Pages
Inventory Landing — Estate Bake-off

Source: qdash /inventory landing — c100 bake-off #3, 2026-09-30

One question — what is our AI made of, and how concentrated is it? — drawn four ways on the same fixture world: an alluvial of the supply chain, a provider bowtie, a harness × provider switchboard, and Mark's c68 body figure filled one pixel per asset. The blind spots (detected-only installs, the shadow Hermes, personal accounts) get a floor width in every form so they never vanish on a volume lens. Click any harness, provider or model to focus it everywhere.

Lens
Period
Scale
State
Where does the work flow? · tokens in the period · detected-only installs have no telemetry, so they sit at a floor width
7 harnesses · 6 providers · 12 models — 26 installs unseen (9 Hermes), 2 personal accounts
Form 1

Alluvial — harness → provider → model

Strong at the join itself — which harness reaches which model through whom, and how concentrated it is, as one image; blind spots ride along as dashed ribbons.Weak at exact reading (widths are approximate) and crossings on the left once harnesses multi-home; needs the fold rule to survive real scale.
HarnessProviderModel2 unseen4 unseen7 unseen4 unseenClaude Code1.2BCursor670.1MCodex258.1MCopilot259.8MWindsurf132.8MGemini CLI40.2MHermes Agent9.6MAnthropic1.7BOpenAI513.8MAzure OpenAI193.7MGoogle135.1MMistral59.7MOpenRouter9.6Mclaude-sonnet-51.0Bclaude-opus-5479.8Mclaude-haiku-4.5187.9Mgpt-5365.0Mgpt-5-mini102.0Mo446.8Mgpt-5 (Azure)145.3Mgpt-4.1 (Azure)48.4Mgemini-3-pro74.2Mgemini-3-flash60.9Mcodestral-359.7Mhermes-4-70b9.6M
Cross-foot: harnesses Σ 2.6B = providers Σ 2.6B + unseen stubs (17 installs, no tokens) · models Σ 2.6B
seen on the wire — width = lensdetected-only stub / shadow ribbon — floor width on tokens & spendpersonal-account share (tokens; people on installs)unknown accountproviders by weight · harnesses by barycentre · models grouped by provider
Form 2

Provider bowtie

Strong at reading one provider — who routes in, what it serves — with zero crossings at any size; the natural provider-detail header.Weak at the harness → model continuity is cut at the spine; linear wings shrink small providers to slivers; reads as a bar table.
← harnesses routing inprovidermodels it serves →CursorClaude Code1.7BAnthropic1.7Bclaude-sonnet-5claude-opus-51.7BCodexCursor513.8MOpenAI513.8Mgpt-5513.8M193.7MAzure OpenAI193.7M193.7M135.1MGoogle135.1M135.1M59.7MMistral59.7M59.7M9.6MOpenRouter9.6M9.6MNo telemetry17 inst.no telemetry — provider and model unknown
Both wings share one scale (1.7B = full wing). Left wings Σ 2.6B · right wings Σ 2.6B — the gap is the no-telemetry row, which has no right wing.
segment = one harness (left) or model (right)detected-only / shadow — 6px floorstrip under the provider = personal-account share
Form 3

Switchboard — harness × provider

Strong at scale and precision — rows and columns cross-foot, personal accounts land on the exact pair, 30 × 8 fits the same shape.Weak at the model hop is only a header strip; it is a grid you scan, not a picture you see — the least "30 ft" form.
Anthropic1.7BOpenAI513.8MAzure OpenAI193.7MGoogle135.1MMistral59.7MOpenRouter9.6MNo telemetry17 inst.model unknownClaude Code1.2BCursor670.1MCodex258.1MCopilot259.8MWindsurf132.8MGemini CLI40.2MHermes Agent9.6M92474Σ columns1.7B513.8M193.7M135.1M59.7M9.6M17 inst.2.6B
dot area = lens (seen)hollow = detected-only / shadow, count beside itring = personal account on that pair (n if more than one)column strip = that provider's models · row and column totals cross-foot to 2.6B
Form 4

Pixel body (c68, Mark's figure)

Strong at the census — every class of asset, one pixel each, blind ones hollow; the brand voice and the c68 card in one; the only form with MCP, skills, tools, workspaces.Weak at concentration and joins — it says what the estate is made of, not how it flows; pixel size varies per limb.
HeadModels124 new & unvetted
Left armMCP servers166 unvetted
TorsoHarness installs26826 unseen
Right armSkills93 unvetted
ForearmTools23035 from unvetted servers
HipsWorkspaces140no vetting signal yet
FeetProviders62 new or no org account · 1 with personal

Click a row to isolate the limb; click a pixel to focus its harness, model or provider across all four forms. Pixels are one per asset within a limb — each limb packs its own count, so pixel size is not comparable across limbs.

Ruling to make — feet = providers, “what it stands on”. Machines left Inventory in c21, which orphaned the c68 legs, and providers had no body part. This draws the legs out entirely: hips (workspaces) stand on the providers. Two smaller asks ride along: the torso counts installs, not the 7 harness kinds, so the blind installs have a pixel each; and the head lights for new unvetted models (c68 kept it calm until asset_vetting exists — the fixture assumes it does). Hips stay calm: no workspace vetting signal.
seen & declared / vetteddeclared, idle in the fixture (never called / invoked)hollow = unvetted or detected-onlydashed = shadow (Hermes, OpenRouter)rose inset = provider with personal accounts
Read

My read

The alluvial is the hero. It is the only form that draws the whole join — harness to provider to model — as one image, and that is the thing no child page can show. At this estate's size (seven harnesses, six providers, twelve models) the barycentre order keeps the crossings to a handful on the left and none on the right, because every model belongs to exactly one provider. The c67 “loom” warning holds at 34 people-ribbons, not at twenty-odd class ribbons. The fold rule (top N models + others, blind spots never fold) is what keeps it honest at 1,500.

The pixel body is the census, not the flow. It is the only form that shows MCP servers, skills, tools and workspaces at all, and the brand's pixel voice makes the blind spots read as missing pixels rather than a percentage — Mark's “5 blind installs, not 87%” made literal. It says nothing about concentration, so it belongs on the Home Inventory card (its c68 home) or as the band under the hero, not instead of it.

The switchboard is the scale answer and the child-page answer. It cross-foots exactly, it takes 30 harnesses × 8 providers without changing shape, and the personal-account ring lands on the precise pair. It loses the model hop (a strip in the header is a hint, not a join), so it belongs on /inventory/providers. The bowtie is the provider detail page's header: one row, both wings, no crossings — but as a landing it is a bar table, and its linear wings make Mistral and OpenRouter slivers.

Judge

What to try, and what to look for

Flip the lens to Installs

The no-telemetry band grows from a floor to its true width, and Hermes gets thicker than its token trickle. Which form makes 26 blind installs feel like 26?

Flip the scale to 1,500

Models fold to the top 7 + others; hermes-4-70b never folds. Watch the left-side crossings in the alluvial and whether the switchboard dots still read against a 358-install no-telemetry column.

Flip the state to All clear

Every amber mark leaves; Hermes and OpenRouter vanish. Does each form look finished, or just emptier? The body should read as one calm grape figure.

Flip the state to First run

The demo: the forms build in, the verdict says "first sweep found…", and the no-telemetry band is the sensors not yet reporting. Is this the moment a prospect says "holy…"?

Click Claude Code, then Anthropic, then hermes-4-70b

Focus travels across all four forms (the body lights the harness installs, its models and its provider). Which form shows the footprint of one thing best?

Change the period

Tokens and spend follow the pill; installs do not — inventory is not period-scoped. At 24h the alluvial thins but keeps its shape; the body does not move at all.

Qpoint Brand Style Guide