THE METHOD
How it works
An advisor who always agrees with you is the risk, not the help. CounterBrain puts every decision through structured, repeatable, verifiable challenge.
- 3 isolated passes
- 7 output sections
- Δ-CSI 0–100
- fail-closed
An advisor who always agrees with you is the risk, not the help. CounterBrain puts every decision through structured, repeatable, verifiable challenge.
Language models tend to please whoever prompts them. For a high-stakes decision that is the most dangerous property: they confirm what you already think. CounterBrain is built for the opposite — its job is to attack your thesis, not endorse it. A flattering output, to us, is a defect.
Every analysis goes through three isolated passes. Separating the roles is what makes the challenge rigorous instead of complacent.
A generalist analyst produces the dominant reading: what a good analyst would conclude. It is only the starting point — not the contradictor.
An isolated pass, with no mandate to be balanced, attacks the thesis: it surfaces implicit assumptions, builds the counter-intuitive scenario, proposes falsification tests and questions that raise the burden of proof.
It calibrates confidence in light of the challenge, declares what is not known, and cites sources with their provenance. Without flattering.
The three passes can run on models from different providers: independent, uncorrelated blind spots. A single model checking itself tends to repeat its own mistakes; distinct models expose each other's.
The output structure is fixed: the voice changes by sector, the schema does not. Seven elements, always.
The Δ-CSI is a number from 0 to 100 that measures how much pressure the challenge put on the assumptions of the starting thesis. It is an honest proxy for the intensity of the challenge — not a measure of the decision's quality nor of the actual change in judgment. It is computed deterministically and explainably from five signals: surfaced assumptions, falsification tests, burden questions, sources, and the divergence between thesis and synthesis — measured semantically (cosine between embeddings), not by word overlap.
On an acquisition decision: 3 assumptions surfaced, 2 falsification tests, 2 burden questions, 1 source and a high thesis↔scenario divergence → Δ-CSI = 54.23 ("moderate pressure"). The numbers come from the same engine that runs in production.
If the challenge is empty, the engine retries harder; if it still doesn't hold, it fails. It never emits an analysis without challenge.
When it cannot produce a valid, complete result, it stops and says so — instead of returning something plausible but fabricated.
Every source carries its origin. Without access to documents or search, the engine does not invent sources: it returns an empty list.
We apply the contradictor to ourselves too.
CounterBrain is a cognitive-contradiction infrastructure: instead of agreeing with a decision, it stress-tests it. For every decision it produces implicit assumptions, a counter-intuitive scenario, falsification tests, burden-of-proof questions, a calibrated confidence, cited sources and the Δ-CSI. It is a contradictor, not an oracle.
The Δ-CSI (Cognitive Sovereignty Index, delta) measures the intensity of the contradiction the decision was subjected to: how much pressure the challenge put on its assumptions. It is computed, stored and calibrated for every analysis, so the intensity of the contradiction is verifiable over time.
Yes. CounterBrain is built for on-premise or air-gapped deployment: nothing leaves your perimeter. The engine sits behind a model-agnostic gateway, so it never depends on a single model provider.
Generic assistants tend to please the user; in CounterBrain flattery is treated as a defect. The output follows a fixed, anti-sycophancy schema and every analysis comes with the Δ-CSI and its cited sources.
It works on any high-stakes decision (generalist-first). Sector- or client-specific tuning happens through configuration "Packs", not by forking the code.
Yes. The Sovereign Decision Ledger records decisions, challenges and outcomes, isolated per tenant and local-first: it is the memory that powers calibration.
In the demo you watch the three passes stream live, the Δ-CSI breakdown, and the provenance of every source.