Brennelhove AI applies predictive modelling to market data in real time, producing allocation recommendations that are logged and published for independent review. No retrospective edits. No selective disclosure.
Each recommendation passes through four discrete stages. The process is fixed; only the input data changes.
The model consumes order-book depth, volatility surfaces, and on-chain flow data from multiple independent feeds, cross-checked for discrepancy before use.
Candidate positions are scored against historical pattern libraries and current volatility regimes, producing a probability-weighted expected return for each.
Position sizing is capped by pre-set drawdown and exposure limits. No recommendation can exceed these thresholds, regardless of model confidence.
Approved decisions are executed and written to the public log within the same reporting cycle, timestamped and unalterable.
Figures below reflect the current reporting period and update on each logging cycle. Historical periods remain accessible in full.
The verification badge confirms that each entry in the log has been cross-referenced against exchange-side execution records within 24 hours of posting. No entry is published without this check.
Log VerifiedMitigation is structural, not advisory. The following parameters apply to every recommendation without exception.
No single asset position may exceed a fixed percentage of total allocated capital, set independently of model confidence scores.
If cumulative drawdown within a rolling window breaches a set threshold, new positions are suspended pending manual review.
Position size is reduced automatically as realised volatility rises, independent of the underlying signal strength.
We do not publish selectively. We do not amend figures after the fact. Every input, decision, and outcome used to evaluate Brennelhove AI is drawn from the same record available to the public.
Governance detail: the logging infrastructure is maintained independently of the modelling team, with write access restricted to an automated pipeline. Manual intervention in the log requires a recorded override and is itself logged as an event.
Training data consists of historical order-book data, volatility series, and on-chain transaction data spanning multiple market cycles. No forward-looking or non-public data is used.
The underlying model is retrained on a fixed quarterly schedule. Each retraining event is dated and logged, with the prior model version retained for comparison.
No. Entries are append-only. Corrections to erroneous entries are recorded as new, dated entries rather than alterations to the original record.
Drawdown circuit breakers and volatility-adjusted sizing apply automatically. If conditions exceed defined thresholds, new position entry is suspended until manual review is complete.
No. Price and liquidity data are aggregated across multiple venues to reduce dependency on any single exchange's data quality or uptime.
Account data and API credentials are held under restricted access controls. No capital is transferred during initial integration.