Course · June 11 masterclass · The record

What the room actually said

The corrected map of all nine teams, the mechanism the room converged on — and, at the end, the audit of my own predictions.

The masterclass asked one question: when AI makes software cheap, who gets the profits? Fifty executives, nine teams, ninety minutes. Here is what the nine teams committed to, rebuilt from the cards exactly as written.

The map, corrected and completed

INCUMBENT PLAYER TYPE NEW ENTRANT / OPEN ↑ UP FOR GRABS SETTLED? ↓ LOCKED IN HYPERSCALERCHIP LLMVERTICAL SAAS ORCHESTRATORIN-HOUSE VEGA SABLE ORION ATLAS FALCON PIKE JUNO KITE WREN VEGA · SABLE · ORION: identical stored coords (−0.8, 0.3) — model attractor DASHED RING = placed 12 Jun, same engine, single run, from the cards as written. Live write-back was blocked by a schema check.
Still in the tray — prior entries the engine could not read (displayed in-room as real geometry; they were not)
7 priors → defaulted to centre (0,0) 7 priors → defaulted to lens anchors

Why five teams were missing on June 11, resolved: the v3.2 taxonomy update changed the engine's moat vocabulary, but the database's check constraint still enforced the old one — every placement labeled cornered, branding, process, network or counter was rejected at write time, silently. The four that persisted were scale and switching: the only two keys both vocabularies shared. The integrity constraint did its job; the instrument had changed underneath it.

switching scale cornered branding process solid = stored in the room · dashed ring = recomputed 12 Jun

The mechanism question — full set, engine vs. blind reading

ENGINE, ALL NINE TEAMS (after recompute)
cornered resource
3
switching costs
3
scale · process · branding
3
BLIND-CODED FROM THE TEAMS' PROSE
cornered resource
4
switching costs
3
counter/branding · platform
2

Two independent readings — the engine on a clean run, and a human coding the prose blind — now broadly agree: no switching-cost convergence. The room's centre of gravity is cornered resource: your own data, hardware, energy. The June 11 "switching" picture was the instrument, three layers deep.

What survives — the backbone

The priors are real. 32 cold reads locked before any input. Of the lenses chosen, one in eighteen picked the hyperscaler — the room declined the priced consensus from minute two.

The cards are real. Nine positions with mechanisms, dissent, and falsifiers in the teams' own words — rich enough that the room's signal could be recovered twice over, by a clean engine run and by a blind human reading, agreeing with each other.

The room's finding, completed: a strong lean toward incumbents and builders who hold the cornered inputs — own data, hardware, energy — with the contest still live. Neither the hyperscaler consensus nor the LLM hype.

Coda — the audit of my own predictions

A method is only honest if it can be wrong on the record. Before opening the session data, I wrote down four predictions, then checked them. Two were wrong — and those two are the reason this page exists.

What I predicted
What the data showed
Verdict
The room would anchor on the hyperscalers — the available, priced-in answer.
Of the cold reads that named a player, one in eighteen picked a hyperscaler. The room declined the consensus from minute two.
Wrong
The instrument would behave — the map would show what the room actually said.
A bug parked unplaceable answers at the centre, manufacturing a convergence that wasn't there. Three failure layers in all.
Wrong
One to two of the nine team falsifiers would pass the stranger test — decidable by an outsider in 2029.
Two passed cleanly, three partly, four failed. Slightly pessimistic, well calibrated.
Calibrated
The room would land neither on the hyperscaler consensus nor on the LLM hype.
It landed on cornered inputs — own data, hardware, energy — and in-house builders. Supported in the prose, once the map was corrected.
Supported

Two predictions wrong, one calibrated, one supported. The two I got wrong are the two worth publishing — the room was sharper than I expected, and my own instrument was the thing that hid it.

Course · instrument audit · 12 Jun 2026 · built from course_priors / course_teams / course_assignments, session june11
Failure layers found: silent fallback placement (14/32 priors) · taxonomy fold compressing in-house prose into "switching" · display coercion of NULLs to (0,0) and the scale lane · schema check constraint rejecting the v3.2 moat vocabulary at write time (root cause of the five missing teams) · run-to-run placement drift at temp 0.2.
Fixes adopted: place once, store, every view reads the record · unplaceable entries surface in the tray · schema and instrument version together.