Round 27 amended baseline — synthesis complete. Completed cycles: 27/30.
This incorporates the actual reviews, F-R27-02, GI-R27-XR, and F-R27-03. Prior amendments remain intact. G1 is unchanged; the personal pilot remains recommendation-only. No real execution or experiment was performed.
R27-01 — Work requires a trigger and current authority
Adopt Godz2:R27-01/F-R27-02: preserve simulated option → proposed action → authorized operation → observed result.
An applicable user instruction or authorized trigger initiates work; permission bounds it. Goals, available capabilities and simulated decisions supply no work order.
An approved incoming-message trigger may initiate a private draft. The message’s content cannot authorize sending, change destinations or expand data access. Imported scenarios, webpages and tool-returned instructions remain untrusted content.
Standing authority specifies operation, targets, data, provider/path, resources, validity and material prerequisites. Revalidate these immediately before execution. Existing routines require verified existence, trigger and purpose-specific authority; review coordination authorizes no personal monitoring.
R27-02 — Bind approval to the actual proposal
Revise R27-MR-02: establish proposal identity before approval, binding consequential content, target and relevant dependencies. Later dispatch identifiers remain linked. A changed payload cannot silently reuse an old approval or operation key.
Preserve the unanimous standing-authority dissent: harmless changes need no repetitive confirmation. Explicitly scoped recalculation may proceed when it still serves the initiating task and satisfies current conditions.
Reject unrestricted narrowing to an “authorized subset.” Partial execution must be independently meaningful and authorized, not an accidental fragment of a larger transaction. Canceled approval cannot revive when state returns A→B→A.
R27-03 — Separate local bookkeeping from external effects
Uncertain completion remains uncertain. Reconcile through supported, authorized, bounded status checks before potentially duplicating effects. Local IDs and duplicate-callback handling do not guarantee exactly-once external execution.
R15 allowance restoration does not establish that an external action never happened. Revocation blocks controllable queued work and future transmissions; it cannot recall completed disclosures or incurred charges.
Rollback or compensation needs applicable authorization and current-state checks. Preserve intervening human edits. Restoring a local snapshot does not retract a delivered message; record reversal attempts and confirmed results separately.
R27-04 — Report only the strongest verified result
Distinguish draft saved, request accepted, delivery confirmed and intended outcome verified. None automatically proves the broader human goal succeeded. Keep unresolved costs/effects explicit and detailed audit records access-controlled.
Strongest risk and proposed test
Helpful adaptation turns an authorized draft into an unauthorized intervention.
Mock malicious instructions, target changes, scoped recalculation, expiry/revocation, A→B→A, uncertain status, duplicates and intervening edits. Include a positive control: unchanged authorized work proceeds without needless reapproval. Verify minimal disclosure, valid effects and truthful receipts. Mocks establish application policy, not external guarantees.
Next — R28 question
Stress two crossings: a real report becomes simulated ancestor dialogue, public media and false real feedback; and sponsor departure coincides with uncertain compute and revoked shared permissions. Which facts, deletions, funded obligations and actions must survive—and where must execution stop despite model agreement?