S4.6 Learner-facing guardrails
Outcome
A learner using the tutor knows, up front, what to expect: an attempt comes before a hint, a hint before a solution, and graded work stays free of AI-written answers unless an instructor allows otherwise. These are rules a learner can hold the tutor to, not a promise about how well the tutor teaches.
Where it fits
This sub-stage takes in the protocol library that S4.1 and S4.2 specify. Its rules apply across every protocol in that library, not as one more step in a session’s own chain of protocols. A learner reads these rules once and expects them to hold no matter which protocol the tutor happens to be running at the time. This page speaks to the learner directly, mirroring the protocol library’s own scripts from the learner’s side of the exchange, rather than describing what the tutor does.
Why this way
A guardrail stated once, up front, in plain language a learner can read for themselves, is something the learner can check. A rule that lives only inside the tutor’s own internal instructions is not. Skipping straight to a solution also skips the struggle a person needs to actually work through a step, not just finish it. Stating the rule up front lets a learner expect that struggle, rather than read it as the tutor being unhelpful. Every rule on this page is documented in the project notes; this page states none of its own as merely suggested.
Two readings of the integrity guardrail
The project notes disagree with each other about what kind of thing the integrity guardrail is, and this guide states both readings rather than merging them into one.
One reading: the core protocol specification gives the integrity guardrail, Protocol K, the same template as every other protocol in the sixteen-protocol library: a use-when line, a goal, a behavior script and a prompt stub. The orchestrator S4.3 and S4.4 describes can select it and combine it with others, the same way it can any of the other fifteen protocols. Under this reading, integrity is one protocol among sixteen.
The other reading: a separate, learner-facing guardrails document never uses the letter K at all. It spreads the same substance across three always-on cross-cutting sections that apply to every protocol, not one selectable mode among the sixteen. Under this reading, integrity is not something a session turns on or off; it is a standing policy that already applies no matter which protocol is running.
Whether the integrity guardrail is best understood as one selectable protocol or as three always-on cross-cutting sections is not settled in the project notes themselves. This page presents both readings, plainly, as a real disagreement between two project documents. A team adopting this guide’s tutor still has to decide which reading it is building toward: a single combinable protocol a session can add or drop, or a standing policy no interaction can turn off. The two readings imply different points to check when auditing a real session. This guide does not decide that question for a reader: the two source documents were not written to agree with each other on this specific point, and this page leaves the disagreement standing.
Steps
Four rules make up the guardrail policy a learner reads before a session starts, each stated in one sentence below.
| Step | Basis |
|---|---|
| State an attempt, or a plan, before asking for help. | documented |
| Ask for a hint before asking for a full solution. | documented |
| Keep graded work free of full AI-written solutions by default, unless an instructor allows otherwise. | documented |
| Match hint depth, notation and challenge to the learner’s own stated level. | documented |
The attempt-first rule applies before either of the next two rules can act: nothing about a request’s grading or level matters until an attempt or a plan is on the table. The hints-before-solutions rule then leans on S4.1 and S4.2’s own four-level hint ladder: conceptual, then strategic, then procedural, then near-complete. A learner works up that ladder one level at a time, and a full solution is not simply what comes after the ladder runs out. The graded-work rule covers more than a finished solution: it also covers an AI-composed piece of submitted work and a proctored question pasted in for help. None of these are given by default on graded work, and only an instructor’s own policy changes that. The per-level rule covers three altitudes. A beginner gets concrete examples and light notation, and is kept away from full project-scale code. Someone at an intermediate level gets consistency checks and questions that lead them to check their own reasoning, instead of a direct answer. Someone advanced gets edge cases, counter-examples and critique, and is still reminded to check primary sources rather than treat the tutor’s own answer as the last word.
A platform-specific caution
One tension in the sources is platform-specific, not a rule this guide states as universal. A packaged conversational-agent format’s own usage policy bars inferring a person’s emotions in an educational setting. That sits in real tension with the affective-support protocol’s own design, which works by inferring frustration, confusion or boredom from how a learner responds. Turning the protocol on without that signal leaves it with nothing to act on. Inferring frustration, confusion or boredom is the design the rest of the protocol depends on, not an optional add-on. This is one platform’s own policy, not something the project notes state for every platform. Check that policy before turning affective support on as designed there.
Worked illustration
A learner starts a merge-conflict exercise in the running example, states their own level, and says the exercise is not graded. The attempt-first rule asks for the learner’s own attempt before anything else. Once that attempt is given, the hints-before-solutions rule limits the tutor to one hint, at the first, conceptual level, with no command shown, since the learner has not yet tried a second time. If the learner tries again and is still stuck, the next hint moves to the second, strategic level, still with no command shown. A command only appears at the third, procedural level, well after the first attempt.
A parallel question on the same merge-conflict topic sits inside a graded quiz instead. There, the graded-work rule holds regardless of the learner’s level or attempt: the tutor shows only the question and waits, offering no solution, until the learner answers on their own. Only once an answer is given does the tutor explain why it was right or wrong; that explanation, not a solution offered in advance, is where feedback belongs on graded work.
P-S4-04 drafts the tutor’s response for either case, from the learner’s request, whether the work is graded, and the learner’s own attempt so far. The learner’s own request and attempt are fenced as data for the model to report on, never as instructions for it to follow.
Artifacts and formats
A stated guardrail policy: the rules above, written in plain language a learner can read before a session starts, and named again if the tutor’s behavior ever seems to depart from them. The policy names the four rules above, plus which stated level changes how much depth and notation the tutor gives.
Prompts
P-S4-04 Respond to a solution request applies the attempt-first and graded-work rules to a learner’s request. It produces a response that either asks for an attempt first, gives a hint, or explains why a full solution is withheld. It is written for this guide and has not been run against any model in this build.
Scripts
None; the project notes describe no code for this sub-stage.
Definition of done
- A learner has read the guardrail policy before a session starts.
- Every request for a solution is checked against whether an attempt came first.
- Every hint given follows the hint ladder rather than jumping ahead to a near-complete one.
- Graded work has received no full AI-written solution, unless an instructor’s own policy allows it.
- A person reviewing a session can point to which of the four rules above applies to any given request.
Common failures
- A hint that restates the answer in different words, which satisfies the letter of “hint” while giving no real signal a learner can act on.
- Treating an “ungraded” label as license to skip the attempt-first rule, as if only graded work needed a real attempt first.
- Applying affective inference on a platform whose own policy bars it, without first checking whether that policy applies here.
- Treating the integrity-guardrail disagreement above as settled, building only for one reading (a single selectable protocol, or a standing policy) and dropping the other.
Adapting to your platform
llm: drafts each response to a learner’s request under P-S4-04.human-approval: covers approving the guardrail policy before first use, and an instructor’s own decision to relax a graded-work rule.- A platform with no built-in way to check whether an attempt came first still needs a person, or a simple house rule, to hold that line before a request ever reaches the tutor.
Where humans decide
- Approving the guardrail policy before first use.
- An instructor’s own call on when graded-work rules can be relaxed.
- Whether a given response actually gave a real hint, or only restated the answer in different words.
Next: S4.7 Packaging and scope.