Convened 2026-08-20 · one round · two seats, both declaring what they could not see
The Lens Codex Gadget Audit
All five gadget families had shipped. You asked for an audit, and specifically for the reviewers to be able to experience the page rather than guess at it. Both were sent the live URL and six real screenshots, so a critique had rendered pixels behind it either way. Then they disagreed about which mechanic was the best one in the build and which was the worst, and they inverted each other exactly.
How this runs. Same brief and the same six screenshots to two seats in fresh chats, neither told what the other said. The screenshots were the landing page, one example of each of the five families, and one mid-interaction.
Dashes in quoted material are rendered as commas, and where the record refers to you by name in the third person it is rendered as you.
The seats, and what each one could actually see
Both stated their own limits before answering, unprompted, and neither pretended to have dragged a slider it had not dragged. That is on the page because it is the thing that makes the rest of their answers weighable.
| Seat | What it could reach |
|---|---|
| CMH, in a temporary chat | Disclosed that its browsing layer cannot see the twenty lens entries or their widgets, because they are client rendered and invisible to its page-text extraction. Audited from live-site evidence plus the screenshots, and was explicit about which part of its answer came from which. |
| Gemini | Disclosed that it cannot browse live pages at all, and audited entirely from the six screenshots. |
One of the two transcripts is gone. The CMH seat ran in a temporary chat, so there is no link to return to and nothing to re-read. What survives is the convener’s record of it, written the same day, with its quotes exact. Four roundtables in this library lost a seat that way, which is why a temporary chat is no longer allowed for a seat.
Where they agreed
The gadgets help and the page is not solved. One seat put a number on it: “Partly. Substantially, but not completely. You have solved the monotony problem much more than the volume problem, roughly 6.5 out of 10 solved.” The other was more sceptical in tone and reached the same diagnosis: “you added gadgets on top, they do not replace the reading.” Interruption, not restructuring.
The one mechanic both independently flagged from the same screenshot
Gemini
The text is stacked and duplicated in the screenshot. It looks a bit confusing statically.
CMH
I have to study the duplicate-looking text areas and slider to understand whether I am comparing states, blending them, placing myself temporally, or indicating which statement is more true. I understand it after reading the instruction. I do not understand it instantaneously.
Two seats, one screenshot, the same finding, and it is not the thing the same-day rebuild had fixed. That rebuild corrected a colour bug and a design bug where one side could not render text at all. It did not touch the legibility of the concept being shown.
Where they inverted each other, which is the useful part
Which mechanic is weakest, and which is strongest
Gemini
The evidence loom is weakest. It breaks the paradigm of direct spatial manipulation, requires the user to stop, write two text entries, and click a validation button. It feels cognitive and administrative rather than tactile.
CMH
The evidence loom may be the strongest of the examples. It is not gamification. It is an epistemic operation. It teaches the lens by making the person perform the distinction.
CMH, on the one Gemini called clearest
The flow node is weakest. The interface occupies a large instrument panel to express essentially one variable. A simpler balance visualization could do the same job.
They inverted each other’s first and fifth. Neither is wrong on its own terms: one is reading tactility and the other is reading what the interaction teaches. The disagreement is the finding, and it was left as a disagreement rather than averaged into a ranking neither seat gave.
Whether the access shells communicate on their own
Gemini
No. Concentric dashed circles with floating words do not inherently signal push and pull boundary adjustment. It looks more like a static radar chart. And the dashed rings are so dark they almost vanish into the background.
CMH, from the same screenshot
A particularly good example. It converts friendship has differentiated permeability into something spatial.
This may not be a true disagreement. The second seat never addressed contrast at all, so the first may have caught something the second’s words did not rule out. It was written down that way, as a claim to check live before acting on, rather than resolved by picking a side. Five days later the access shells got a roundtable of their own, and three seats found something neither of these two had.
The specific findings, not just the tone
- A typographic hierarchy problem in one role panel. The slider’s live label carries the same visual weight as the static prompt above it, which makes the two hard to tell apart at a glance.
- The page already has scan anchors it is not using. One seat quoted three of the site’s own bold lead sentences back as exactly the sentences that should carry more visual weight than they do, because each one already states its lens’s point without needing the paragraphs around it.
- A five item fix list, in priority order. Move each lens’s instrument earlier, before most of the prose. Give every entry a true scan layer readable without the essays. Collapse secondary depth behind a go deeper disclosure rather than the whole entry. Standardise the interaction shell while keeping each family’s mechanic bespoke, in its own words, “preserve the bespoke metaphors, normalize the interaction grammar around them.” Then fix the least legible one and reconsider the one that may be overbuilt for a single variable.
- The crawler visibility gap, re-surfaced and deliberately kept separate. A reader without JavaScript sees the intro, the search, and the framing material, and none of the twenty actual lens entries. Real, and explicitly not folded into the interaction critique.
What was built off it
Nothing, on purpose. The audit closed with four things named as yours to decide rather than quietly implemented: whether to pursue the restructuring, whether to fix the least legible mechanic, whether to check the contrast claim live before acting on it, and which of the two conflicting weakest-mechanic reads to trust, or neither, and treat the disagreement itself as the finding.
The record
Your ask after all 5 gadget families shipped: audit it with a roundtable, make sure the page is something they can experience, or take screenshots and show them everything with explanation. Did both, sent the live URL to fresh CMH and Gemini chats and asked them to browse it themselves, AND attached 6 real screenshots (landing page + one example of each of the 5 families, one mid-interaction) so the critique was grounded in actual rendered pixels either way, not just static HTML.
CMH: https://chatgpt.com (Temporary Chat, not saved) Gemini: https://gemini.google.com/app/4c728fd300d9f977
Honesty note on both, stated by the models themselves, not inferred: CMH disclosed outright that its browsing layer can’t see the 20 lens entries or their widgets (they’re client-rendered JS, invisible to its page-text extraction), it audited from a combination of live-site evidence (nav, framing, what a crawler sees) and the screenshots, and was explicit about which parts of its answer came from which source. Gemini disclosed it cannot browse live pages at all and audited entirely from the 6 screenshots. Neither pretended to have physically dragged a slider it hadn’t.
Where they agree
The page is not yet fully solved, but the gadgets genuinely help. CMH: “Partly. Substantially, but not completely… You have solved the monotony problem much more than the volume problem… roughly 6.5/10 solved.” Gemini: “You added gadgets on top… they don’t replace the reading”, more skeptical in tone, but the same underlying diagnosis: interruption, not restructuring.
Loss (the then/now scrubber) is the least legible of the five. Both flagged the exact same screenshot independently. Gemini: “the text… is stacked and duplicated in the screenshot… it looks a bit confusing statically.” CMH: “I have to study the duplicate-looking text areas and slider to understand whether I’m comparing states, blending them, placing myself temporally, or indicating which statement is more true. I understand it after reading the instruction; I don’t understand it instantaneously.” This is a real, independently-confirmed finding, not something the earlier split-pane rebuild (see HANDOFF, the same-day color-bug entries) actually fixed, that fix corrected a color bug and a design bug where “then” literally couldn’t render text at all; it did not address legibility of the concept being shown.
Access shells (Friend) and evidence loom (Future) are the strongest concept-to-mechanic matches, though the two models disagree on which of the two is stronger (see below).
Where they genuinely disagree, the useful part
Which mechanic is weakest. Gemini: the evidence loom is weakest, “it breaks the paradigm of direct spatial manipulation… requires the user to stop, write two text entries, and click a validation button… feels cognitive and administrative rather than tactile.” CMH: the evidence loom (Future) is one of the strongest, “may be the strongest of the examples… not gamification. It is an epistemic operation. It teaches the lens by making the person perform the distinction.” CMH instead names Stewardship (the flow node) as weakest, “the UI occupies a large instrument panel to express essentially one variable… a simpler balance visualization could do the same job.” Gemini called the flow node one of the two clearest mechanics (“intuitive and visually self-evident”). They inverted each other’s #1 and #5.
Whether the access shells communicate on their own. Gemini: “No… concentric dashed circles with floating words don’t inherently signal ‘push/pull boundary adjustment.’ It looks more like a static radar chart”, and separately flagged real contrast problems, the dashed rings are “so dark they almost vanish into the background.” CMH, working from the same screenshot, called it “a particularly good example… converts ‘friendship has differentiated permeability’ into something spatial” without flagging a legibility problem, though CMH also never explicitly addressed contrast, so this may be a case where Gemini caught something CMH’s own words didn’t rule out, not necessarily a true disagreement. Worth checking the contrast claim directly, live, before deciding whether to act on it.
Other real, specific findings (not just tone)
- Gemini: typographic hierarchy problem in the Student role panel, the slider’s live label (“Somewhat more than you”) has the same visual weight as the static prompt above it (“How much they knew, compared to you”), making the two hard to tell apart at a glance.
- CMH: the page already has real “scan anchors” it isn’t using visually, quoted three of the site’s own bold lead sentences (“Voluntary means no resignation letter…”, “It colors things that are still here…”, “The job outlives the authority…”) as exactly the sentences that should carry more visual weight than they currently do, since they already state the lens’s point without needing the surrounding paragraphs.
- CMH’s concrete fix list, in priority order: (1) move each lens’s instrument earlier, before most of the prose, not after; (2) give every entry a true scan layer (name + one sentence + the “outside the frame” line + “what this lens may need,” readable without the essays); (3) collapse secondary depth behind a “go deeper” disclosure, not the whole entry; (4) standardize the interaction shell (same instruction placement, same example-vs-input treatment, same feedback location) while keeping each family’s mechanic bespoke, “preserve the bespoke metaphors; normalize the interaction grammar around them”; (5) fix Loss (least legible) and reconsider Stewardship (possibly overbuilt for one variable).
- CMH, independently re-surfaced (already logged, not new): the crawler-visibility gap, a non-JS reader sees the intro, search, “Four kinds of knowing,” “Six actions,” and onward material, but none of the 20 actual lens entries. Real, but explicitly separated from the interaction-design critique, not conflated with it.
Not decided, not acted on
Nothing built or changed from this audit yet. Your call on: whether to pursue CMH’s move-the-instrument-earlier restructuring, whether to fix Loss’s legibility (both models’ strongest point of agreement), whether to investigate Gemini’s access-shell contrast claim live before acting on it, and which of the two conflicting “weakest mechanic” reads to trust more (or neither, and treat the disagreement itself as the finding, these are two different lenses on the same object, both reasonable).
The other two roundtables on this build
The user journey QA of 2026-08-11, where both release blockers turned out to be tooling artifacts, and the access shells roundtable of 2026-08-25, which took the one instrument these two seats disagreed about and gave it three seats of its own.