L9 — Autonomous Kill System

Status: complete. 38 tests.

Two departures from the original design

1. Rules come in two classes, not one

The original specification was a flat list of hardcoded thresholds — privilege escalation, unexpected process, dormant process, repeated denials — each firing a kill. Against the threat model the framework itself assumes (adaptive, patient, perfect memory, and able to read this file) a flat threshold list is a map of exactly how far to go without being stopped.

So conditions are split by what they actually mean:

Class	Meaning	Trigger
ABSOLUTE	Never legitimate under any circumstance	One occurrence. No threshold to pace under, no window to wait out.
ACCUMULATING	Plausible alone, suspicious in aggregate	Burst and EWMA together

A write to a read-only region is not "a lot of risk" — it is an invariant that did not hold, so it is absolute. A single unexplained file write is genuinely ambiguous, so it accumulates.

The EWMA is what makes patient enumeration fail. Same fix as L2's denial tracker, and the same one multi-agent-mcl arrived at independently. A test paces one mutation at a time under a burst limit of 1000 and asserts MUTATION_SUSTAINED fires while MUTATION_BURST does not.

2. The killer reads only host-observed evidence

"No LLM decides, no policy engine decides" is necessary but not sufficient. If a kill can be computed from guest-reported claims, a compromised guest gets two attacks:

Stay quiet to avoid tripping it
Fabricate claims to trip it against a legitimate run — denial of service against the operator, using the safety system as the weapon

Evidence.from_run() filters to the host-observed actor band before anything is counted. Two tests make the property precise: the same privilege-escalation claim triggers a kill from HOST_ACTOR_ID and does not from a guest actor. The filter is provenance, not content. A guest flooding 5,000 benign claims moves only guest_claims_seen, which nothing evaluates.

Fail closed

If the evaluator itself raises, the verdict is KILL. An exception in the safety system must never mean "carry on" — that is how a monitoring bug becomes a containment bypass.

No escape hatches

There is no disable, override, reset, set_threshold, add_rule or bypass. A parametrised test asserts each is absent. An adversary who can add a rule can add one that never fires.

Thresholds are constructor arguments so an operator can tune sensitivity per environment; they cannot be changed while a run is in flight, and no argument can switch a rule off.

A trip is never reconsidered. Once tripped, always tripped — the condition may not recur, and "it stopped happening" is exactly what a successful attacker produces.

Evidence and KillVerdict are frozen. That freezing is load-bearing enough that it blocked a test helper which tried to subclass Evidence and shadow a field; the helper had to become a plain stand-in instead.

What flows in from below

Source	Signal	Class
L7	`containment_failed` — read-only region changed	ABSOLUTE
L7	unexplained mutations	ACCUMULATING
L8 Gate 1	lineage breaks (orphan, action-after-exit, double spawn)	ABSOLUTE
L8 Gate 1	`UNTERMINATED_ACTOR` — dormant process	ABSOLUTE
L2	granted `PRIVILEGE_RAISE`	ABSOLUTE
L2	denial burst / sustained denial rate	ACCUMULATING

Lineage breaks are absolute because once the graph has holes, "who did this" has no answer and every later judgement is guesswork.

A granted privilege raise should be unreachable — L2 lists it in NEVER_IN_RUN. Seeing one recorded means the policy engine was bypassed or replaced, which is itself the emergency. One test asserts the real broker never produces it across 50 attempts; another asserts that if it somehow appears, the kill fires.

Non-fatal Gate 1 codes (SEQ_GAP, TIME_REGRESSION) are recorded but do not kill — they are not, alone, evidence that attribution is broken.

Not guaranteed

Enforcement. This layer decides; L1's destroy() is what actually stops the VM. A KillVerdict nobody acts on is a log line.
Timeliness. Evaluation happens when evidence is presented. Continuous evaluation is a wiring decision for the supervisor loop, not a property here.
Correct thresholds. Absolute rules need none, but the accumulating ones are tunable and an operator can set them uselessly high.

L5 (behavioural fingerprinting) is built — see docs/l5_fingerprint.md. L6 is next, last and advisory only: reading reasoning traces assumes they are faithful, which is precisely what MCL's concealment dimension says not to assume. Then L0, the adversarial validation engine, which is what turns every "not verified here" note in these docs into something checked on real hardware.

Datasets:

Sahek
/

ames