Altiera

How it works

Running a question

Six steps, from a folder of documents to an answer you have pushed back on. This is the whole loop, not a summary of it.

Step 01

Give it context

Upload whatever you already have: decks, research, exports, policy documents, past campaigns. Altiera reads them to understand your world. None of it leaves Europe and none of it trains anything for anyone else.

Step 02

Ask the question

One decision, phrased so it can be wrong. We start from a decision with a real outcome and a date attached, not a research topic. If the question cannot be wrong, it cannot be simulated.

Step 03

Parameters, proposed

Agents are constructed from observed behavioural histories and calibrated to the population you care about. Altiera proposes the setup from what it read, and every value is visible, sourced, and yours to override. Nothing is decided behind the panel.

Step 04

It runs

The population meets the scenario and interacts. The run repeats until the distribution stops moving.

Step 05

First read

You get the range, the confidence, and which segments break from the median. This is the point where most tools stop.

Step 06

Push back, it re-runs

Ask a follow-up and the population re-runs against it. The agents deliberate again on the new condition and come back sharper. The population persists between questions, so the second answer costs hours rather than a new project.

Altiera consoleStep 01 / 06

Context ingested

pricing_deck_q3.pdf	PARSED
churn_export.csv	PARSED
community_guidelines.md	PARSED
2025_research_readout.pptx	PARSED

4 FILES · 1.2M TOKENS · HELD IN EU

Question

If we raise the mid tier from £9 to £12 in Q1, what happens to retention?

SIMULABLE — OUTCOME DEFINED, DATE ATTACHED

Proposed parameters

Population	30,000 agentsSOURCE: DEFAULT	EDIT
Segments	4SOURCE: churn_export.csv	EDIT
Runs	1,000SOURCE: DEFAULT	EDIT
Horizon	90 daysSOURCE: pricing_deck_q3.pdf	EDIT
Confidence floor	0.70SOURCE: DEFAULT	EDIT

Running

RUN 001 ADOPTED 000 HELD 000

ONE RUN OF MANY. THREE SEEDS, THEN CONTAGION. ~20% NEVER ADOPT, AND THAT IS THE RESULT.

Retention at £12, 90 days

MEDIAN −4.1PP · SPREAD −9.2PP TO −0.6PP

CONFIDENCE: MODERATE-HIGH

SEGMENT: LAPSED-RETURNERS−14PP

SEGMENT: UNDER-25+6PP

Follow-up

If we take the aggressive option, how does it land with the under-25 cohort specifically?

RE-DELIBERATING · 400 RUNS

Under-25s absorb £12 with a median of +2.4pp retention, spread +0.1pp to +4.6pp. The cohort is narrower than the population and moves with its peers rather than on price. CONFIDENCE: HIGH (0.86) SENSITIVITY: RESULT HOLDS UNLESS PEER CHURN EXCEEDS 8%.

ROUND TRIP ON A FOLLOW-UP: HOURS. THE POPULATION STAYS LIVE BETWEEN QUESTIONS.

The engine

The simulation engine. It maintains the behavioural substrate, constructs populations, runs the interactions and calibrates its own confidence. Five capabilities matter to you as a client.

Histories, not labels

Agents carry behavioural histories rather than a demographic tag. Two people with identical demographics who behaved differently are modelled as different people, because they are.

Groups, not averages

Agents influence each other during a run. Contagion, pile-ons, silent majorities and tipping points appear because they are simulated, not because they were assumed.

Confidence, stated

Every result carries a confidence level. A trend-break guard refuses to extrapolate through a level shift, because the most expensive simulation error is a confident straight line drawn through a discontinuity.

Reads the room

Sentiment models trained on generic text misread real communities. Ours arbitrate over model confidence with a hand-built lexicon, and detect code-mixing that off-the-shelf language identification gets wrong.

Auditable and European

Every result traces back to the reasoning that produced it. Data is generated and held in Europe, which is what makes it usable for European enterprise and public-sector buyers.

Deliverables

What you actually receive

Outcome distribution	The range of results across repeated runs, not a single number. Median, spread, and the shape of the tail.
Confidence level	A stated confidence for each result, with the reasons it is lower where it is lower.
Segment breakdown	Which parts of the population diverge from the median, and by how much. This is usually where the decision actually changes.
Sensitivity view	Which variables moved the outcome and which turned out not to matter. Useful for deciding what to test in the real world.
Reasoning trace	An auditable account of how the engine reached the result, for internal review or regulatory questions.
Re-runs	The population persists. Alternatives and follow-up questions are hours of work rather than a new project.
Working session	A session with the people who built the model, walking your team through what to trust and what to treat carefully.