Sister-LLMs: ensemble reasoning over a shared ontology

A single LLM asked for a probability gives you a number shaped by one model’s biases, one context window, and one set of training data. It is a strong opinion delivered with confidence. We did not want a strong opinion. We wanted a calibrated distribution — a thing that says not "70%" but "70%, and here is how much I disagree with myself about it." That is what the five sister-LLMs produce.

Five personalities, one ontology

Each sister is a personality — analyst, contrarian, disruptor, historian, institutionalist — loaded from a TOML file at runtime, no recompile to change one. They are not five random seeds of the same model; they are five deliberately different lenses, each with its own prompt version stamped on every run for reproducibility. What makes them an ensemble rather than five separate answers is that they all reason over the same versioned GNL world model. They disagree about interpretation, not about the facts.

Merge, don’t vote

The easy wrong answer is to vote — take the median of five numbers. Voting throws away the disagreement, which is the signal. The Oracle merges the sisters’ outputs into a normalized Ensemble: probabilities sum to one, scenarios sorted descending, entropy in nats. Normalization lives in exactly one place (deino-oracle) so every consumer can rely on the invariant. The shape of the disagreement — high entropy, low entropy, bimodal — is the thing a CRO reads to know whether the forecast is sharp or spread.

Voting throws away the disagreement, and the disagreement is the signal. Merge into a distribution; let the entropy speak.

Why a shared ontology is the whole trick

An ensemble of five LLMs each reasoning over a private, unaligned set of facts is five forecasts of five different worlds. The GNL is what makes them five forecasts of the same world: every sister reads the same NodeVersions, the same version of the graph. The merge is meaningful because the inputs are identical. That is also why the manifest records which NodeVersions were read — so the ensemble is reproducible, not just rerunnable.

• Personalities are data (TOML), not code — edit a lens without recompiling.
• Prompt versions are stamped per run — the ensemble that produced an output is reconstructable.
• Normalization happens once, in deino-oracle — sum(probability) ≈ 1.0 is a guarantee, not a hope.
• Entropy is the honesty field — it tells you how spread the disagreement is.

Calibration is the test, not the promise

An ensemble is not calibrated by construction. It is calibrated by measurement: shadow runs against a realized timeline, scored with Brier and log-loss via deino-eval. Gate 1b requires the shadow run to beat a disclosed baseline on one forecastable series before any forecast side is positioned as accurate. The sisters are the inference layer; calibration is the loop that earns the word "oracle." We are pre-revenue and building the POC; the sisters run, and the calibration loop is inert until it is filled.