Intelligence
shadow only · advisory
Advisory by architecture, not by policy.
The analytical layer is not trusted to behave. It is structurally prevented from acting: actionable fields are removed by a validator before anything downstream can read them.
The shadow contract
Evidence gets through. Instructions don’t.
A neural gateway can send analysis. If its payload contains a field that looks like an order — a multiplier, a block, an apply — that field is stripped by an allowlist and a warning is logged. Every decision is then forcibly rewritten to Mode="SHADOW_ONLY", Applied=false.
Accepted as evidence
Descriptive fields. They inform an operator; they instruct nothing.
- signal
- confidence
- entry_quality_score
- family
- regime
- model_version
- reason_codes
- data_quality
Stripped on arrival
Imperative fields. Removed by allowlist, logged as ignored_actionable_neural_field.
- action
- reduce_risk
- block
- block_new_entries
- lot_multiplier
- multiplier
- override
- apply
- boost_risk
Payloads claiming applied=true, operational_use=true or shadow_only=false are rejected outright rather than sanitised. The legacy POST /api/ai/apply route is permanently quarantined and always answers blocked.
Health scoring
Every score shows its arithmetic.
Each strategy starts at 100 and loses points against explicit, inspectable rules. There is no embedding, no opaque weighting, no “the model felt uneasy”. Every deduction carries the written reason that produced it, so you can argue with the score instead of merely accepting it.
- ≥ 78 HEALTHY
Performing in line with its baseline.
- ≥ 60 WATCHLIST
Drifting, but within tolerance.
- ≥ 42 DEGRADING
Materially below baseline on more than one axis.
- < 42 CRITICAL
Multiple confirmed deteriorations.
- NO_LIVE
No live evidence to score against.
| Condition | Penalty |
|---|---|
| Live profit factor below 55% of baseline | −26 |
| Allocator already wants a block | −16 |
| 30-day expectancy negative against a positive baseline | −18 |
| Sample under 3 trades, or confidence under 0.45 | −18 |
| Negative expectancy across 7d, 14d and 30d | −14 |
| Loss streak of four or more | −12 |
| No trades in 30 days | −12 |
| Execution friction (spread + slippage) ≥ 25 pts | −8 |
Restraint
Built to under-react.
Most monitoring fails by crying wolf until operators stop listening. Four mechanisms exist specifically to keep that from happening.
0.35–0.92
confidence is clamped at both ends
it can never claim certainty
3 windows
degradation must persist across 7d, 14d and 30d
a single bad week doesn’t escalate
2 cycles
a state must repeat before it is called confirmed
consecutive-state persistence tracker
120 min
cooldown between automated defensive actions
max 2 actions per run, 2 confirmations each
Auto-defense ships switched off. Where the on-premises platform can act defensively, it is disabled by default, requires two confirmations, honours the cooldown, and caps itself. Recommendations built on thin evidence are explicitly flagged provisional rather than presented as conclusions.
Shadow allocation
We show you what we would do, then don’t do it.
Family-level and cross-account allocators publish a target weight beside the current one, with the delta, the implied multiplier, and the reason. You read the counterfactual and decide. Nothing is applied, and there is no button that would apply it.
Rehearsal without a live tick. Eight synthetic scenarios — drawdown pressure, correlated-cluster stress, block failure, execution drift, offline peer, news shock — let an operator exercise the system’s judgement over a closed weekend market. In that mode nothing reaches the backend at all.
{
"mode": "SHADOW_ONLY",
"applied": false,
"operationalUse": false,
"appliesTradingActions": false,
"signal": "degrading",
"confidence": 0.71,
"reason_codes": ["pf_drift", "streak_4"],
"warnings": [
"ignored_actionable_neural_field: multiplier"
]
}
What this is not
The part most vendors leave out.
There is no trained model inside Quantisentry. The scoring engines are deterministic C#. “Neural” refers to an external service you supply, reached through a GET-only proxy. Direct prediction is disabled by default behind a feature flag.
The knowledge base does not learn at runtime. Per-strategy baselines, in-sample and out-of-sample segments, loss maps by hour and weekday, and redundancy relationships are loaded from curated data packs. They are a historical reference, not an online learner.
This layer is on-premises today. The commercial API exposes enrolment, configuration and telemetry ingest. Health, copilot and shadow-allocation surfaces are dashboard capabilities, not contractable cloud endpoints. Treat them as platform capability when you evaluate us.
Language we won’t use. Not self-learning. Not proprietary deep learning. Not trained on millions of trades. Not an AI that trades for you. None of those would be true, and each one would be easy to check.