Naledi surveyed data journalism, motion graphics, spatial interfaces, game UI, and scientific visualization for patterns that already demonstrate real generative and structural ambition — but were never applied to business communication. Mateo then built four working prototypes, each one deliberately designed so the agent has to make the real generative decision, not just fill in text around hand-built scaffolding.
The interesting part usually isn't a new chart shape — it's a new relationship between the reader and the artifact. Who controls pacing, what gets disclosed and when, whether the data lives inside a depiction of its subject rather than beside it. Naledi's own framing, though Ingrid scoped it down: only about half of the 11 surveyed patterns actually support this claim — carried forward as a partial finding, not a proven generalization.
Every prototype forces the real generative decision, not a fake version of it. Ingrid flagged a concrete risk before any code was written: it's easy to build something that looks generative — hand-author the interesting part, have the agent only fill in text — without testing the capability that actually matters. Each of the four below is structured so that decision is unavoidable for the model.
Diegetic display was re-risked, not dropped. Generating a full "world model" of a subject (what its entities are, what they look like) is a categorically harder problem than the other three, which only need to sequence, interpret, or classify content the agent already has. Ingrid's fix: pre-specify the scene by hand, and only let the agent decide the data-to-visual-property mapping — a deliberately narrower test of a real but smaller capability.
These are demonstrations of feasibility, not evidence they communicate better. All four run end-to-end on real generated content below. Whether any of them actually outperforms a conventional chart for a given audience and goal is untested — and still is. That validation, using the same paired-scenario method as Sprints 6 through 8, was originally planned as Sprint 10; it has been deferred (see the note below).
Known limitation: not independently re-verified in a browser beyond the founder's own pass — the same class of rendering gap that caused a silent container-autosize bug in Sprint 8. Founder's verdict on this round: functional, some rough edges, closed as a working proof of concept rather than debugged to production quality.
Reader controls pacing by scrolling; a pinned chart updates in response, instead of a presenter advancing static slides.
Content: support-ticket volume across five product lines after a release. The agent had to identify which single insight each narrative beat should surface and which bar the pinned chart should highlight at that moment — not just write prose alongside a static chart.
Scroll inside the frame above to advance the narrative — the chart highlight updates with each beat.
What this tests: whether the model can sequence a narrative and bind each beat to the correct visual state, rather than a human hand-authoring the highlight logic.
The chart commits to one insight and annotates it directly, instead of presenting all the data and describing it separately underneath.
Content: monthly churn rate, one anomalous month. The agent had to identify the single data point that actually needs explaining and generate a real annotation layer around it (callout, reference line, comparison average) — a genuine single-insight identification and justification task, not a templated chart.
What this tests: whether the model can pick the one thing that matters and justify it, instead of rendering a complete, undifferentiated chart.
Business communication has no standard slot for network-shaped questions — they get forced into bar charts or org charts even when the actual insight lives in the connections.
Content: real engineering-team collaboration structure. The agent had to first argue why this specific question is relational rather than a magnitude comparison, then extract the actual node/edge structure from the content itself — draggable, force-simulated below.
Nodes are draggable — the simulation responds live.
What this tests: whether the model correctly classifies a question as relational versus magnitude-based before committing to a form, then extracts a coherent graph structure from prose.
Data embedded directly into a depiction of its own subject — borrowed from game interfaces, where a fuel gauge is part of the dashboard, not a separate chart next to it.
Content: supply-chain flow across a factory, transit, warehouse, and store. Per Ingrid's narrowed scope, the scene itself (the four icons and their layout) is fixed and hand-authored — the agent only decided which metric maps to which stop, and how flow volume and stop intensity should scale visually.
What this tests: the narrower, re-risked question only — can the model correctly map data onto pre-existing visual properties (line weight, opacity, icon scale)? Full agent-generated scene construction ("world modeling") remains a separate, harder open question for later.
A different question from the four patterns above: not "can the agent choose or invent a new visual form," but "given an existing image, can the agent reconstruct it as vector code at all." A prerequisite capability check, not a form-selection test.
The agent sees only the original image and writes SVG markup from scratch to reproduce it — enforced in code: the output is rejected if it re-embeds the source image instead of reconstructing it as vector shapes. Two images, same subject (a rocket), deliberately different difficulty: a flat vector icon, and a tilted monochrome illustration with fine linework.
Includes a real review cycle: the first attempt (v1) was compared directly against the original, zoomed in side by side, and specific gaps were written up — same review-then-revise pattern used throughout this project. That critique, plus the original image again, went back to the agent for a second attempt (v2). Both versions are pure vector reconstructions.
v1 result: the flat icon was reproduced with reasonably close proportions, colors, and composition. The complex illustration was not — it kept the general rocket-with-fins-and-window gestalt but lost the tilted pose (came out upright) and collapsed the fine linework into a heavy black silhouette.
v2 result, after review feedback: the simple icon's specific issues (off-center window, a stray rendering artifact, an unclosed seam near the flame) were fixed. The complex illustration improved substantially on the two biggest gaps — it now sits at roughly the right diagonal tilt, and shifted from a solid black fill toward the source's line-and-white-space style — but is still not a faithful match: the cockpit reads as a gauge-like icon rather than the source's lens shape, and the lower-body silhouette still diverges. One review cycle measurably closed part of the gap; it did not close all of it.
What this tests: raw generative-reproduction fidelity, independent of audience or goal. Says nothing about whether the agent picks the right form for a situation (Sprint 6-8) or invents a genuinely new one (patterns 1-4 above) — a third, separate axis. Pilot only: one image pair per difficulty level, one review cycle, judged by visual comparison, not a scoring rubric. Not yet formally reviewed by Ingrid — flagged openly rather than treated as approved by default.
All four patterns are demonstrated feasible with real generated content, not placeholders — a genuine expansion of what Altair8's pipeline can produce beyond flat 2D chart-and-prose, directly in service of design principle #4 (don't constrain output to legacy formats). None of this is evidence any of them communicates better than a conventional chart for a given audience and goal. That question remains untested.
This page originally stated that the paired-scenario validation of these four patterns would be Sprint 10. It will not be. Sprint 10 was re-scoped, at the founder's direction, to design competing concepts for the communication layer before building or validating any further — the team had never put rival answers side by side, only rival implementations of one answer.
The cost of that decision is named rather than carried silently: the four patterns above remain unvalidated claims, and Sprint 10's concepts are being generated against them in that state. If the validation eventually shows one of these patterns is wrong or context-limited, concepts that depend on it will need revising. The validation is deferred, not dropped, and it is still owed.