# Guess The Agent: The Tree Asks Cheap Questions Instead Of Making You Enumerate, It Guesses Your Grant, And The Product Is The Gap Between What You Predicted And What It Found, Which Is The Third Instance Of The Same Instrument In Three Days **version** v0.33.64 **date** 4 September 2026 **from** Human (project lead) **to** Product, Design, Engineering **type** Dev brief *Eighth of 4 September. The memo proposes a guessing game in the shape of the well-known character-guessing one: the player thinks of an agent, the game asks questions, and it guesses the agent, its mandate and its grant. It asks for two modes, one purely computational over the graph and one with a model adding judgement, and it notes that engagement would let us measure whether users were surprised and gather metrics at scale. One naming collision is caught here, because the word surprise was given a precise and different meaning in a brief written earlier today and the two must not share it. Limitation: the decision tree does not exist yet and its question set is the part most likely to be wrong, so the numbers below about how many questions are needed are arithmetic rather than measurement.* --- ## What This Is A specification for a guessing game whose real output is a measurement: **the design's strength is that it inverts the burden, since every other approach asks a person to enumerate what their agent can do, which is exactly the thing they do not know and the reason the product exists, whereas a guessing game asks cheap questions with obvious answers and infers the rest, so completion rises and the answers get more honest as the questions get easier; the tree is the public artefact and should be published, forked and argued with in the same repository as the probes, because a wrong branch is a contribution rather than a complaint and because a tree nobody can inspect is a fortune teller; the mechanic is ordinary decision-tree induction and the question to ask next is always the one that splits the remaining profile space most evenly, which is why fifteen or so binary questions can separate thousands of configurations and why the game feels uncanny rather than tedious; the two modes the memo asks for map exactly onto a split this estate has now made three times, since the deterministic mode is arithmetic over a published tree and is reproducible, auditable and free to run, while the model-assisted mode handles free text and the setups the tree has never seen and is neither reproducible nor auditable, so they are offered as different instruments rather than as a better and a worse version of one; the measurement is the product and it is the same instrument built twice already this week, because the player predicts their own grant before the reveal, the tree computes what the grant actually is, and the gap between prediction and computation is the finding, exactly as the gap between a model's self-report and its transcript was the finding on Wednesday and the gap between a grant and a mandate is the finding of the whole product; that gap needs its own name because surprise was defined precisely in a brief written this morning as an action outside the grant, which is a measurement failure rather than a feeling, so the game measures a prediction gap and the word surprise stays where it was put; and the metrics ambition has a constraint the memo does not mention, since a brief written today committed the site to collecting nothing by default, so the game runs entirely in the browser and submits at most an anonymous tuple of answers, prediction and result, offered rather than taken, which is enough to test the hypothesis and not enough to identify anybody.** New contributions: **the inversion of burden as the reason this works, the tree as a public forkable artefact, the split-evenly question rule with the arithmetic, the two modes as different instruments, the prediction gap as the product and the third instance of the same shape, the naming collision caught, and the client-side metrics design.** ## Why A Guessing Game Beats A Questionnaire The product's central problem is that people cannot enumerate what they have granted. A questionnaire asks them to do precisely that, which is why questionnaires in this space return a copy of the vendor's marketing. **A guessing game asks questions whose answers people actually know.** Do you use it in a browser or a terminal. Have you ever clicked something that said always allow. Can it read your email. Does it have a key to anything that costs money. Those are cheap, and from perhaps fifteen of them the tree can place somebody in a profile. Three consequences, and the third is the one that makes it a product rather than a toy. **Completion rises**, because the burden per question is near zero and the player is being entertained rather than audited. **Honesty rises**, because an easy question invites a true answer, and a hard question invites the answer that makes the respondent look competent. **And the player commits to a prediction.** Before the reveal, the game asks what they think their agent can do. That prediction is the measurement, and no questionnaire ever gets one because a questionnaire's respondent has already told you what they think. ## The Tree Is The Artefact, And It Is Public The tree ships in the same repository as the probes, for the same reasons and with one additional one. **A wrong branch is a contribution.** Somebody whose setup the tree misclassifies has produced a finding: either a profile is missing or a question is ambiguous, and both are pull requests. **A tree nobody can inspect is a fortune teller.** The uncanny feeling is the point of the entertainment and the enemy of the assessment, so the game should offer, at the end, the exact path it took and the question that decided each step. **And the tree is where the estate's knowledge accumulates in a form a non-specialist can improve.** Somebody who cannot write a probe can still say that question four does not apply to their setup. ## The Mechanic Ordinary decision-tree work, stated so the deterministic mode is a specification rather than a vibe. ``` profiles P every published configuration, with its measured grant belief b(p) a distribution over P, starting at the population prior question q a predicate over profiles, with an expected answer per profile choose q maximising the expected reduction in uncertainty about P update b on the answer, allowing for wrong answers rather than eliminating outright stop when one profile dominates, or when the next question buys almost nothing ``` **Choosing the question that splits the remaining space most evenly is the whole trick.** A perfectly balanced binary question halves the candidates, so ten reach about a thousand configurations and fifteen reach tens of thousands. Real questions are not balanced and answers are sometimes wrong, so **budget twenty and stop early when the belief concentrates**, which is why the game feels short when somebody has an ordinary setup and long when they have an unusual one. Taking longer on an unusual setup is correct behaviour and should be said out loud rather than hidden. **Never eliminate on one answer.** A player who misremembers should be recoverable, which means answers update a belief rather than prune a branch, and the game can offer to revisit the question that mattered most. ## Two Modes, Two Instruments The memo asks for a deterministic mode and a model-assisted one. They are not a lesser and a greater version of the same thing, and presenting them that way would repeat a mistake this estate has now avoided twice. | | **Deterministic** | **Model-assisted** | |---|---|---| | Runs on | The published tree and the profile set | The same, plus a model | | Handles | The configurations somebody has mapped | Free text, unusual setups, "it's a bit like that but" | | Reproducible | **Yes.** Same answers, same result, forever | No | | Auditable | **Yes.** The path is the explanation | Only as far as the transcript | | Costs | Nothing to run | A model call per turn | | Where it runs | **The browser** | A call out | | Its answer is | A profile match, or an honest "not in the set" | A best reading, labelled as one | **The deterministic mode is the default and the model is the fallback**, which is the opposite of how these things are usually built. The reason is that the tree's failure is informative and the model's is not: a tree that cannot place somebody has found a gap in the profile set, which is a contribution, while a model that cannot place somebody produces a plausible guess and the gap is never recorded. This is the third time the same split has been drawn this week, after the passive-mode analyst that needs no model and the status block that needs two producers. **It may be worth naming the pattern once**: computation where the answer is checkable, a model where it is not, and never a model in front of an arithmetic step. ## The Prediction Gap Is The Product The memo's most valuable line is that engagement would let us measure whether users were surprised. That is the whole instrument and it should be the first thing built rather than analytics added later. ``` 1 the questions cheap, fifteen or so 2 THE PREDICTION "before we show you: what do you think it can do?" a short multiple choice over capability families 3 the reveal the tree's computed grant for the matched profile 4 THE GAP what you did not predict, with the irreversible items first 5 the reduction what to change, immediately, with what each costs ``` **Step two is the experiment and step four is the finding.** Without step two the game is a lookup with a costume on. And it is the same instrument as the other two built this week: | Instrument | Belief | Fact | The gap is | |---|---|---|---| | Status block | The model's account of its own usage | The transcript | Whether a model can be trusted to account for itself | | Conformance layer | What a team says it does | What is attested and at what tier | Where the standard is unevidenced | | **This game** | **What a person thinks their agent can do** | **The measured grant for their profile** | **Whether people know what they have granted** | Three instruments, one shape: **a belief, a measurement, and the distance between them as the product.** That is the unified theme the earlier memo was reaching for, and it may be the clearest single sentence about what this whole estate builds. ## The Naming Collision, Caught A brief written this morning defines a **surprise** precisely: an action outside the grant, which is a measurement failure rather than a feeling, and whose count is the assessment's own validity test. The memo uses the same word for the player's reaction. **Those are different things and must not share a word**, or the falsifiability argument dissolves the first time somebody reads a metric called surprises and thinks it means people were startled. | Term | Means | Where | |---|---|---| | **Surprise** | An action outside the grant. A measurement failure | The assessment, and only there | | **Prediction gap** | What a person did not expect their grant to contain | The game, and only there | The game measures a prediction gap. It never reports surprises. This is the sixth naming collision this corpus has caught, and the cheapest, because both terms are two days old. ## Metrics Without Surveillance The memo wants metrics at scale and a brief written today committed the site to collecting nothing by default. Both hold, with one design. **The game runs entirely in the browser.** The tree, the profiles and the computation are static files, so nothing about a player's answers needs to leave their machine for the game to work. **At the end it offers a submission**, which is a tuple and nothing else: the answers given, the profile matched, the prediction made, and the gap computed. No address, no identifier, no free text, no timestamp finer than a day. Offered, with the exact contents shown, and declined by default. That is enough to answer the questions worth asking. Is the median gap large. Which capability families are most consistently unpredicted. Do people who use one product predict better than people who use another. Does prediction improve on a second visit. **And it is not enough to identify anybody**, which is the condition under which people will actually submit. **Publish the aggregate as it accumulates, including if it disproves the hypothesis.** A game that reports the median gap is small would be the most credible thing this estate could put on the internet. ## What The Game Must Not Become The most likely failure is that it works too well as entertainment and the reveal becomes a scare. - **Not a fear appeal.** This corpus already found that high threat without efficacy produces denial, so **the reduction arrives on the same screen as the gap**, never a click later. - **Not a shaming mechanic.** A large gap is the normal state and the copy should say so before the number appears. - **Not a fortune teller.** The path is always available, and a wrong guess is met with an invitation to correct the tree. - **Not a lead-generation form.** No address is requested, at the end or anywhere. - **Not a claim about the player's actual environment.** It matched a profile from answers, which is the weakest tier available, and the result says so and points at the probe that would establish it properly. That last one deserves emphasis. **The game produces a hypothesis about somebody's grant, not a measurement of it.** The probe repository produces the measurement. The honest arc is: the game gets you interested in five minutes, the probe tells you the truth in ten, and the difference between those two results is itself worth showing. ## Build Order | # | Step | Produces | |---|---|---| | 1 | The profile set with measured grants, from the repository | Nothing to guess without it | | 2 | Twenty questions, each with an expected answer per profile | The tree, as data | | 3 | The deterministic engine in the browser, with the path shown | A playable game, no model, no server | | 4 | The prediction step | **The instrument.** Without it there is no finding | | 5 | The gap and the reduction on one screen | The value, and the guard against a scare | | 6 | The optional tuple submission, contents shown | Metrics without surveillance | | 7 | The model-assisted fallback for unmatched setups | Coverage for the unusual, labelled as a reading | | 8 | The correction path: a wrong guess becomes a pull request | The tree improves from being played | **Steps 1 to 5 are the whole product.** Six is the research, seven is coverage, and eight is what makes it get better. ## What This Does Not Try To Be - **Not a measurement.** It matches a profile from self-reported answers, which is the weakest tier, and the result says so. - **Not a replacement for the probes.** It is the five-minute front door to the ten-minute truth. - **Not a personality quiz.** The output is a capability list and a reduction, not a type. - **Not a data collection exercise.** Nothing leaves the browser unless offered, and the tuple is shown before it is sent. - **Not a finished tree.** The question set is the part most likely to be wrong and the correction path exists because of it. ## Honest Tensions | Tension | Note | |---------|------| | A game about somebody's exposure | It is the lowest-friction way to make the point and it makes a serious subject playful, which some audiences will resent | | Guessing rather than asking | It gets honest answers and it produces a hypothesis where people will hear a verdict | | Deterministic by default | The failure is informative and it means the game will refuse to guess for exactly the people with the most unusual setups | | The prediction step | It is the entire finding and it adds friction at the moment the player most wants the answer | | Publishing the tree | It invites correction and it lets anybody see how to answer their way to a flattering result | | Metrics by offer only | It is the condition for people submitting and it will produce a sample biased toward the curious | ## Open Questions | Question | Notes | |----------|-------| | How many questions does it actually take? | Fifteen to twenty is arithmetic over a balanced tree, and the real answer depends on a question set that does not exist | | What is the prediction step's format? | A short multiple choice over families is the guess; a free-text answer would be richer and unscoreable | | Does the gap shrink on a second play? | It is the most interesting longitudinal question and it needs a returning player who submitted once | | Should a good result be shareable? | The precedents say a service spreads only if good news is worth posting, and a small gap is a good result | | Who arbitrates a corrected branch? | Same problem as a contested probe, and the same answer is probably two profiles rather than one | | Is the population prior honest? | Starting from what most people have makes the game feel accurate and biases the first questions toward the common case | ## Relationship To Previous Briefs | Date | Document | Relationship | |---|---|---| | 4 Sep | `v0.33.64__dev-brief__the-grant-mandate-repo-ships-probes-not-tables-grants-are-measured-per-tool-and-contributed-mandates-are-elicited-locally-and-never-leave.md` | The profile set and the measured grants this guesses at, and where the tree lives | | 4 Sep | `v0.33.64__strategy-brief__name-the-question-not-the-concept-a-surprise-is-an-action-outside-the-grant-so-the-assessment-is-falsifiable-and-the-scan-cannot-run-from-outside.md` | **The collision caught here**: surprise is defined there and the game must use prediction gap | | 4 Sep | `v0.33.64__research-brief__the-precedents-lead-with-a-grade-and-scale-by-selling-the-same-thing-larger-and-the-probe-and-finding-vocabulary-already-exists.md` | Why a good result must be shareable, and the verdict-first pacing the reveal follows | | 3 Sep | `v0.33.63__dev-brief__the-status-block-is-a-self-report-by-design-which-makes-it-a-calibration-instrument-rather-than-a-meter-and-the-gap-per-model-is-the-finding.md` | The same instrument built for a model, and the first of the three | | 3 Sep | `v0.33.63__dev-brief__passive-mode-was-specified-in-may-and-what-is-new-is-that-the-sensor-sits-inside-the-thing-observed-the-derived-band-is-a-curve-nobody-but-a-person-can-choose-from-and-simulation-is-the-step-the-memo-drops.md` | The analyst that needs no model, which is the same split as the two modes here | | 20 Aug | `v0.33.61__dev-brief__user-section-is-a-conformance-test-store-the-choices-not-the-answers-high-threat-without-efficacy-produces-denial.md` | Why the reduction shares a screen with the gap rather than following it | --- ## Key Claims | # | Claim | |---|-------| | 1 | A guessing game inverts the burden, asking questions people can answer instead of asking them to enumerate what they do not know | | 2 | Completion and honesty both rise as the questions get cheaper, and a hard question invites the answer that looks competent | | 3 | The tree is a public forkable artefact, because a wrong branch is a contribution and an inspectable path is what separates a game from a fortune teller | | 4 | The next question is always the one that splits the remaining profile space most evenly, which is why fifteen or so questions reach thousands of configurations | | 5 | Answers update a belief rather than prune a branch, so a player who misremembers is recoverable | | 6 | The deterministic mode is the default and the model is the fallback, because a tree's failure is a recorded gap and a model's failure is a plausible guess | | 7 | The player commits to a prediction before the reveal, and that step is the entire experiment | | 8 | The prediction gap is the product, and it is the third instance this week of a belief, a measurement, and the distance between them | | 9 | Surprise is already defined as an action outside the grant, so the game measures a prediction gap and never reports surprises | | 10 | The game runs in the browser and submits at most an anonymous tuple, shown before it is sent and declined by default | | 11 | The reduction shares a screen with the gap, because high threat without efficacy produces denial | | 12 | The game produces a hypothesis about a grant and the probes produce the measurement, and the difference between the two results is itself worth showing | ## Sources - Public explanations of the character-guessing game's mechanic, read on 4 September 2026 to confirm the approach is ordinary decision-tree induction over a belief distribution with questions chosen for information gain, rather than anything proprietary. https://medium.com/@inemri/how-akinator-reads-your-mind-unveiling-the-games-algorithmic-magic-c8ee86dbc1d3 --- This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).