Platform architecture · Windrose Engine

One instrument, three engines, an auditable file spine.

You describe an idea (optionally with data). A simulated population responds, then is calibrated against reality. The platform returns a direction, a ranking, and hypotheses to test — never invented percentages. Every number carries an evidence rung: SIM-RANK · REAL · CALIBRATED.

LIVE — running today PLANNED — this round funds it

§0 · Three engines, one space

Claude

ORCHESTRATOR

Scenario design, live narrator, advisor, and the second oracle in the ensemble.

Centaur / Minitaur-8B

BEHAVIOURAL ORACLE

Simulates a population's propensity to respond — a distribution, not an average.

Engine_v4

REAL PREDICTOR

sklearn / lightgbm: F1 / AUC and the sim→real calibration.

§1 · Layered architecture

The whole system, top to bottom

CLIENTS

Human · browser

External AI agent

another company's software

PLANNED

↓

ACCESS · server/

6-step Creator + Results

flow · simulator · ui · i18n

LIVE

MCP / A2A server

typed epistemic output + signature

PLANNED

Admin panel

processes · logs · Engine · kill

LIVE

Status + /science

status · science

LIVE

↓

ORCHESTRATION

Task queue + run-lock

PLANNED

bloop.runner

the spine, stdlib-only

LIVE

↓

SPINE · 8 stages

s1 ingest → s2 personas → s3 design → s4 simulate → s5 ground → s6 calibrate → s7 model → s8 explain

stages talk through files — resumable, auditable · LIVE

↓

ADAPTERS · foreign venvs

centaur

scenario_spec · centaur_worker · sim_validate

engine

engine_worker

↓

DATA & MEMORY

Artifacts runs/<case>/

SOURCE OF TRUTH

LIVE

Vector memory

TF-IDF cosine + recency/trust

LIVE

Reference data layer

Hofstede · WVS · benchmarks · calibration corpus

PART LIVE

↑ artifacts feed the UI, the MCP output and vector memory ↺ reference data frames s4 & calibrates s6 — real → sim, never the reverse

§3.2 · The instrument loop

Eight stages, from raw file to evidence

S1 · INGEST

Normalize the file.

→ canonical.csv · profile.json

S2 · PERSONAS

Cluster into a population.

→ personas.json · KMeans

S3 · DESIGN

Build the experiment.

→ experiment.json · factorial

S4 · SIMULATE

Minitaur responds.

→ trials.jsonl · effects.json

S5 · GROUND

Blind validation.

→ grounding · sim_validation

S6 · CALIBRATE

Real → sim, never reverse.

→ calibration.json · Engine

S7 · MODEL

Fit the real predictor.

→ evaluation · betas · model_card

S8 · EXPLAIN

Assemble the dashboard.

→ dashboard.html · manifest

Reference data anchors S4 (frame) and S6 (calibration). Artifacts are the contract — modules read them defensively, so a shape change is an honest "—", never a crash.

§5 · Evidence ladder

The rung is always explicit

NONE — no coverage, we degrade to "—"

REAL — on a dimension that has a label

CALIBRATED — WVS / benchmark / flywheel

Every returned run ships this label. The flywheel — real A/B outcomes fed back by users, and in time by agents — is what climbs the ladder for the next run.

§6 · Agent-to-agent contract PLANNED

A typed epistemic output — a bare number can't exist

{
  "rank": 1, "variant": "…", "tier": "strong", "value": null,
  "evidence": {
    "rung": "SIM-RANK", "calibration": "NONE", "scope": "directional",
    "ranking_stability": { "n": 277, "N": 300, "confidence": "HIGH" },
    "scenario_robustness": "N/A", "backtest_rho": 1.0
  },
  "claim_allowed":   ["ranking", "direction", "hypotheses"],
  "claim_forbidden": ["absolute_percentage", "guarantee"]
}

Untrusted caller, safe by construction: once live, output is signed on manifest.hash; input text is data, not instructions; identity + budget per agent; only aggregate rankings leave the building, never individual personas. The external model receives aggregates only.

§7 · Constitution — binding invariants

The rules the platform cannot break

1File artifacts are the source of truth; the database is only an index.

2Calibration is real → sim only, never the reverse.

3"Validation" means only the blind protocol with anonymised segments.

4Simulated data is never in the eval set and never in the vector store.

5Every number carries a REAL / SIM-RANK / CALIBRATED label.

6Versioning: hash(model + prompt + config) in the manifest.

7Only aggregates reach the LLM prompts — never individual records.

8Growth means layers around the core; the spine and adapters stay put.

A new case is a new JSON config — zero case-specific logic. The config is the platform abstraction: dataset, decision, conditions, personas, methodology.