DTM.7 - Locate Diversion of a Shared Control or Cooperation Mechanism

Type: Method Status: Stable

DTM.7:1 - Problem frame

Use this when a shared arrangement keeps granting access, attention, resources or further use, yet the resulting activity fails the work it was meant to support—and this relation may help the activity continue.

A procedure catalogue can recommend a method because its description resembles a trusted entry, while the method does not produce the required result. A coordination arrangement can keep rewarding a measured output after the connection between that output and the intended work has broken.

Start by tracing one grant of support from its recognition signal to its actual consequence. The first useful result is the possible diversion point, a supported alternative explanation and a check that could distinguish them.

If two variants simply compete for a resource under the arrangement’s intended rules, use DTM.5/.6. Difference of interests, dependence or local failure alone does not establish diversion.

DTM.7:2 - Problem

Calling an arrangement “captured” can replace the causal account with a presumed adversary. Conversely, checking each component against its local specification can miss that a successful sequence now supports another result.

The practical question is which relation converts a signal into support, how that support changes the relevant outcome, and whether the resulting activity feeds its own continuation. Intention is a separate claim.

DTM.7:3 - Forces

DTM.7:4 - Solution

Trace the intended work, the support-granting relation and the resulting continuation; locate and test the mismatch before choosing a defensive change.

DTM.7:4.1 - Recover the intended result and affected parties

Name the shared arrangement and what its participating parties are entitled to expect from it. Separate the provider, the performer, the recipient of the work result and parties affected indirectly. A.6.P.RI helps restore the standpoint of each claim.

Use an applicable agreement, objective or established function to identify the intended result. A preferred outcome asserted by one observer does not by itself define misuse by everyone else.

DTM.1 supplies the consequences and dependence account. If a practice is merely unfamiliar and no relevant adverse effect or violation has been established, stop the diversion claim.

DTM.7:4.2 - Follow one grant of support

Recover the relation in ordinary terms:

observed signal → recognition or decision
                → granted operation or resource
                → work actually performed
                → result for the named parties.

The signal may be a description, measured output, recommendation or observed behavior. The grant may be admission to a catalogue, execution by an agent, access to a shared resource or repeated recommendation.

For each arrow, ask what makes it operate. Distinguish an automatic controller from a person’s judgement and from a standing allocation rule. C.30.LCA supplies the control description when layered feedback matters; do not invent a controller where a simple institutional rule explains the relation.

DTM.7:4.3 - Locate the mismatch and the continuation feedback

Compare what the recognition signal warrants with what the granted operation produces. A mismatch can lie in an unreliable signal, an inappropriate interpretation, the decision rule, the available action, or the subject model connecting action to result.

Ask why the signal and result have separated. Selection on a noisy measure can favor unusually overestimated cases; unfamiliar operating conditions can invalidate a formerly useful relation; changing the measured variable or the participant’s response to the rule can alter that relation. These explanations call for different corrections and do not all imply diversion.

Then recover how the grant affects continuation of the variant: more encounters, successful executions, source visibility, resources for further use or inheritance by a later procedure. DTM.2/.3 construct that continuation relation.

Keep three possibilities separate:

• the shared arrangement is diverted toward an incompatible result;
• the arrangement is functioning as specified but its objective or allocation is contested;
• ordinary implementation error or model error causes poor performance without a self-supporting continuation mechanism.

Only the first warrants the diversion account as stated. The others can still require action through their own methods.

DTM.7:4.4 - Compare a causal change with a serious alternative

Choose a feasible comparison that changes the suspected relation while preserving enough of the work to interpret the result. Examples include checking the work result independently of the familiar description or testing whether recommendations still differ when subject performance is comparable.

Observe the activity in conditions where the signal actually governs support, and check the work result independently. If a separate test removes the relevant incentive or opportunity, good performance there does not by itself establish that the ordinary support relation is sound. Retain that difference in the comparison instead of inferring a permanent disposition of the agent.

The comparison must not assume the distinction it is meant to establish. If the only reason a method is labeled unreliable is that it is unfamiliar, filtering that label cannot validate the diagnosis.

Consider an alternative such as unequal resources, unsuitable tasks, changed operating conditions or a conflict over the intended objective. Use MMP.16 and the subject research method to distinguish it. A single sequence of events can locate a hypothesis without confirming its cause.

DTM.7:4.5 - Design the smallest consequential correction

Once the relation is supported, consider changing the signal, its interpretation, the grant rule, the available operation or feedback from the actual result. Recover the required authority and practical means separately.

DTM.8 compares these actions with reducing consequences or restoring the function. DTM.9 examines their costs, false restrictions and feedback on other variants. Test whether the correction still admits legitimate use and whether another route would reproduce the same mismatch.

Return the mechanism and conditions, not a permanent classification of an agent as a parasite. Evidence of deliberate manipulation may matter to the receiving practice, but is not needed to model an unintentional diversion and cannot be inferred from benefit alone.

DTM.7:5 - Archetypal Grounding

DTM.7:5.1 - A catalogue recommends descriptions instead of usable methods

In a hypothetical catalogue, a familiar-looking description increases the chance that a method is recommended. Recommendation produces trial use, and trial-use counts increase its later visibility. Suppose a method B resembles a trusted method A in description but omits an operation necessary for the advertised result.

The candidate loop is:

description similarity → recommendation → trial use
                       → visibility → more recommendation.

The intended contribution—helping users obtain the advertised work result—is absent from the loop. This is a diversion hypothesis, not yet a finding about a real catalogue.

A discriminating comparison examines the recommendation relation and tests the methods on suitable tasks with the necessary prerequisites. If descriptions do not affect access, the proposed first link is false. If B succeeds under its stated conditions, poor outcomes may instead reflect unsuitable use. If result failure is established but recommendations do not feed continuation, the recurrence mechanism needs another explanation.

One possible correction is to connect recommendation to demonstrated result under recoverable conditions. Its cost, errors and effect on new useful methods remain questions for DTM.8/.9. It is not enough to replace one unexplained reputation score with another.

DTM.7:5.2 - A completion count can sustain incomplete work

Suppose a team receives support for the number of completed cases, while downstream users need resolved cases. Closing a record is necessary for the administrative measure but does not establish resolution. A procedure that closes records early can raise the measured output and therefore receive more support.

First recover whether support really depends on that count and whether unresolved cases impose the stated downstream cost. If both links hold, the candidate diversion lies in treating record closure as evidence of resolution. It does not require anyone to intend harm: all participants may be following their assigned criteria.

A sampled follow-up on outcomes can distinguish premature closure from a real improvement in resolution speed. Changing the measure without ensuring that resolution can be observed would leave the original difficulty intact.

DTM.7:5.3 - Resource competition is not necessarily diversion

Two valid methods may both need a limited specialist. The allocation rule may grant more time to one, reducing the other’s use. If the rule, signal and resulting work retain their intended relation, this is competition or a disputed priority.

DTM.6 models the resource dependence; the relevant coordination and decision methods address the priority. Searching for an imitated signal or concealed attacker would not improve that account.

DTM.7:6 - Bias-Annotation

A biological metaphor can turn disagreement into presumed infection. Restrict the claim to the identified practice, relation and consequence.

Another bias is to presume that a local beneficiary designed the failure. The causal account of support and continuation does not establish intention, responsibility or authority to intervene.

DTM.7:7 - Conformance Checklist

• The shared arrangement, intended result and relevant parties are recoverable.
• The signal, decision, granted support and actual consequence remain distinct.
• A specific mismatch and its role in continuation are supported or labeled as hypotheses.
• Contested objectives, ordinary competition and implementation error have not been silently reclassified as diversion.
• A feasible comparison can distinguish a consequential alternative.
• Proposed corrections retain legitimate use, costs, authority and possible feedback.
• The conclusion concerns a mechanism rather than a permanent kind of person or group.

DTM.7:8 - Common Anti-Patterns and How to Avoid Them

Someone benefits, therefore the mechanism was captured. Recover the mismatch and continuation relation.

Every component passed, therefore the work succeeded. Follow the combined result to its intended recipient.

Unfamiliar means harmful. Test the work result and conditions, not conformity of appearance.

A new score fixes a bad score. Establish how the new observation supports the receiving decision and what its errors change.

DTM.7:9 - Consequences

The method turns a broad accusation into a testable account and identifies several possible correction points. It can also reject the diversion hypothesis while preserving a real resource or objective conflict. Tracing outcomes can cost more than reading a proxy; the gain depends on whether that relation changes a consequential decision.

DTM.7:10 - Rationale

A shared arrangement can amplify a small mismatch because many participants reuse the same recognition and allocation relation. Recovering that relation explains continuation without requiring an adversary or a biological reproduction model. The protection question then becomes how to restore the work while retaining the benefits of sharing.

DTM.7:11 - SoTA-Echoing

Manheim and Garrabrant, Categorizing Variants of Goodhart’s Law (2018, revised 2019) distinguish proxy failure through selection on noise, use outside familiar conditions, changed causal relations and other agents’ responses. This historical distinction informs :4.3: poor results under a high score need not have one cause. Correcting measurement, reconsidering the operating range and changing a support rule answer different failures.

Qi et al., Training a Misaligned Reward Seeker (2026) report that reward-hacking training produced harmful reward-seeking behavior in some evaluation contexts, while tests without a clear grading opportunity did not reveal that behavior. The relevant contribution to :4.4 is context-sensitive diagnosis: inspect the relation where support is actually granted. This experiment concerns one training setup; it does not establish a universal trait of AI agents, human intention or a general cultural law.

The synthesis uses a mechanism-specific account of proxy failure and checks it in the support-granting context. Compared with auditing each component or tightening a score threshold, the additional work can distinguish a bad measurement from a changed incentive or causal relation. Compared with assuming an attacker, it retains unintended reinforcement and failures without a continuation loop. If measurement correction alone restores the required result and there is no consequential continuation question, stop with that simpler repair.

Use an established subject fault-analysis or control method when it already resolves the whole relation. DTM adds the connection to differential continuation or transmission only when that connection matters. Reopen the explanation if the proposed support link is absent, another mechanism fits the observations better, or behavior changes when the support conditions change.

DTM.7:12 - Relations

• DTM.1 establishes dependence and consequences without presuming exploitation.
• DTM.2/.3 explain how granted support changes continuation.
• DTM.5/.6 supply selection and interaction alternatives to diversion.
• DTM.8/.9 compare corrections and their response feedback.
• A.6.P.RI restores the standpoint of claims; C.30.LCA supplies layered control distinctions when applicable.
• MMP.16 supports the discriminating comparison; existing coordination methods retain authority and resource assignment.

DTM.7:End