Deck work fails in layers: the visual layer (untitled charts, flat images), the structure layer (missed requirements), the message layer (long, unspecific titles), and the evidence layer (claims the cited source does not support). A decision-grade record keeps them separate and gates on the evidence layer, because a beautiful deck with unsupported claims is a liability with good formatting.

This report reads a July 2026 third-party assessment as dated discourse: two systems deep-tested on three briefs, six decks each, thirteen measurement families, scored against a style library of strategy-firm decks. A style library is a production key, not a truth key. Where this report derives per-deck edit floors from published slide rates, the arithmetic is labeled modeled and inherits the three-brief scope. Their cells never become our measurements.

The published cells: 84.7% of Perceptis charts carry a title and 82% a unit signal, against 50% and 50%; 3.5 native charts plus 4.5 editable diagrams per deck against 1.5 charts and no diagrams; deck-level visual consistency 97.2% against 94.4%; template adherence 83.3% against 75%. Neither system produced axis titles — a shared category gap. This is the layer a junior designer repairs in an hour, and Perceptis largely ships it repaired.

Structure is parity: requirement coverage 94.9% against 94.8%, conformance 97.6% against 96.3%, content delivery 93.8% against 95.2%, slide count 100% for both. A 0.1-point spread on three shared briefs is parity by construction. Following the brief is the category's competence floor — it decided screening — and both survivors clear it together. Nothing here differentiates.

Message economy reverses the picture: concision 54.4% against 85.2%, presentable density 55.6% against 79.8%, internal consistency 58.3% against 91.7% on a six-deck directional read. The bake-off's own phrasing is exact: almost everything lands one edit short of tight.

Slide-level sourcing is a formatting discipline: Perceptis puts a source note on 89.8% of slides against 65.3%. Claim-level support is an evidential discipline: whether the note's source actually supports the claim. The bake-off reports Perceptis lower on it, with spread across briefs too wide to publish a number. Argument soundness — do conclusions follow without overclaiming — is 73.1% against 84.1%, roughly three slides of twelve overclaiming against two.

A deck that is 90% source-noted while its claim support will not stabilize is worse than it looks, because the notes signal verification that has not happened. In this report a gate whose measurement does not stabilize is not established — for both systems — and the operational consequence is mandatory: every generated deck gets a full claim-check pass before a partner sees it, which caps what the generation layer can save. The one bright evidential cell is honesty: when Perceptis goes beyond its sources it labels the addition 92% of the time.

Resolution bounds the rest. Three briefs, six decks, seventy-two slides per system, with briefs sharing source packs, so deck-level observations cluster. A 100% slide-count record on six decks has an exact 95% floor of 54%. The internal-consistency gap rests on six decks a side and is directional, as the bake-off itself labels it. Deltas smaller than several points on this design are noise, which is why this report treats structure as parity rather than a 0.1-point lead.

The failure mode that matters for Perceptis is treating a production-mechanics lead as partner-readiness. A deck can ship titled charts and still fail a claim-to-source check.

Mapped onto this report's comparison set, the bake-off ran the vendor surface and one raw-model add-in cell, screened out two others, with no human-build cell and no isolated comparisons. Thirteen families against a style key measure production quality well. They cannot say a deck's content is true, and they cannot attribute the visual lead to anything separable from the renderer. The pairs that would price the product: the same briefs with a compression pass bolted onto each system; a claim check bolted onto each; and an analyst-plus-template cell for the human floor.

Twelve slides per deck. At the published rates (modeled), a Perceptis deck arrives with about 5.5 slides failing concision, 5.3 failing density, and 3.2 overclaiming; the comparator's with about 1.8, 2.4, and 1.9 — but also with 1.5 visuals against 8, so its repair is building six or more visuals rather than trimming prose. These are floors derived from artifact rates: they exclude the claim-check pass, which the unstable support measure makes mandatory for both systems.

The buyer estimand is reviewer minutes to a deck the partner will present, including the claim-check pass, at all-attempt cost. No system's minutes are measured anywhere. A compression pass and a claim check are cheap harness layers; whichever side ships them first erases the other's published lead. That is a reason to price the category's direction of travel, not the current artifact gap.

The screening result is the most structurally loaded cell in the bake-off: the incumbent bundle could not carry a brief and source pack to a finished deck in July 2026. Today the entire moat is that Perceptis does the end-to-end job reliably at all. That is an execution lead, dated and version-specific. PowerPoint is the customer's surface; Microsoft and the model labs ship into it by default, and the comparator that tied or beat Perceptis on the message and evidence layers is an add-in inside it.

There is no exclusive corpus: briefs and source packs are the client's, and the style key is the industry's own deck corpus. Vendor materials describe editable PPTX output, template support, add-in, API, MCP, and managed quality-control features. No export test establishes portability, services dependence, or defensibility.

Current marketing and posted terms appear to describe different training rights. Without an executed DPA, lawful reuse remains unknown. Confidential client material should remain out of the product until updated terms and plan-level behavior resolve the conflict.