Workspace · Guardrails

The limits are the product

An agent platform without hard limits is a liability for a SEBI-registered advisory. Four of the guardrails below are locked at the runtime level — they cannot be raised from this screen, by anyone, including DeployOne.

Locked guardrails

Tunable

Fully autonomous agents

of 15

Archive retention

5 years

immutable

Guardrails

Locked rows are enforced in the runtime, not in this UI

Guardrail	Setting	State
Money moves above ₹2L	Human approval required	locked
Any member-facing asset	Compliance Guard must pass it	locked
Outbound voice window	9am–8pm IST, DNC honoured	locked
Discount authority	Agents may not discount — extensions only	tunable
Autonomy ceiling	Support Resolver + Academy Tutor fully autonomous	tunable
Model routing	Reasoning model for advice-adjacent, fast model for ops	tunable

Autonomy per agent

Raising a ceiling needs two weeks of held eval score

@stoxoStoxo Deskeval 91

@guardCompliance Guardeval 96

@keepRetention Sentineleval 84

@voiceVoice Conciergeeval 79

@resolveSupport Resolvereval 88

@activateActivation Piloteval 83

@forgeCreative Forgeeval 87

@spendMedia Buyereval 85

@radarMarket Radareval 89

@coachAdvisor Coacheval 86

@scoutCreator Scouteval 81

@b2bStoxo B2B Stewardeval 92

@tutorAcademy Tutoreval 88

@factoryResearch Factoryeval 90

Model routing

Chosen per task, not per vendor loyalty

Advice-adjacent

Reasoning model with citations enforced and a refusal boundary

Operations

Fast model, tool-heavy, cost-optimised for volume

Voice

Realtime speech with Indian-language pronunciation tuning

Creative

Multimodal generation on the StockGro brand kit

Compliance

Policy-tuned model with the rulebook in context

Evals

A separate judge model, never the model being scored

Investments in securities markets are subject to market risks. Every member-facing output in this OS is checked by Compliance Guard before it ships.