Boardroom · the AI brain · in ViP, AiP and PiP
An adviser that cites its sources and will tell you when you’re wrong.
Most AI assistants agree with you. Boardroom is built the other way, and not by asking it nicely. Anti-sycophancy is an architectural constraint enforced in three layers that no product on the platform can switch off, inside an accountability perimeter that checks its numbers, judges its output every night and keeps its score.
Confidently wrong is the enemy.
A hotel owner does not need another voice that agrees. They need one that says the index is relative and the covenant is not, that a three percent variance is noise, that the number in the pack was extrapolated. Zero sycophancy on adversarial benchmarks is a release criterion for every product we ship, not a target we report.
- 1
The base persona no product can remove
Every HotelStack.ai product inherits a priority-zero layer in its system prompt: never confirm a claim without evidence, correct misconceptions with citations, admit uncertainty, no false reassurance, raise risks unprompted. Products can add rules on top. None can take this one off.
- 2
Every answer is evaluated before you see it
Output passes an evaluation pipeline. Injection patterns and cross-tenant data are blocked outright. A claim about your business without a citation is blocked. Declared confidence that outruns the evidence is flagged.
- 3
Confidence is capped by where the evidence came from
Sources sit in four authority tiers, and the answer’s confidence label follows the tier: Definitive on primary data, Interpreted where reasoning joins two sources, Advisory where the model is inferring. It cannot promote itself.
The accountability perimeter.
The adviser is a model. Everything around it is deterministic, and the deterministic part is where the trust lives. Each row is running today unless it says otherwise.
| Gate | What it does | StatusShipped means running on our own venues today |
|---|---|---|
| Governed metrics | One number, one computer. Every figure Boardroom can say is served from the persisted metric layer with its provenance. The model never queries the database. | Shipped |
| Numeric gate | Each narrated number is re-checked against the row it cites. A number the metric layer did not serve is a number the narration cannot contain. | Shipped · runs on every narration; the block is moving from shadow to enforce |
| Adversarial critic | A structurally separate pass over the output looks for the flattering read, the buried variance, the unearned certainty. | Shipped |
| Nightly judge | A different model from the one that wrote the narration samples live output every night, with a planted error it must catch. If the judge misses the plant, the run fails loudly. | Shipped |
| Claims and outcomes | Every forward-looking sentence is stored as a claim and resolved later against what actually happened. The track record is shown, not asserted. | Shipped |
| Audit log | Every call, every proposal, every approval or rejection, actor and outcome, append-only. | Shipped |
It learns, and it is honest about how.
Self-improving does not mean a model quietly retraining on your data. It means a system that knows what it has not read, keeps its own score, and is held to the same accountability as the advice it gives.
It knows what it has not read
A declared registry of the data domains Boardroom consumes is checked against the schema on every build. A new source that is not mapped, or not deliberately excluded with a reason, fails the build. The coverage gap is a first-class fact, and it is closed by a human approving the wiring.
It is scored, and the score is kept
Claims resolve against outcomes into a calibration band. The nightly judge scores faithfulness. The first live pass on our own venues scored the narration at 29 percent. We fixed it deterministically, with the numeric gate, not with a better prompt.
Advice will be tracked like a prediction
A recommendation becomes an intervention with an expected effect, marked adopted, declined or partial, and resolved only if adopted. Advice efficacy is reported beside prediction accuracy. Roadmap.
The improver is held to the same standard
A loop that proposes improvements to Boardroom itself records what happened to each proposal. An unaccountable improver is worse than none. Roadmap.
The rules it reads by.
Written down, ruled by the people who read the owner’s report, and enforced in code rather than remembered in a prompt.
- Leads with performance against budget, the document the owners signed, and always shows last year beside it.
- Will not inflate a variance under three percent into a story.
- Flags any annualised figure as extrapolated.
- Names the multiplier behind every value claim. A value delta without its cap rate is blocked.
- States which properties a portfolio figure is stated on, and marks anything modelled or partial.
- Proposes; never executes. Every write waits for a person.
Where you meet it.
| ViPThe owner’s canvas and BrewBot | Narrates the month for the operating partner and answers questions from the venue’s own governed metrics, role-scoped. |
|---|---|
| AiPThe Asset Boardroom | Benchmarks, checks covenant headroom, quantifies value impact and drafts the owner’s narrative for the meeting where every number gets challenged. |
| PiPThe portfolio read | One dated paragraph on the whole estate with its receipts, after the deterministic exceptions and never ahead of them. |
| Your assistantThe Portfolio Connector | Six read-only MCP tools so Claude or any MCP client can ask the estate a question and cite the drillable original. |
On the roadmap, labelled honestly.
The perimeter is built. What lives inside it is being built in this order, and each piece ships behind the same gates. If it is in this section it is not shipped yet.
The causal spine
Declared driver trees per root metric with deterministic variance attribution: price, volume and mix splits with the residual always shown. Trees ship draft-labelled until validated against three historical months.
The conversational adviser
Persisted conversations answered through bounded read-only tools, with the numeric gate run per answer over exactly what the tools served.
Watchers and briefs
Declared conditions over streams already computed. A tripped watcher produces a small, fully gated brief. Capped at three per venue per day, operators first, owners after graduation. Quiet is a feature.
Portfolio strategy synthesis
Playbook transfer from measured interventions and capital-allocation prompts ranked by addressable loss. Last by design, because it composes everything above.
Bring us a month of your data.
The fastest way to judge Boardroom is to let it narrate a month you lived through, and see whether it tells you something true you had not noticed, or agrees with you.
Talk to us