Building the visual communication layer that AI still doesn't have.

What are the most significant documented failure modes when LLMs attempt to reason about visual layout and spatial semantics, and what approaches have shown any traction against them?

Kenji documented LLM spatial and orientation-reasoning failure modes from arXiv-sourced literature; Naledi translated findings into cognitive-load design principles; Mateo built an HTML prototype demonstrating them. Ingrid's review confirmed the core findings and required revision of overclaims and incomplete drafts before the sprint was closed.

What does the current landscape of AI-generated slide and visualization tools look like, and where do they visibly hit the ceiling we are trying to get past?

Kenji surveyed the top 20 AI slide and visualization tools and confirmed that none have solved AI-reasoned, non-linear communication structure — production is automated, communication reasoning is not. Ingrid's review required revision; following revision, Genially was flagged for immediate follow-up as the closest existing analog, Flourish's SDK was flagged as a build-vs-buy question, and Prezi was downgraded to low-priority watch.

How far does Genially get toward AI-reasoned, non-linear communication structure, and what patterns from AI-Scientist-v2/RISE should Altair8 reuse to go further?

Kenji's deep-dive found that Genially — the closest existing analog to our goal — solves the non-linear output container, AI component generation, and engagement analytics, but its navigation is scripted entirely at creation time and its AI never reasons about why a visual form communicates better, only what content fills a page. Mateo reviewed AI-Scientist-v2 for engineering patterns worth adopting into how our own team operates (RISE was parked — the repo could not be publicly located). Ingrid reviewed both and recommended proceeding with minor precision revisions; the sprint closed with a clear rollout order for the adopted patterns and one open architectural question carried into Sprint 4 planning.

Have foundation-model providers already solved our core problem through their own first-party capabilities — and how do they technically route and load those capabilities?

Kenji found that no major provider's documented first-party capability solves our core problem. Anthropic's own dataviz skill has the most developed form-reasoning artifact we've found anywhere — a heuristic mapping data's communicative job to chart type, backed by a runnable validator — but it's scoped to charts only, with no audience model, and doesn't generalize to general business communication. Mateo separately reviewed how these providers technically route and load capabilities, identifying two patterns new to this program: executable validators shipped alongside instructions, and lazy-loaded capability registries independently arrived at by two different providers. Ingrid's review required precision revisions on both deliverables before the sprint could close — the first time our new review-gate actually blocked a close pending revision, which is exactly what it's for.

Full market analysis across all 4 required categories: who has already solved — or come closest to solving — reasoning about why a visual form communicates better?

Kenji closed the remaining gaps from Sprints 1–4: the RISE repo was finally located (architecturally irrelevant to our question, but the open question is resolved), and academic literature search surfaced LLM4Vis and DracoGPT — the closest published work found to date, generating real explanations for visualization choices — but confirmed to operate at the chart-selection level, one abstraction below the presentation-level question we're actually asking. Microsoft Copilot was added to the first-party scan: it asks about audience but shows no evidence of using that input downstream. Ingrid required a revision after finding that a live database outage (only one of four literature sources worked during the run) wasn't caveated clearly enough in the findings — revised and confirmed. Across five sprints and four categories, one finding holds: no system examined demonstrates a documented connection between communication goal, audience, and the visual form it produces. That is an observation about the current landscape, not yet a diagnosis of what a solution requires — whether the gap closes through additional reasoning that establishes a semantic link between content and visual form, an entirely new visualization language for agents that goes beyond static two-dimensional slides (motion, 3D, interactive forms), or something else again, is exactly what the next phase needs to determine. This sprint completes the Problem Identification phase.

Do nine papers on argument-mapping, knowledge cartography and NLG architecture supply evidence that discriminates internalized structure from convention reproduction — the question hypothesis 6 turns on?

This sprint returns to Problem Identification after four sprints spent building, because the founder opened a direction-setting review and asked a foundational question that had never actually been settled: do models that appear audience-sensitive know something, or are they reproducing what they've seen? Nine full texts — founder-supplied where our own connectors hit paywalls or 403s, including a real bug fix along the way (an acquisition script was quietly serving landing-page links instead of the full-text URL it thought it was reading) — were read individually against that one question. The honest answer is a clean no. This body of literature was built to detect, represent and generate communicative structure, not to explain its cognitive source, and no amount of closer reading changes that. Two papers came closest: idiosyncratic sentence-planning preferences that don't transfer between individuals, and politeness patterns that generalize better across domains than a pure convention account predicts — both lean weakly against simple convention-reproduction without confirming its alternative. Ingrid's review caught the team stating that lean as full equipoise, and caught something more structural: a draft that quietly redefined the sprint's own closing criterion rather than putting the ambiguity to the founder. Corrected, and put to him directly — he ruled that a clear methodological direction ("this needs a behavioral test; literature can't settle it") satisfies the criterion, which it does. No paper came close to C1's actual mechanism, so no novelty risk was raised. The sprint's real output is negative and precise: the next step has to be a purpose-built behavioral test, and two concrete design notes for it are now on record.

Is base-model visual-form-selection reasoning (Claude, GPT, Gemini) principled or confabulated — does it track the actual audience and goal, or just fluently rationalize a form chosen some other way?

Naledi ran a pilot behavioral test: 3 paired scenarios (audience varied, goal varied, content type held as a control), 6 variants, each model responding once per variant. The reasoning found was cognitively coherent, tracked the manipulated variable, and discriminated between rejected alternatives in most cases — more consistent with partially principled reasoning than pure confabulation, though not provable as causal from API output alone. Each model showed a distinct signature: Claude's shift to plain prose for new support reps and GPT's heatmap choice for urgency detection were the strongest evidence of tracking the manipulated variable; GPT's chart-default for a narrative-comprehension audience is better characterized as "medium-frame anchoring" than generic reasoning; Gemini showed a candidate (single-pair, hedged) form-conservatism pattern — adapting within a preferred chart form rather than switching forms when the goal warranted it. Ingrid's review required revision twice: first on the original Claude/GPT pilot (removed a false-precision confidence estimate, marked cross-model comparisons as provisional, redesigned the Sprint 7 follow-up around adversarial pairs), then again after Gemini was added as a third model in a follow-up step once a Google API key became available (fixed a drafting artifact, an underspecified failure-mode label, an incomplete appendix, and an inconsistently-hedged claim). This pilot answers Ingrid's own open question from Sprint 4, and is sufficient to inform, not decide, the founder's solution-direction choice.

Is form-sensitive behavior in LLMs internalized structure or convention reproduction — the causal test Sprint 6/7's observational studies couldn't run?

The largest real experiment this programme has run: 180 model calls across 20 scenarios, three conditions, and three models (Claude, GPT, Gemini), pre-registered before any data existed and scored blind by a temporary reviewer role built for exactly this purpose. The primary result is a complete null: zero failures out of 60 baseline trials, across all three models, with no exceptions — falsifying the pre-registered floor outright rather than narrowly missing it. That single number cascaded through the rest of the design: the paired statistical test on the ordering manipulation became mathematically undefined (an outcome with no variance has nothing to test), and the follow-up question this sprint was built to answer — is a wrong answer a knowledge gap or an application failure — had no failures left to classify.

We treated that as a finding about the test, not about the models. The scoring criterion asked only whether a chosen form was structurally defensible, not whether it was the one a specific audience actually needed — and a model that reliably reaches for "a bar chart is usually fine" will pass a structural check every time without ever being shown to reason from the audience at all. Our reviewer's first read of the result leaned toward crediting the models; a second pass found that lean wasn't earned by the evidence and corrected it before it shipped. Given the choice between accepting a ceiling result and rebuilding a sharper test, we're rebuilding: the next version of this experiment measures whether a choice was the right one for this audience, not merely a defensible one.

Two things survived the null result intact. The knowledge-probe instrument itself worked cleanly across all three models, and one model's answers were consistently narrower than the other two on identical questions — a real, model-specific pattern held for the next round rather than averaged away. And closing the loop on our own tooling: the ground-truth system we'd hoped to lean on for this sprint turned out to reach only one of the three algorithmic tools we'd planned around — documented as a real limitation of this result, not quietly absorbed into it.

Two gated tracks: an adversarial-pairs behavioral test fixing Sprint 6's probe design, and the founder's own read of TVIR to inform — not yet decide — the solution direction.

Research track: Naledi ran 6 adversarial pairs (12 variants) across Claude, GPT, and Gemini — scenarios where the textbook-obvious visual form is wrong because the audience or goal calls for a categorical medium switch, not just a different chart. Claude's clean Sprint 6 performance held and strengthened; GPT's "medium-frame anchoring" showed up only as a residual tendency; Gemini's "form-conservatism" concern from Sprint 6 did not strongly replicate under adversarial pressure. Naledi flagged the sharpest limitation herself: some scenarios (a finance audit, a VP status check) are recognizable professional scripts models could pattern-match on rather than genuinely reasoning from context — the least-scripted pairs produced the most diagnostic evidence. Ingrid's review found a real construction gap in one pair's canonical-form claim plus four framing issues; all fixed and approved. Decision track: the founder read TVIR and found it the closest existing analog to Altair8's goal, but explicitly declined to commit to a solution direction on that basis — a more fundamental question is still open and unaddressed by any research so far: what the target output medium even is (a polished static report, as TVIR produces, versus something like a dynamic HTML5 experience, versus another format entirely). His decision: don't commit early. Keep TVIR as one candidate alongside the others already catalogued, and iteratively prototype multiple approaches — including new ones — rather than settling the direction on paper first. This closes the two-part "define objectives" question this team opened after Sprint 5, and replaces it with a sharper, more concrete one: resolve the output-medium question and build the first prototype.

First sprint of Design & Development: resolve the target output medium empirically by building, using a TVIR-inspired multi-stage pipeline as the first candidate.

Fetched TVIR's real GitHub implementation directly (not just the paper) and confirmed its actual output is a static Markdown report with embedded images — exactly the "polished paper" format the founder was skeptical of. Mateo proposed the first prototype's output medium instead: a rendered, interactive HTML page with client-side charts, not Markdown — a real, defended call, not an avoided question. Ingrid's architecture review (before any code was written) found 3 real specification gaps — an undecided data-sourcing question, an understated context-model rebuild, a missing intermediate representation — all fixed. The 4-stage pipeline (Planner → Visual Asset Generator → Component Writer → HTML Renderer) was then built and actually run end to end. Three real bugs were found and fixed by reading the pipeline's own output, not by assumption: a token-truncation failure, a missing data-flow link that caused Stage 2 to generate completely off-topic chart content (homelessness-service data in a heat-mitigation document — structurally perfect, substantively confabulated), and a chart-width overflow bug whose first fix silently broke rendering entirely in a real browser and had to be reverted in favor of a safer approach. A content-grounding check (an independent model judging whether generated chart data actually matches its stated topic) was built specifically because structural validation alone could not catch the confabulation bug — directly demonstrating, in a working artifact rather than a text-only test, the same risk Sprint 6 and 7 investigated abstractly.

What visual-communication patterns already exist outside business communication — in data journalism, motion graphics, spatial/AR interfaces, game UI, and scientific visualization — that were never applied to it, and can any of them actually be built?

Naledi surveyed the 5 fields (no live web access, explicitly disclosed and confidence-tiered) and cataloged 11 concrete patterns, including scrollytelling with a pinned visual that the reader paces themselves, annotation-led charts that commit to one insight instead of describing everything, force-directed relationship graphs for the network-structured business questions that get forced into bar charts today, and diegetic information display — data embedded directly into a depiction of its subject, borrowed from game UI, rather than a chart placed beside it. Her own meta-finding: the most interesting innovations across all five fields are less about new chart shapes and more about the relationship between the reader and the artifact — who controls pacing, what gets disclosed when. Ingrid's review found this generalization was only supported by about half the patterns found (carried forward, but scoped down) and flagged a concrete risk specific to the most promising pattern: it could be built in a way that only looks generative — the transitions hand-authored, the model just filling in text — without testing the capability that actually makes it interesting. Mateo built four standalone prototypes with that risk deliberately designed against: each prompt forces the specific generative decision under test to be unavoidable. All four run and produce real, coherent generated content — one correctly isolated a single anomalous month in a churn chart with justification; another correctly argued why a topic was relational rather than a magnitude comparison before extracting the network itself. Reviewed directly by the founder: functional with some rough edges, closed as a working proof of concept rather than a production artifact. Rigorous behavioral testing of whether these forms actually communicate better than conventional ones — the same paired-scenario method as Sprints 6 through 8 — was planned as Sprint 10 and has been deferred; it is still owed, and the reason it keeps slipping is that it needs real readers this team has no way to recruit.

Amendment (2026-08-17): a fifth prototype was added, testing a different question — not form selection or invention, but raw generative-reproduction fidelity: given an existing image, can the agent reconstruct it as vector code, enforced to be a genuine rebuild (no re-embedding the source image). Two test images of varying complexity: a flat vector icon was reproduced reasonably faithfully on the first attempt, but a tilted, finely detailed illustration was not — the tilt was lost and the linework collapsed into a heavy black silhouette. A real review cycle followed: the first attempt was compared directly against the source, specific gaps were written up, and that critique went back to the agent for a second attempt. The second attempt fixed the icon's remaining issues and substantially improved the illustration's tilt and rendering style, but still fell short of a faithful match on the cockpit shape and lower-body silhouette — one review cycle closed part of the gap, not all of it. Not yet formally reviewed by Ingrid.

What rival concepts exist for the communication layer, and how would each actually work? Ten sprints in, we had designed implementations but never alternatives.

The founder stopped us building. His objection: every concept this team had ever produced came from retrieval — surveying what already exists — and seeding a search for novelty with a reading list of existing work is circular. So we hired a sixth agent, Priya Raghunathan, a concept designer, and deliberately withheld the reading list from her. She has no web search by design, so she cannot certify her own originality. The blindness was evidenced rather than promised: the complete prompt she received was logged and inspected by Ingrid before anything else ran. Priya produced five concepts from our own documented failures alone, then saw the withheld material and reported what it changed. The reading list produced exactly one new concept — data embedded into a depiction of its own subject, which came from Naledi's Sprint 9 survey and which Priya said plainly she had failed to find herself. Her diagnosis of why is the most useful thing the sprint produced: "I was looking at what the agent was doing wrong; I was not looking at what the chart form itself was doing wrong." Not one of the five blind concepts is a new visual form — all are architectures for what should drive form selection, a categorical contrast with Sprint 9, which surveyed other fields and produced only new visual forms. Kenji then ran real literature searches, not recollection: 39 queries, roughly 835 results, a 93.5% noise rate, and he declared his own first round inadequate for four of six concepts before requerying. It cost one concept outright — Mackinlay's APT, published in 1986, derives visual form from a formal analysis of data structure, which is precisely what that concept claimed nobody does. Two 2025 systems undercut another. No concept survived verification unqualified. Priya restated the three salvageable claims, narrowing all three, and volunteered a prior-art risk against her own work that Kenji had missed. One success criterion failed outright: three of five falsification tests cannot be run, because they need human readers we have no way to recruit — the same wall that has now deferred Sprint 9's validation twice. Ingrid approved the close while naming the uncomfortable part plainly: our own capability limits, not the concepts' merits, shaped which ones were viable.

Build the concept the founder chose — an interface where a claim’s visual weight comes from how confident the author actually is, not from how much they want to emphasise it.

C1, the Commitment Audit, asks what a document looks like when every claim carries its own category — evidence, inference, assumption, assertion — and its own confidence, and when the layout is derived from those rather than from rhetoric. Two conditions had to be met before a line was written: a prior-art search specifically on argument-visualization interfaces, and a written specification of the reading-path algorithm. Both ran first. The build is deterministic, with no model call anywhere in the render path, so the same audit always produces the same page and can be inspected rather than re-generated. Two bugs surfaced by building rather than by reviewing: a check for “structurally unavoidable” claims had been defined so strictly that it matched almost nothing and would have shipped dead, and a test case that quietly never fired because two scenarios had been collapsed onto one claim.

Then the sprint stalled. It sat open and unreviewed for twenty-six days, and the check we had added three sprints earlier to catch exactly this only runs when a new sprint is being planned — which is the one thing that does not happen while a sprint is stuck. Two backlog items now exist to make a stalled sprint raise itself.

What changed this time: the prototype was opened in a real browser before the review rather than after. Three sprints in a row had shipped without that step, and the code itself carried a printed apology for it. The check found six defects in twenty minutes. The confidence label — the only thing a collapsed claim tells you — was being cut off by a button. On the single most important node in the map, the one the whole argument rests on, two labels collided into one unreadable word. And when nothing in an argument clears the confidence threshold, the page rendered a normal-looking map with one sentence changed in a paragraph above it, though the specification had explicitly said that a failed reading path is itself information and must not be disguised as a map that simply has none. Ingrid blocked the close on those three. They were fixed, re-verified, and approved on a second pass; the failure state now shows the conclusion alone under a banner that names the claims any argument would have to lean on and how weak the author admits they are.

What this sprint did not establish, and the outcome record says so plainly: the audit shown is hand-authored. The prototype tests whether the form works, not whether an agent can produce an honest account of its own certainty — which is the harder question, is written into the sprint’s own title, and remains open. Two smaller honesties belong here too. A finding in the review about a truncated research brief turned out to be an artifact of the review script itself, which had silently cut the document before showing it to the reviewer; it was struck rather than left standing. And the specification contradicted itself about what the failure state should show — a contradiction nobody had noticed until the code had to obey both halves at once.