CAIN-42 Evolution 16 of 42: the durable mission runtime (2026-10-12)

AN AGENT'S MISSION NOW SURVIVES A CRASH, NEVER REPEATS AN ACTION IT CANNOT PROVE WAS NOT DONE, AND LEAVES A JOURNAL ANYONE CAN CHECK WITHOUT CAIN'S CODE.

Project Sovereign Frontier, evolution 16 of 42. Not to be confused with the earlier "Evolution 16" world-state work of September 2026.

What it does

Every request to a running mission and every decision about it is written to a signed, hash-chained journal on disk before it takes effect. After a crash the runtime rebuilds itself from the journal and refuses to start if any rebuilt decision differs from the recorded one.
Before a tool is called, CAIN-42 re-checks that the agent still holds its delegated authority, that the authorization has not expired or been revoked, and that the mission was not cancelled. The intent is recorded first.
If the tool times out or the process dies, the action is not retried blindly. It waits until a pinned, independent observer signs whether the effect happened.
A running mission's contract can be changed only as a signed chain of versions: no old version can come back, two simultaneous edits cannot both win, and anything that widens what the mission may reach, spend or run until needs the human principal's signature on exactly that change.
A compensation (an undo) is a new action with its own authorization, and counts only when an observer confirms it. Irreversible actions are never reported as rolled back.

What was measured

1,230 of 1,230 scenarios executed, 0 inconclusive. 0 of 690 fault and attack scenarios ended badly on the new runtime. On the 420 cases where the previous code has a comparable mechanism, it ended badly in 330.
Crash after every journal entry of a complete mission: 71 of 71 crash points recovered and completed, 0 duplicate writes, 0 missing effects.
Each of the runtime's 25 safety checks was removed in turn: 25 of 25 removals were caught by a scenario.
Every sequence of up to 4 events was run on the real runtime: 16,104 sequences, 0 invariant violations.
The clean-room verifier returned the expected verdict in 570 of 570 cases, including forged, truncated, re-signed and incomplete journals.
90 of 90 ordinary missions completed, 0 refused a legitimate step.
Two defects found in existing CAIN-42 code and fixed: a workflow step that ran twice after a timeout or crash (10 of 10 runs before, 0 after) and a withdrawn fact that stayed trusted (10 of 10 before, 0 after).

Check it yourself

Download verify_mission_journal.py (Python 3 and the cryptography package only; no CAIN imports) and the sample mission in sample/, then run:

python3 verify_mission_journal.py sample/mission.journal.jsonl --trust sample/trust.public.json --head sample/head.json --require-final

Expected: "verdict": "VERIFIED" with 31 distinct checks performed, exit code 0. Change one byte of the journal and it fails. Leave out --head and it says INCOMPLETE, because a deleted tail cannot be ruled out without a head kept by someone else.

Status

Label	Value
IMPLEMENTED	YES
TESTED (local simulator)	YES
INDEPENDENTLY_REPRODUCED	NO
EXTERNALLY_REVIEWED	NO
WIRED_TO_MCPGATE	NO
PRODUCTION_DEPLOYED	NO
PRODUCTION_RELEASE	NOT APPROVED

Limits, stated plainly

Tested in a deterministic local simulator: no language model, no real tool, all keys in one process. Not connected to MCPGate or the hosted gateway yet.
A pinned observer that signs something false is not detected (30 of 30 such cases). A signature proves who signed, not that the statement is true.
Durability costs speed: one decision takes 5.0007 ms at the median with every entry forced to disk, against 0.1771 ms in memory (150.1 vs 5391.3 decisions per second, ext4).
Written and tested by the CAIN-42 project itself. Nobody outside has reproduced it or reviewed it.

IP register

Register v47 adds 6 mechanism families. Standalone net IP base, BASE scenario: 16,404,530 before, 16,457,330 after, rising at each implementation step: 16,415,090 → 16,425,650 → 16,436,210 → 16,443,250 → 16,450,290 → 16,457,330. FIRME AI company value (v42), BASE: 18,939,330 investor / 19,939,330 strategic. Internal analytical scenario, NOT an appraisal, NOT a financing valuation, NOT a guarantee.

Files

Everything Evolution 16 produced is here, including the mechanism specifications (mission_contract_spec.md, recovery_execution_spec.md), the architecture and the prior-art research. Source commit 465ef253. Implementation source is not published; SHA-256 commitments to it are in EVOLUTION16_MANIFEST.json.