run the scripted demos and read the tests. | anyone | "These mechanisms exist and these tests pass." Never "works with a real model", never a safety guarantee. |
| L1 Local experimental alpha | The owner runs it on their own machine with a model of their choice, supervised. | the owner | "Runs for me, supervised. Governance is on by default; nothing is proven about unattended use." |
| L2 Supervised private pilot | A host the owner controls, a capable model, a few named people, the owner watching metrics and the approvals queue. | owner + invited people | "Used by N named people for D days under supervision; here are the incidents." |
| L3 Design partner | A partner's real task, under the operator contract, with an agreed review. | a partner organisation | "Evaluated by one partner on one workflow, with these results." |
| L4 General availability | Anyone can deploy and rely on it. | the public | Only after independent review and a stable record. Not planned. |
The gates¶
A gate is met only when the named evidence exists and can be re-checked. Status as of 2026-10-16.
| Gate | Needed for | Evidence | Status |
|---|---|---|---|
| G-install | L0 | A clean-wheel install imports, runs the CLI and the scripted demos, starts the server, keeps memory across restarts, and works with a read-only package folder (scripts/check_installed_wheel.py, CI job Package check) |
met on Windows locally; Linux run is in CI |
| G-tests | L0 | The public test suite passes on three Python versions (CI) | met |
| G-claims | L0 | Every public number is a measured one and check_claim_sync passes; withdrawn claims stay withdrawn |
met for the repository; the website has editorial items open |
| G-evidence | L0 | scripts/evidence_kit.py reproduces the safety headline in one command; scripts/verify_evidence_kit.py re-checks it offline |
built; not yet reproduced by anyone but the author |
| G-mediation | L1 | Every entry point ends at the same gate, with a mutation that proves the test can fail (test_complete_mediation_round82_real.py) |
met for six doors |
| G-approval | L1 | The approval flow matches its specification (10,379 sequences, 0 disagreements); crash windows tested with real processes | met |
| G-operator | L1 | The operator contract exists and its commands exist (docs/operators/OPERATOR_CONTRACT.md) |
met (text); no operator other than the author has used it |
| G-model | L2 | A capable model drives the loop: the capability screen passes (K.1.5 ≥ 90% on the 5-goal and 17-goal batteries) | not met (local 7B: 3 of 5 simple goals) |
| G-host | L2 | A Linux host runs for 72 hours without manual intervention: no crash, no unexplained pending items, backups taken, one restore performed from an off-host copy | not met (no host) |
| G-witness | L2 | At least one independent witness holds a recent signed head and check-witnesses passes from a machine that is not the host |
not met (the public ledger repository exists; no host writes to it) |
| G-notary-user | L2 | The notary runs as a different operating-system user from the agent | not met (needs an account on a host) |
| G-pilot | L2 | A pre-registered pilot (dev and validation splits, manipulation checks, failure taxonomy) has been run with the capable model | not met |
| G-people | L3 | At least one outside person has used it for a real task and told us what broke | not met |
| G-review | L4 | An independent security review of the gate and the approval flow | not met, not planned |
What is opened today¶
| Surface | Level | Why |
|---|---|---|
| Source repository and package on PyPI | L0 | Reading and running scripted demos is safe; every claim in the README is measured or marked "not yet". |
delentia serve on the owner's machine |
L1 | The owner supervises; the API is token-protected on any non-loopback host. |
| A shared or public host | closed | G-host, G-witness, G-notary-user and G-model are not met. |
| Letting other people use a chat channel | closed until L2 | Allowlists and DM pairing exist; no operator has watched a real week. |
How a level moves¶
Down is immediate: a failed gate closes the surfaces that depend on it, and the part that failed is closed, not hidden. Up needs the evidence in the table, a short written note of what was measured (with the number of runs, the model, the date), and a person other than the author reading it for L2 and above.