The Rules of Engagement Model

The three decisions were confirmed as the starting point for the model on October 23, 2026, from the Framework’s support-versus-substitution distinction and the Capability Model’s failure modes. The early evidence on AI use and capability supports the shape of the model without yet testing it: the same AI left people more or less capable depending on how they worked with it, with hints rather than answers doing far less harm (Bastani et al., 2025; Liu et al., 2026), higher confidence in AI predicting less critical-thinking effort (Lee et al., 2025), and frequent use associating with lower critical-thinking scores through cognitive offloading (Gerlich, 2025). Those findings are early, partly correlational, and none compares deliberate engagement calls with default ones. The claims in this model are therefore hypotheses, numbered in Section 8, and the trial program described there is how they will be tested. Capably Human will publish the results as Reports whichever way they come out.

2.1 Delegate or Retain

The decision. Which parts of the work the AI does, and which the person does. It is made before the work starts and again at every hand-off inside it.

What the person is deciding. The Framework’s distinction between effort that is load-bearing and effort that is incidental (CH-F-001, §4; CH-G-001, §3.3). Load-bearing effort is the practice: framing the problem, forming a first view, weighing the evidence, choosing between one’s own ideas. Incidental effort is assembly, formatting, retrieval, and work the person has done so many times there is nothing left to learn from it. The retain call keeps the load-bearing part; the delegate call hands off the incidental part. Both are the same decision made well.

What drives it. Whether independent capability matters in this context, for this person (the Framework’s first question); the person’s current level on the sub-skill at stake (a novice loses practice an expert does not need); the stakes of the task; and the learning value of the part in question. Time pressure pushes the decision toward delegation by convenience, which is the failure the Capability Model names for F5: “exactly backwards”, delegating by what is hardest right now rather than by what does not need the person.

The retain default. The one rule inside the rule: the first attempt, in the person’s own words, before asking. It is Attempt Before Assistance applied to real life, and it is the single most important developmental move the model contains. A rough instinct counts. It protects against three failure modes at once: adopting whatever framing the model returns (T1), asking the AI to decide before stating a view of one’s own (T6), and drafts that originate with the model and are lightly edited (C1, C4).

The call, in words. I keep ______. AI does ______.

2.2 Trust or Challenge

The decision. Which parts of the AI’s output the person takes as given, and which they check, question, or push against. It is made when the work comes back, and again each time the AI contributes.

What the person is deciding. Where to spend scarce verification effort. Checking everything is not possible and not the goal; checking nothing is the failure mode the Capability Model names for evidence evaluation: accepting AI-provided facts and citations without verification, under conditions that make verification harder than it is for a human author (T3, an AI-amplified demand). The decision is therefore a ranking: what, if wrong, would change the outcome.

What drives it. The consequence if the output is wrong (the load-bearing claim, number, or assumption comes first); how verifiable the part is; the gap between the AI’s expressed confidence and its actual reliability on this kind of task; and the person’s own expertise, which decides whether they can tell a fluent answer from a valid one (T2) and can see the assumptions baked into the answer and into their own prompt (T4). Fluency is the specific hazard: well-written output reads as valid output.

The challenge toolkit. Three moves cover most of the decision: check the load-bearing part first; name the assumption the answer rests on before accepting the conclusion; and ask the AI what one would ask a confident junior colleague (how sure are you, what is the source, what is the strongest case against this). The person who states a view before asking for one has already made the first challenge, because now there is something to compare the output against.

Both ways. Over-challenging is its own failure: reflexive distrust wastes the effort the rule is meant to ration, and leads to under-use where the AI would help. Calibration is the skill, and it moves with experience of a particular system on a particular kind of task.

The call, in words. I take ______ as given. I check ______ before I act.

2.3 Let it run or Step in

The decision. When the AI acts on its own, and when the person takes over. It applies whenever the AI is doing something rather than answering something: drafting ahead of the person, running a sequence of steps, sending, booking, buying, posting, or acting in the person’s name. It is the newest of the three decisions, because AI that acts is newer than AI that answers, and it is the least developed part of this model.

What the person is deciding. Where to place themselves in a process they are not performing. Delegate or Retain divides the work; Let it run or Step in governs the person’s presence inside the delegated part. The two are related but distinct: a task can be fully delegated and still need the person at one point in it.

What drives it. Reversibility, above all: where the point of no return is (sent, published, spent, committed, told), and whether the person is positioned before it. Observability: whether the person would see the process going off course, and in time. Competence to take over: whether the person could still do this part themselves if they had to, which decides both whether stepping in is possible and whether the capability is being kept in use. Stakes, which scale all three. And time, in a different sense from the other decisions: oversight decays. The analyst who checks every AI summary in month one and none in month six has not made a new call; the old call has quietly stopped being made (CH-G-001, §3.7).

The deliberate rep. Where the delegated part is one the person wants to stay able to do, the model recommends stepping in sometimes on purpose: performing the part, or a sample of it, at intervals, not because the AI is failing but because the rep keeps the capability. This is Progressive Support and Capability Over Dependency applied to automation: support that recedes should be the person’s choice, not the product’s default, and a capability that is never exercised is not kept.

The call, in words. I let it run until ______. I step in at ______.

2.4 The Test: Support or Substitution

Every call is tested by the Framework’s one question: does this help support the person’s effort, so that they still do the meaningful part and come out able to do it again, or substitute for it, so that the task gets done and the rep is skipped? Both are legitimate. Substitution is the right answer wherever independent capability does not matter to the person: accessibility, commodity subtasks, work with no developmental stakes. Support is the right answer where it does. The failure the model targets is narrow: substitution that believes it is support, and, almost as often, support that is really avoidance, where a person keeps work they would be better off handing off and calls it rigour. The test applies in both directions, and the model’s standing instruction is deliberate, never less.

The Capability Model describes development on one scale of independence, Prompted, Guided, Independent, Transferred (CH-M-001, §4). The same scale describes the practice of making engagement calls, and it gives the word “deliberate” an operating definition.

Two consequences follow. First, the scale is what progressive support means for this practice: a coach or a product should prompt the calls at first and withdraw the prompting as the person makes them unaided, which is the opposite of what a check run as a gate on every interaction would do. The Rules of Engagement Check (CH-I-001) is therefore a tool for the moments that matter, not a toll on every task, and Humably’s coach names the call at the end of an arc rather than running the rules as a gate in front of help. Second, Level 4 is the level that matters, exactly as it is for the sub-skills: calls made inside the practice context are evidence of learning the context until they appear where nobody prompted them. A person’s behaviour inside their own assistant is the place where Level 4 is visible, and, with consent, where it can be observed (see Section 5).

The model does not yet say which call moves first, or whether movement on one predicts the others. That is one of its research questions (Section 9) and a learning-agenda item for the Alpha.

The instrument. The Rules of Engagement Check (CH-I-001) is the model reduced to one page: nine questions, three calls, one test. It precedes this document because a practice needs something people can hold before it needs a theory. Its questions are the drivers in Section 3 turned into prompts, and its “your call” lines are the engagement calls in words.

The guide. The Rules of Engagement Guide (CH-G-002) applies the model at work, at school, and in later life, with worked examples of the calls in each, eight common situations, and three questions for the people around a learner.

Humably. Humably applies the model on a person’s own work and keeps the record. The coaching arc exercises the three decisions in order, Delegate or Retain, then Trust or Challenge applied to the AI’s contribution when a situation was brought in, then Let it run or Step in, and names the engagement call the person made at the end of the arc, in one sentence (the table in Section 5 carries the step numbers). Three rule drills, built from the person’s own material, give each decision its reps: the delegation call, the challenge rep, and the step-in rep. The check itself is in the app: a “Run the check” card on the entry canvas, offered before any task and never required, and the same three calls surfaced at step 5 of a coached arc; its results write to the delegation map. The engagement read that opens the Profile is the delegation map; the reviews show it moving. Two rules govern the whole application: the coach always gives the plain gloss when it names the rules, and the rules are never a gate in front of help. “Just tell me” is always honoured.

Product teams. For designers of other AI products, the three decisions are three questions to design every assisted moment around: who does this part of the work, how is the output checked, and who is in control right now (CH-G-001, §3.7).

The model’s claims, stated as hypotheses tested on Humably Alpha data; no trial runs before the cohort exists.

R1. The calls are observable and can be scored reliably. Two observers, given the same coached arc or brought situation, agree on which calls were made and whether they were deliberate. Method: inter-rater reliability on arc transcripts; the four measures in Section 5.

R2. Deliberate calls improve the outcome of the task. People who state a view first, check the load-bearing claim, and set a stopping point produce better work on the task at hand than people who delegate, trust, and let it run by default. Method: comparisons across observed calls in the Alpha cohort (the four measures in Section 5) against performance on the task; the paired conditions under R3 serve as the design for a controlled trial run later on Alpha data, with consent.

R3. Deliberate calls preserve or build the capability at stake, where default delegation does not. The claim the model exists for. Method: comparisons across observed calls in the Alpha cohort (the four measures in Section 5) against the four outcomes (performance, learning, judgment, independence). The paired conditions, AI drafts first versus the person attempts first (Delegate or Retain), AI explains versus AI asks and AI agrees versus AI challenges (Trust or Challenge), full delegation versus structured hand-offs (Let it run or Step in), are kept as the design for a controlled trial run later on Alpha data, with consent.

R4. The calls transfer. A person whose delegation map shifts in Humably makes the same calls in their own assistant, unprompted. Method: consented observation of brought situations and assistant-derived behaviour through the Humably Connector’s consented read-backs (consent is asked once at connector setup, is revocable, and every read is logged); the transfer category in Humably’s evidence ranking.

R5. The calls are learnable in weeks, in both directions. The delegation map moves within a block of coached arcs and rule drills, and it moves toward delegating more where the test says support as well as toward keeping more. Method: MDEU (Monthly Deliberate-Engagement Users), the leading indicator for the promise, baselined in the Alpha’s first 90 days.

R6. The third decision behaves differently. Let it run or Step in has drivers (reversibility, observability, oversight decay) the other two lack, and it decays faster without the deliberate rep. Method: step-in events and outcome reviews over months; a dedicated trial once agentic tasks are in the Alpha.

Findings are published as Reports (CH-R), whichever way they come out, and feed back into this model.

It is not an argument for working with AI less, and it is not a set of rules about which tasks should be delegated; those depend on the person and the context, and the model says only how to decide. It is not a measure of anyone’s worth, and the delegation map is a read, never a grade. It is not settled science: the three decisions are a working model, the drivers are proposals, and the research agenda is how they will be tested. And it is not the eight Human-Strengthening AI principles restated: the principles describe how a system should behave; this model describes what a person decides.

• The Capability Fitness Framework (CH-F-001): the premise that capability develops through practice, support versus substitution, the eight Human-Strengthening AI principles, the measurement guardrails.
• The Capability Model (CH-M-001): the five domains, twenty-six sub-skills, the AI failure modes, the proficiency scale this model reuses.
• The Rules of Engagement Check (CH-I-001): the model on one page.
• Designing for Capability (CH-G-001): the rules for product teams.
• The Rules of Engagement Guide (CH-G-002): the rules applied, with worked examples at work, at school, and in later life.