ABERNATE

A release gate for automated work.

Automated workers are taking real actions. ABERNATE decides whether each action is permitted, at run time, and leaves a record that settles who allowed it.

A REAL RUN, VERBATIM. THE SHIPPED SAMPLE MISSION, RELEASE REQUESTED.

{
  "mission_id": "recover-sample-001",
  "operational_status": "PASS",
  "steps_planned": 4,
  "steps_executed": 4,
  "receipt_chain": "VALID",
  "release_required": true,
  "released": false,
  "squad_leader": "BLOCKED: no SQUAD LEADER clearance for mission recover-sample-001: run the go/no-go board and record a GO before release",
  "final_receipt_hash": "ec18a705d090f8d9c5b5ece7...",
  "metrics": {
    "completion_ratio": 1.0,
    "successful_ratio": 1.0,
    "authority_coverage": 1.0,
    "receipt_integrity": 1.0
  }
}

The work succeeded. Every step ran, the chain verified, every metric came back at 1.0. It still did not ship, because no human had recorded a decision. That is the product.

The problem this answers

The failure organizations are insuring against is a single one: the system did something nobody authorized, and afterwards there is no record that settles what happened or who allowed it.

Most tooling answers that at deploy time, with configuration and review. ABERNATE answers it at run time, on every action.

Authority is an intersection, not a setting

Four layers decide whether an action may happen. No layer can grant what a higher layer forbids.

1What the worker is capable of

2What the command permits

3What standing policy allows

4What the mission actually asked for

execution is conjunctive · one failed condition stops the action rather than being averaged away

A human release gate sits at the end. It can only block, never grant. A policy that omits the gate is refused at load, so it cannot be disabled by being forgotten.

Every step emits a hash-chained receipt. Whoever receives that ledger can verify it with a shipped command, without our help and without trusting us.

What is unusual about it

It refuses by construction, not by threshold A worker with five strong governance properties and one near-dead one does not average out to acceptable. It is refused, and the failing property is named.
It ships the record of its own defeats Eight forgeries were run against its evidence chain. Five were caught outright, two more were missed and then fixed, and one is structural. That last one is stated rather than buried, and asserted as a passing test so it cannot be quietly dropped. All eight were re-run against the shipped verifier before this page was published.
Two prior versions had their flagship claim measured false Both were repaired and both repairs were reversion-proven. Those records ship with the package rather than being cleaned out of it.

rewrite one receipt's status	CAUGHT
reverse the whole ledger	CAUGHT
delete a receipt from the middle	CAUGHT
swap two adjacent receipts	CAUGHT
remove the head receipt	CAUGHT
drop the last receipt	CAUGHT, missed in v9.1
empty ledger	CAUGHT, missed in v9.1
tamper then recompute the chain	UNDETECTED, structural

What it is not

Not tamper-proof It is tamper-evident against a party who does not hold the ledger. It is not evidence against whoever does hold it. Closing that gap needs a signature or an external anchor, which is the obvious first joint project.
Not an agent framework, and not a competitor to one It is the layer that decides whether an action is permitted, and what record exists afterwards.
Not portable across machines, yet Runs are deterministic at a fixed path and reproduce byte-identical receipt hashes there. Receipts embed absolute paths, so the same run on another machine produces different hashes. Named in the brief.

Before you spend time on this

Nobody has run it but us. No production deployment. No external user. We would rather you hear that here than find it.

There is no patent and there will not be one. The work was assessed against prior art and a filing was judged not warranted. The reasoning was written down rather than left as a gap. This is a copyright, trade-secret and contract asset.

44 of 44 tests pass cold on Python 3.12, standard library only, no network, no dependencies, offline. The internal manifest is generated from the tree and refuses when the two disagree, because a hand-typed count is what went wrong twice.

Terms

Available for outright sale or license. Delivered as a technical packet with provenance, hashes, and the full defect history. Evaluation copies are confidential and carry no patent license.

hello@abernate.com