# Levels And Variants Are Two Axes Rather Than One: The Generator This Programme Needs Was Specified On 9 August With Its Governing Rule Already Written, And The Most Persuasive Scenario Is The One The Reader Is Sitting Inside

**version** v0.33.61
**date** 20 August 2026
**from** Human (project lead)
**to** Product, Ambassador, the pki.sgit.ai site agent, Design

**type** Dev brief

*Fourteenth of 20 August. The prior persona machinery and its governing rule were located in the repository and are quoted from it rather than recalled. The worked instance in this brief was measured inside the session that produced it, including which capabilities were actually exercised today and which were held and never used, and it is reported with the uncomfortable half intact. Limitation: no interface exists to test, so this specifies a programme rather than reviewing one, and the sample-size argument is arithmetic rather than a claim about any planned study.*

---

## What This Is

The experimentation programme the memo asks for, the machinery it already has, and the reframing that makes the scenarios land on people who have actually used an agent: **the memo wants three to seven scenarios spanning non-technical through advanced users, arranged like levels of a game, varied and tested and fed back on, with something shareable at the end, and the driving question is what the right altitude is for explaining a gap that the memo honestly admits its own author does not always hold in mind; the machinery is not new, because on 9 August this corpus specified a persona generator that compiles a view of underlying facts for a named audience with every element resolving back to the evidence it came from, and it set the rule this whole programme turns on, that a persona may change emphasis, ordering, vocabulary and format and may never change what is being accepted, so what is needed is that generator pointed at a grant tree rather than a risk register; the memo then conflates two things that must be separated or the experiment cannot be read, since levels are depth of detail for one reader and variants are alternative renderings of one level, which are orthogonal, and a programme that mixes them cannot tell whether a poor result came from the wrong depth or the wrong wording; levels are also not expertise tiers, because expertise predicts vocabulary and does not predict whether somebody knows their own grant, and the advanced user is the hardest case rather than the easiest since they hold the largest grant and the greatest confidence that they already understand it, which is the combination the fear-appeal work makes most likely to produce denial; two of the memo's organising words are already load-bearing here, since altitude means stakeholder level in the risk chain and persona means the generator described above, and using them loosely will collide with machinery this programme wants to reuse; the scenarios should be ordered by grant size rather than by job title, which resolves the memo's own hesitation about whether these are personas or use cases, because the personas fall out of the ordering; and the reframing that matters most came from measuring this session, because the memo's stated mandate for it was to produce a memo and the capabilities beyond that were not merely held, they were exercised, repeatedly, and they improved the output, so the honest scenario is not what could go wrong but which parts of your grant you actually used, which is derivable from a transcript and is the only version that survives contact with somebody who has used an agent and knows they needed it.** It is the fourteenth document of 20 August (cross-ref: the v0.33.57 executive view brief, the v0.33.61 user section brief, the v0.33.61 end-to-end flow brief, the v0.33.61 shared drive brief, and the v0.33.61 analogies brief). New contributions: **the generator located and its rule applied to variants, levels separated from variants, expertise separated from awareness, two naming cautions, the scenarios ordered by grant, the exercised-versus-held reframing, and the stopping rule as the answer to the altitude question.**

## The Machinery Exists, And Its Governing Rule Is Already Written

The memo asks for a way to generate and vary explanations programmatically. **That was specified on 9 August**, for a different subject.

The executive view brief established that a view for a named audience is **compiled rather than authored**: technical detail is lifted into something somebody can act on, every element resolves back to the evidence it was compiled from, and the technical view is always one step away. It then treated framing as unavoidable, which makes it a governance question rather than a rendering one, and produced the line that this entire programme depends on:

> **A persona may change emphasis, ordering, vocabulary and format, and may never change what is being accepted.**

That is the rule for variants, written eleven days early. And it is mechanically checkable in exactly the way this corpus likes: **every variant renders the same underlying fact set, so a diff of the facts across variants must be empty.** A variant that introduces or drops a fact is not a variant, it is a different assessment, and the difference is detectable without anybody's judgement.

So the programme is: point the existing generator at a grant tree instead of a risk register, and inherit the source-map requirement so that any claim in any variant is one click from the node it came from.

**The 9 August brief also named this programme's central hazard before it existed.** It warned that a system which learns what an audience responds to **will optimise for acceptance rather than understanding unless the measured outcome is chosen deliberately.** That is precisely what a rounds-and-feedback loop does, and the previous brief today already chose the outcome: visitors who take a named action, not visitors who felt alarmed.

## Levels And Variants Are Two Axes, And Conflating Them Wastes The Experiment

The memo hesitates over what these things are. The project lead: **"it's not A/B, it's not use case, but it's like it's sort of like the A/B testing workflow"** and **"maybe levels of a game is a good example."**

Both are right and they are different dimensions:

| | What it varies | Held constant | Question it answers |
|---|---|---|---|
| **Level** | Depth of detail | The reader | How much does this person need? |
| **Variant** | Wording, ordering, emphasis, visual form | The facts, by rule | Which rendering is understood? |

**They are orthogonal, so the design is a grid rather than a ladder.** Five levels by three variants is fifteen renderings, and that is fine because they are generated rather than written.

The reason to separate them is diagnostic. **A conflated programme cannot tell whether a poor result came from the wrong depth or the wrong wording**, which are opposite fixes: one adds or removes content, the other changes how the same content reads. A round that changes both learns nothing, and this corpus has been here before, in the publishing matrix of 16 August, where the move that made the space tractable was separating what must hold everywhere from what genuinely varies.

## Levels Are Not Expertise Tiers, And The Advanced User Is The Hardest Case

The memo's most honest sentence undermines its own ladder, usefully. The project lead: **"anybody who uses an agent, we should assume that they are they today don't have a full understanding. Even I would argue me, sometimes have a full understanding of what the agent can do."**

**Expertise predicts vocabulary. It does not predict whether somebody knows their own grant.** Those are different things, and the ladder metaphor quietly assumes they are the same by starting advanced users higher up.

Worse, the direction is inverted:

| | Non-technical user | Advanced user |
|---|---|---|
| Size of the grant | Small. One hosted assistant | **Large.** Several agents, tokens, local execution, permission-skipping flags |
| Prior confidence they understand it | Low | **High** |
| Vocabulary needed | Plain | Technical |
| Likely response to being told they are exposed | Curiosity | **Reactance** |

So the advanced user holds the biggest gap and the strongest prior that there is nothing to learn, which is the combination the fear-appeal work identified as most likely to produce defensive response rather than change.

**The consequence for the design is concrete: everybody starts at level one.** What changes with audience is the language, not the entry point. An advanced user who is dropped into a deep technical view gets confirmation that they already knew this, and the one thing the programme needs them to feel is the surprise that level one delivers.

The memo says this itself and it should be a rule rather than an aside. The project lead: **"even advanced users needs to have something simple."**

## Two Of The Memo's Organising Words Are Already Taken

The fifth naming check in a fortnight. Neither is a hard collision and both will cause confusion if used loosely, because in each case the existing meaning belongs to machinery this programme wants to reuse.

| Word | Means here already | The memo's use | Recommendation |
|---|---|---|---|
| **Altitude** | Stakeholder level in a risk chain, where each altitude has its own language and its own acceptor | Depth of explanation | **Keep altitude for stakeholder. Use depth or level for detail** |
| **Persona** | The generator that compiles a view for a named audience, with a source map back to evidence | A user archetype for testing | **Keep persona for the generator. Use scenario or archetype for the test subject** |

The altitude one is worth a second look because the two meanings are close enough to feel interchangeable and are not. **Stakeholder and depth are two axes**: a board member wants low detail in board language, and a platform engineer wants high detail in engineering language, but a board member investigating an incident may want high detail in board language. Collapsing them into one word makes that third case impossible to express.

## The Scenarios, Ordered By Grant Rather Than By Job Title

The memo asks for three to seven and hesitates about what organises them. **Order them by grant size**, and the personas fall out rather than having to be invented, which answers the memo's own uncertainty about whether these are personas or use cases: they are neither, they are positions on one axis.

| # | Scenario | Mandate | Grant | Why it is at this position |
|---|---|---|---|---|
| 1 | **Dictation.** A hosted assistant turning voice notes into documents | Produce a document | The surface's whole capability set, including search, memory and file handling | Smallest gap, most relatable, and it is the memo's own example |
| 2 | **Document work.** An assistant connected to mail or a drive | Summarise last week from one team | The entire mailbox or drive, including everything anybody ever shared | First time the grant contains other people's material |
| 3 | **Vibe coding.** A desktop or web coding tool on a personal project | Build the feature | The user account on that machine | First time the grant includes execution |
| 4 | **Professional development.** A command line agent with a code host token | Commit to one branch this afternoon | Every repository the person can reach, plus the machine, plus whatever permission-skipping flags are set | Largest grant most people actually hold |
| 5 | **Operations.** An agent with production access | Correct three rows | The estate | Where the gap stops being personal |

Five is the right number and the memo's range allows it. **Each is a level in the memo's sense and a scenario in its own right**, so the two things it was trying to reconcile turn out to be the same list read two ways.

## The Worked Instance Is This Session, And The Uncomfortable Half Is The Useful One

The memo supplies its own example and it is the best asset in the programme. The project lead: **"even when I create these voice memos, the mandate fundamentally at this moment in time is just to create a memo... the agent, even when a web, should not have access to anything else apart from the past knowledge, and apart from this information, and the ability to create a document. But I, for example, I actually think that the Claude agent on the web already is able to do searches, for example, or even has the memory feature."**

**That is correct, and it was measured rather than assumed**, in the session that produced this brief.

| Capability | Within the stated mandate? | Exercised today? |
|---|---|---|
| Read the uploaded memo | Yes | Yes |
| Produce a document | Yes | Yes |
| Web search and page fetching | **No** | **Yes, repeatedly** |
| Execute arbitrary shell commands | **No** | **Yes** |
| Write files anywhere in the workspace | **No** | **Yes** |
| Clone a public repository over the network | **No** | **Yes** |
| Read the session's own transcript directory | **No** | **Yes** |
| A configured commit signing identity | **No** | Held, not used |
| Passwordless escalation | **No** | Held, not used |
| Read other accounts on the machine | **No** | Held, not used |

**And here is the half that makes the scenario credible rather than preachy.** The capabilities outside the mandate were not merely held. They were used, all day, and **they are the reason the output is any good.** Searching found that a proposal was already specified in May. Cloning the repository found a registry design from June. Fetching documentation corrected a claim about what ships. Reading the environment produced the measured grant trees. Every one of those is excess authority, exercised, with a better result.

That is the objection every experienced reader will raise, and the scenario library has to raise it first:

> **The gap is not only where the danger is. It is also where the value came from.** That is precisely why nobody closes it, and a scenario that pretends otherwise will be dismissed by the only people who could act on it.

Which produces the reframing that should organise the whole library.

**Do not ask what could go wrong. Ask which parts of your grant you actually used.** The exercised set is derivable, because the session transcript is an ordinary file on disk, established twice today from two directions. So a scenario can show three sets rather than two: what you were asked for, what you used, and what you held and never touched. The third set is the one nobody can defend, and it is the one that shrinks without anybody losing anything.

| Set | What the reader thinks |
|---|---|
| Mandated | Obviously fine |
| **Exercised beyond the mandate** | **I needed that. Now we are having an honest conversation** |
| **Held and never exercised** | **Why does it have that?** |

That third row is the product. It is measurable, it is undeniable, and it costs the reader nothing to agree with.

## The Experiment Will Not Detect What It Is Looking For, And The Corpus Says Why

Two warnings about the rounds-and-feedback plan, both from work already done here.

**The visual will wash out the framing difference.** The 9 August brief recorded, among the research it drew on, that **presenting a visual alongside text eliminated a framing effect that was significant in the text-only condition.** This programme's output ends in a rendered grant tree. So an experiment varying the wording of a page that culminates in a picture may find no difference between variants, and the correct conclusion would be that the picture is doing the work, not that framing does not matter. **Test the visual variants first**, because that is where the effect lives.

**And the sample will not support the design.** Five levels by three variants is fifteen cells. A round of the size the memo describes, where the project lead hands it to people he knows and asks what they thought, is a few dozen people at most. At that scale this is **qualitative research and should be run as such**: watch somebody use it, ask them to say back what they think it told them, and count nothing. The corpus already holds this shape from 9 August, where the caution was that a percentage implies calibration nobody has measured and the recommendation was to show the structure of the evidence until a track record exists.

Naming that now prevents a false-precision result, which is the expensive failure mode here because it would be used to pick a design.

## The Stopping Rule Answers The Altitude Question

The memo asks what the right depth is. **There is not one, and looking for it is the wrong shape.** The testable property instead:

> **Each level must be a complete answer, not a step towards one.** A reader who stops at any level has gained something whole.

That is checkable without instrumentation. Show somebody only level three and ask what they would do about it. **If the answer is that they would keep reading, level three failed.** Not because it was too shallow, but because it was written as scaffolding rather than as an answer.

That also removes the need to guess the audience. A reader finds their own depth if every stop is a landing rather than a stair, and the programme's job is to make each landing complete rather than to route people correctly.

## Shareable Is Already Solved By Yesterday's Rule

The memo wants an artefact people can pass around. The project lead: **"have something that they can share, something that they can collaborate, or something they communicate how it get went."**

**The obvious version is dangerous and the previous brief already prevented it.** The natural thing to share is a picture of what your agent can reach, which is a description of your own estate. The choices-only storage rule means the stored artefact is a set of references into a public library, so **a share discloses nothing personal by construction** rather than by care.

And the mechanism exists. The corpus settled on 16 August that a fragment is the key channel because it is never sent to the server. **The same channel carries a selection**: a share is a link whose fragment holds library identifiers, so the sharing party's choices never reach any server, and the recipient's browser reconstructs the same view from public material. One pattern, two payloads.

## What This Does Not Try To Be

- **Not a new generator.** It is the 9 August persona machinery pointed at a grant tree.
- **Not a ladder by expertise.** Everybody starts at level one and the language changes, not the entry point.
- **Not an A/B programme.** At the described sample size this is qualitative research and reporting it as anything else would be false precision.
- **Not a catalogue of what could go wrong.** The library shows what was actually exercised, because that is the version that survives an experienced reader.
- **Not a settled vocabulary.** Two words are flagged and neither is decided here.

## Honest Tensions

| Tension | Note |
|---------|------|
| Levels as a grid rather than a ladder | It makes results readable and it multiplies what has to be produced, even generated |
| Everybody starts at level one | It is right for awareness and it will read as patronising to the exact audience that most needs it |
| Showing what was exercised | It is the honest and persuasive version and it concedes the reader's best objection before making the argument |
| The transcript as the source of the exercised set | It is complete and it is the artefact that carries every secret the session saw, which today's other briefs establish |
| Qualitative rather than measured | It is what the sample supports and it will not settle an argument between two people who prefer different designs |
| Reusing the persona generator | It brings a written rule and a source map, and it binds this programme to machinery built for a different subject |

## Open Questions

| Question | Notes |
|----------|-------|
| Who authors level one? | Everything depends on it and it is the hardest thing on the list to write |
| Is the exercised set derivable without a scan? | It comes from a transcript, which is local, and getting it into the assessment reopens the storage rule |
| What does a variant diff check compare? | The rule needs a fact set with identity, which the grant tree gives and the prose does not |
| How are the two taken words settled? | Cheap now, and the fifth such decision in a fortnight |
| Does the share link survive a library version change? | A recipient opening it next month may reconstruct a different tree, and neither party would know |
| What is the feedback instrument? | Watching somebody use it, and nobody has time to watch forty people |

## Relationship To Previous Briefs

| Date | Document | Relationship |
|---|---|---|
| 9 Aug | `v0.33.57__dev-brief__sg-send-executive-view-is-compiled-framing-is-governed-grc-connects-simulation-finds-gaps.md` | The persona generator, the source map, the emphasis-not-substance rule, the visual mitigation, and the optimisation hazard, all specified there |
| 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` | The choices-only rule that makes sharing safe, and the outcome measure this programme must optimise |
| 20 Aug | `v0.33.61__arch-brief__end-to-end-flow-is-the-august-worked-example-with-an-agent-as-the-twin-grant-tree-needs-control-labels.md` | The grant tree these scenarios render, and the library of pre-computed trees they select from |
| 20 Aug | `v0.33.61__dev-brief__shared-drive-is-a-watcher-a-queue-and-three-vaults-batching-was-specified-in-may-ignore-file-blocks-it.md` | The session transcript on disk, which is where the exercised set comes from |
| 16 Aug | `v0.33.59__dev-brief__publishing-matrix-five-invariants-fourteen-tests-local-folder-needs-a-server.md` | Separating what holds everywhere from what varies, and the fragment as a channel never sent to a server |
| 20 Aug | `v0.33.61__strategy-brief__analogies-only-work-if-you-name-what-is-missing-hotels-solved-the-bounded-grant-the-law-already-assigned-the-loss.md` | The audience-by-audience guidance these scenarios operationalise |

---

## Key Claims

| # | Claim |
|---|-------|
| 1 | The generator this programme needs was specified on 9 August, compiled with a source map back to evidence |
| 2 | Its rule already governs variants: emphasis, ordering, vocabulary and format may change, and substance may not |
| 3 | That rule is mechanically checkable, since a diff of the fact set across variants must be empty |
| 4 | Levels and variants are orthogonal, and conflating them makes a poor result undiagnosable |
| 5 | Expertise predicts vocabulary and does not predict whether somebody knows their own grant |
| 6 | The advanced user holds the largest grant and the strongest prior, which is the condition for reactance |
| 7 | So everybody starts at level one and only the language changes |
| 8 | Altitude and persona are both already load-bearing here, and both belong to machinery this programme reuses |
| 9 | Scenarios ordered by grant size produce the personas rather than requiring them to be invented |
| 10 | Measured in this session, capabilities outside the stated mandate were exercised repeatedly and improved the output |
| 11 | So the library should show mandated, exercised and held-but-never-used, and the third set is the product |
| 12 | A visual eliminated a framing effect significant in text alone, so testing wording around a rendered tree may detect nothing |

---

## Sources

- The 9 August executive view brief, establishing that a view for a named audience is compiled rather than authored with every element resolving to the evidence it came from, that framing is unavoidable and therefore a governance question, that a persona may change emphasis, ordering, vocabulary and format and may never change what is being accepted, that presenting a visual alongside text eliminated a framing effect significant in the text-only condition, and that a system learning what an audience responds to will optimise for acceptance rather than understanding unless the measured outcome is chosen deliberately: the project repository, cloned and searched on 20 August 2026, with the document named in the relationship table above
- The capabilities present and exercised in the session that produced this brief, distinguishing those used from those held and never invoked: measured directly on 20 August 2026, held with the project record

---

This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).