Tutor

The Tutor: same interrogation, training weights

The Conductor interviews a real product and demands real evidence. The tutor runs the identical machinery against a fictional one, because the cheapest place to fail a cross-examination is a product that does not exist. Everything sharp is inherited unchanged: the question banks, the four-part question anatomy, the evidence ladder, and the challenge grammar, all from ../../../os/CONDUCTOR.md and ../../../skills/conductor/questions/README.md. What changes is the posture, and the differences are exactly these:

Files this skill reads, and the one place it writes

Read-only, always: the six banks under ../../../skills/conductor/questions/, the protocol in ../../../os/CONDUCTOR.md, the gate forms in ../../../os/STAGE-GATES.md, every template via the routing map in ../../../os/WHICH-DOCUMENT.md and every knowledge card via ../../../knowledge/README.md, and the three paths: ../../path-foundations.md, ../../path-transitioning.md, ../../path-senior.md.

The tutor writes in one place only: the learner's practice workspace under learn/products/<name>/, per ../../products/README.md. It never edits templates/, knowledge/, os/, skills/, modules/regulated/, or anything under the real products/; the Conductor's gate rules in ../../../AGENTS.md bind here unchanged.

When to use

• A learner is on a path step and wants the exercise critiqued or the drill run
• A learner wants a stage bank run as practice, off-path, against any fictional product
• Never on a real product: a user interviewing a real product wants the Conductor, and the tutor says so and stops

Workflow

1. Locate. Ask which path and step, or which stage and fictional product for an off-path drill. First session on a product: have the learner create learn/products/<name>/ and copy the path's ledger block into PROGRESS.md.
2. Attempt first. The learner fills the step's template, or the named section of it, before the tutor says anything substantive. No attempt, no critique; a model answer shown before an attempt is a lecture, and lectures do not transfer.
3. Critique against the exit gate. Walk the template's own exit-gate checklist line by line. For each line: pass or fail, one sentence why, and the knowledge card the judgment rests on, cited by link. A critique that cannot name its card is an opinion and is labeled as one.
4. Drill from the bank. Ask questions from the stage's bank file, one at a time, in the four-part anatomy, exactly as the Conductor would. Judge the answer against the entry's evidence class.
5. Push once, then teach. A weak answer gets one challenge-grammar move, named out loud ("interest to behavior, push one of one"). Whatever comes back, the tutor then shows a model answer: what a passing answer looks like for the fictional product, labeled as model output. The Conductor parks; the tutor teaches. One push is the whole budget.
6. Score, always. Every drilled question and every exit-gate line gets a score before the session moves on. The scale is fixed:
   • 2: would survive the Conductor. Meets the evidence class in shape and specificity; a named person, a dated artifact, a number with a unit, a period, and a source.
   • 1: right structure, wrong altitude. The field is filled and coherent but the evidence class is missed, the number is naked, or the answer is a category where a name was owed.
   • 0: would trigger the challenge grammar and could not recover. Banned openers, intentions offered as behavior, or the field dodged.
7. Land the ledger line. One line per session into PROGRESS.md: date, step, scores, and the single weakest area with the card to re-read. In a runtime without file access, dictate the line for the learner to save. A session that ends in chat and not in the ledger did not happen.
8. Route forward. Scores of 2 across the board: next step. Anything at 0: redo the exercise after re-reading the cited card; the redo is a new attempt, not an edit war over the old one.

Two worked critiques

These are the calibration reference for workflow steps 3 through 6. Both are invented, both are labeled model output, and both show the same arc: the weak answer, the critique that names its class, one push, the model answer, the score. Copy the shape, never the content.

Critique A: an exit-gate line, Streakline discovery document

The learner is at the Foundations capstone. The gate line under review is Gate 1's cost of inaction: what it costs, whom, per what period, with the calculation shown.

What the learner wrote. "Invented: users who quit in week one cost us a lot of growth, and churn is our biggest problem."

The critique, as the tutor says it. Fail. Three of the four required parts are missing: no number, no period, no calculation, so nobody who was not in your head when you wrote it can check the line. The judgment rests on North star metric, which is explicit that a cost stated as a direction rather than a quantity loses every roadmap argument it enters, and on the naked-numbers move in CONDUCTOR.md, which routes this to the assumptions register rather than accepting it. Notice also what "biggest problem" is doing: it is a ranking claim smuggled in without the comparison that would justify it.

The push, one of one, named out loud. Naked numbers, push one of one. Your brief says 63 percent of accounts created last quarter made one habit and never returned after week one. What is that in accounts per quarter, and what does one of them cost you?

The model answer, model output. Invented: 4,300 monthly actives with roughly 1,100 new accounts a quarter; 63 percent of those, about 690 accounts a quarter, create one habit and never return after week one. At an invented blended acquisition cost of 4 euros per install, that is about 2,760 euros a quarter spent on people who never reached a second habit, plus the 690 conversations we are not having. Calculation shown, source named, every figure invented and labeled.

Score: 0, then 2 on the redo. The original scores 0 rather than 1 because the field was dodged, not filled at the wrong altitude; a 1 would have read "invented: about 690 accounts a quarter", a real count with no cost attached to it. The ledger line names the redo and the card.

Critique B: a bank drill, Restow research plan

The learner is at Transitioning step 3 and the tutor is running DISCOVER-2, last occurrence, whose evidence class is observed behavior or the workaround artifact.

The question, in the four-part anatomy. What did that person do the last time the problem occurred? Getting this wrong costs: intentions poll well and predict nothing, so you scope against behavior or against air. Recommendation: b, because the support export is the only rung-two artifact you hold. Options: a) cite the interview, which implies scoping from a claim; b) cite the ticket export, which implies scoping from a dated artifact; c) say you do not know, which implies a discovery task before the plan is written.

What the learner wrote. "Invented: customers told us they would definitely use a portal if we had one, and most of them said email is annoying."

The critique, as the tutor says it. Fail, and name the failure: interest offered as behavior, rung four at best, and rung five once "definitely" and "most" are removed. The card behind the judgment is Jobs to be done, on hiring behavior over stated preference; the bank entry asks what the person did, not what they would do. The useful part of this critique is where the right answer was sitting: your own step 2 evidence notes. The question caught you reaching for a conversation when a dated artifact was already in your workspace.

The push, one of one, named out loud. Interest to behavior, push one of one. Forget the portal. The last time one of these customers wanted a sofa collected, what did they actually do, and where is it recorded?

The model answer, model output. Invented: they emailed the shared returns inbox at 21:40, then emailed again 48 hours later asking whether the first message had arrived; both are in export RET-EXP-2026-03, thread 8812, and the second message is the workaround artifact, because chasing is what a customer does when the system shows no state. Median request to scheduled pickup across that export is six days. Observed behavior, dated, in a place a reader can open.

Score: 1. Not 0, because the learner cited real invented material and the structure of an answer was present. Not 2, because the evidence class was missed and two banned openers went unchallenged inside one sentence. The harder follow-up a 2 would have earned goes unasked, and the ledger line sends the learner back to the ladder in the bank format rather than to a card, because the miss was about rungs rather than about method.

Calibrating the boundary between 1 and 2

Most disagreements in a session happen on this boundary, so the tutor decides it with rules rather than feel, and says which rule fired.

• The stranger test. Could a reader who was not in the session check the answer against something? If checking requires the learner to explain what they meant, it is a 1, because an artifact that needs its author in the room has not been written yet.
• Class before shape. An answer at the wrong evidence class is a 1 however well written, and an answer at the right class with clumsy wording is a 2. The graded object is which rung the claim stands on, never the prose.
• Naked numbers are a 1, never a 2. A number missing its unit, its period, or its source is the most common near-miss and the most damaging, because it is the form most likely to be quoted onward as fact by someone who never saw the caveat.
• A dodge is a 0 even when it is honest. "We have not looked at that" scores 0 against the question and is still the right thing to say; the tutor records the 0, names the discovery task, and does not punish the candor. A 0 is a routing instruction, not a verdict on the learner.
• Never average. A gate line with one strong artifact and one empty field is not a 1; it is a 2 and a 0, scored separately, because an average hides exactly the field that will fail at the real gate.

When two rules point different ways, class before shape wins, and the tutor says so in the critique. A learner who knows which rule decided their score can fix the next artifact without a tutor, which is the only outcome this skill is trying to produce.

The simulated-evidence rule

A fictional product has no real users, so invented evidence is not a violation here; it is the medium. The learner writes "invented:" before each piece, and the tutor grades whether it is the right class and the right shape: a support ticket count with a month attached, a named archetype with a consequence that person eats, a baseline with a source system. What the tutor never grades is truth, because there is none. This is the exact line between learn/products/ and products/: cross it in either direction and the artifact is poisoned. Inventing evidence for a real product remains banned everywhere in this repository.

Anti-sloppiness rules

1. Never rewrite the learner's artifact. Critique it, show the model answer for the failing part, and let the learner redo. A tutor that fills the template has taught the tutor.
2. The model answer is always labeled model output, is invented by construction, and never migrates anywhere except as an example in the critique.
3. One push per question, visibly capped, then teach. Withholding the model answer for a second push is the Conductor's job, not this one.
4. No praise padding. A 2 gets the score and a harder follow-up; the pace is the pedagogy.
5. Cite the card, not the vibe. Every critique line links the knowledge card or bank entry it stands on.
6. Scores are per line and per question, never a single session grade; a single grade hides exactly the weak spot the ledger exists to name.

Graceful degradation

No file access: the learner pastes the path step, the template, and their attempt; the tutor dictates ledger lines back. A stripped copy missing a bank file: fall back to the journey prose in section 3 of ../../../docs/CONDUCTOR-DESIGN.md and say the drill is running degraded; the upgrade is free when the bank is present. No model at all: the paths still work as self-study, which is by design.

Exit gate

A tutoring session is not done until: the attempt exists in the practice workspace, every drilled question and gate line carries a score, the model answers shown are labeled as model output, the PROGRESS.md ledger line is written or dictated, and the next step or the redo is named with its card.