2026-05-01·Glasshouse

Glasshouse enters closed alpha.

Glasshouse rehearses decisions against a simulated public. M&A, capex announcements, crisis responses, policy packages, activist plays, GTM repricings, anything where the cost of being wrong arrives faster than the time it takes to course-correct. Today it opens to a small cohort of design partners.

Creative Human AI · 1 May 2026

We have been running Glasshouse internally for eighteen months. Every public decision Creative Human itself has made: pricing, positioning, two leadership pivots, went through it before it went out the door. The platform is also the engine behind five public-record rehearsals already on file: Alphabet's Q4 capex framing, the UnitedHealth crisis response, Seven & i / Couche-Tard, Broadcom-VMware, and the German Kraftstoffmaßnahmenpaket. Each was constructed from material available before the event being rehearsed; each output is canonically pinned.

Closed alpha means: ten teams, hands-on co-design, named-partner support. The product is not waitlisted-and-forgotten: every alpha team gets a Creative Human engineer attached for the duration, and every run feeds the calibration loop that the next batch inherits.

What a rehearsal produces

One Glasshouse run is five stages. Each stage produces an artifact the next stage depends on, and each artifact is inspectable in the report.

  1. 01

    Seeds

    The decision document is parsed and typeset, the inputs the rest of the run will be evaluated against.

  2. 02

    Graph

    Stakeholders, their relationships, their leverage points. Built from the decision and the public record.

  3. 03

    Personas

    Each stakeholder is given an inspectable persona: disclosure, priors, constraints, vocabulary.

  4. 04

    Simulation

    Reactions are sampled across the cohort. Narrative-share bands form; minority views surface.

  5. 05

    Report

    Verdict, confidence band, pre- and post-decision matrix, pushback phrases, and the cohort coverage audit.

The report is the deliverable. Five surface tile types make it readable: here is what each one looks like, drawn from one of the rehearsals on file.

Decision Matrix

The first tile a decision-maker sees. Three to five framings, each scored against the simulated cohort, ranked by confidence band. The recommended option is not always the highest-confidence one: sometimes the answer is "the second-best option, because the highest-confidence one has a minority view we cannot afford to ignore."

Tile · decision matrix · Alphabet Q4 announcement-sim · n = 240 personas

How should the Q4 capex increase be framed?

"Reinvesting through the cycle"High · 0.71
"AI-led capacity build"Med · 0.54
"Catching up to peers"Low · 0.22 · destructive minority view

canonical run · runs/exp-recover-sprint-g/run-1pin SHA 4f7c3a2

Persona pull-quote

Each tile drills down. Tap a persona, get the verbatim: the line that persona would write, in their voice, about the option in front of them. Useful when the decision-maker needs to sense-check the simulation's read against a stakeholder they actually know.

Tile · persona pull-quote
Sell-side analystPersona #042Tech sector · large-cap coverage · 11 yrs · skeptical of capex narratives

If they cannot point to which workloads the spend underwrites by Q3, this is going to read as discretionary. The market has a tolerance for ambition; it does not have a tolerance for vagueness about return on it.

Confidence verdict

The headline output. One sentence, one confidence value, the exact sample size that produced it. Glasshouse refuses to round. If a verdict is HIGH at 0.71, it says 0.71.

Tile · verdict + confidence band · final verdict · announcement-sim

"Reinvesting through the cycle" is the option the simulated cohort accepts as a credible framing of a costly decision.

Confidence 0.71 · Highn = 240 · 13.4 min runtime

Pushback phrases

The other half of the report. The simulated cohort produces real-language objections, the specific sentences the most adversarial readers would write. Useful for IR scripts, prep documents, and pre-mortems. If a phrase shows up here, it is going to show up on the call.

Tile · pushback phrases · top 4 of 18 · phrases that would land hardest
  • "This is a confession, not a strategy."frequency 0.18 · sell-side analyst cluster
  • "Where is the unit economics on this AI capex?"frequency 0.14 · long-only PM cluster
  • "They are buying optionality at a price the cycle does not justify."frequency 0.11 · macro hedge cluster
  • "Reinvesting through the cycle is the same answer Meta gave in 2022."frequency 0.09 · sell-side analyst cluster

What closed alpha includes

  • One named Creative Human engineer attached for the duration of the engagement, not a support queue.
  • Up to twelve scenario rehearsals over the alpha window, with custom persona libraries per cohort.
  • Methodology disclosure on every run: calibration history, persona provenance, confidence-band inputs, written for the security and audit reviews enterprise buyers actually run.
  • Direct pipeline into the platform roadmap. What alpha partners need next is what we ship next.

Who it is for

Three kinds of decision-makers, mapped to the buyer archetypes Glasshouse was built around. CFOs and IR leads stress-testing announcement framings. CCOs and General Counsels war-gaming crisis response language. Hedge-fund event desks and policy ministries rehearsing decisions where the public record is the adversarial reader.

The methodology disclosure stays. The output is post-cutoff. The personas are inspectable. If you cannot show the work, do not show the answer.

Alpha is by application. The gate is fit, not seniority: we want decisions that are unambiguous, time-boxed, and recordable. Apply via the contact form and reference "alpha"; we will reply within two business days.