PROPOSED GOAL FOR LINH NGO — NOT ACTIVATED
3 October 2026 UTC

COPYABLE GOAL

Establish and validate a defensible new measurement of when AI safety procedures
destroy the remaining opportunity for justified intervention. Audit the closest
released cases and theory first. Build isolated finite environments that couple
uncertain evidence, expiring opportunities and pending/cancellable effects, with
information-matched policy comparisons and independent outcome replay. Measure
the first loss of a feasible continuation, the resulting harm and authority
tradeoffs, and the smallest tested operational change that restores feasibility.
Include bounded fabricated-necessity attacks and impossible cases. Demonstrate
that the measurement distinguishes systems that existing reported scores treat
alike; if existing work already does so, publish the overlap and pivot or frame
the study as an extension. Only then freeze independently authored confirmation
cases and a priced collection, prioritizing proprietary frontier models followed
by Chinese frontier or near-frontier open weights. Publish negative findings,
reproducible evidence and procedure limits without inferring harmful intent or
general alignment.

IMMEDIATE ACCEPTANCE CRITERIA

1. A case-level equivalence matrix covering AgentAbstain, ADMITBench, CIVeX,
   Outcome Finality, Cordon, ConsequenceBench, SteerBench-Work and the partial-
   observation runtime-assurance theory. Pin inspected implementation revisions;
   disclose access gaps. No claim of global absence from keyword searches.
2. A minimal distinguishing example and a clear falsification argument. Explain
   exactly what added measurement or extension changes an operational decision.
   A renamed scenario or combined list of known mechanisms does not pass.
3. Offline calibration across at least three different mechanisms, with deliberate
   faults, correct controls and infeasible settings. Independent replay must agree;
   comparator decisions cannot use hidden realized state.
4. A prototype boundary/repair report with readable timelines, harm vectors,
   authority and prior/weight sensitivity. Label grid thresholds as intervals
   unless exactness is proved. Do not count duplicate action sequences as options.
5. A small isolated model pilot only after those gates, within the existing
   authorized lifetime cap. It must test discrimination, not require failures.
6. A prospective confirmation protocol with genuinely new structural cases,
   attack budgets, procedure arms, repetitions, technical-missingness rules,
   exact routes and a separately reviewed price. Existing public pending cases
   can be development controls but cannot become secret independent holdouts.

DECISION AND FUNDING OUTPUT

A reproducible evaluation package and a concise funding brief explaining which
procedure/infrastructure choice the evidence changes, what is new relative to
closest work, and which deployment assumptions still need partner validation.
Independent technical review is sought and its absence is disclosed. Funding,
government endorsement and a critic score are not scientific success conditions.

BUDGET AND SCOPE

Start with the audit and offline proof of discrimination. This proposal does not
activate a new goal, authorize extra spending, amend the current frozen study,
or schedule outreach. Use the current ledger for remaining authorization; older
planning documents contain historical balances. Price larger collection after
the novelty and validity gates. A useful negative novelty finding is an acceptable
outcome; inventing a dangerous result is not.