Latent Spark Calibrated production and market understanding for the game industry.
Subtle Beacon — Capabilities
Everything your research team needs
to run continuous experimentation.
Subtle Beacon brings live experimentation, preference research, fielding, market simulation, and governed automation into one game-focused system—while preserving the statistical methods, controls, and integrations research teams need to trust the result.
- 01
Run rigorous live experiments
Design, operate, and monitor Bayesian experiments in the live game without giving up guardrails, diagnostics, or control over stopping and allocation.
Experiment types & measurement
- A/B and multi-arm experiments
- Conversion, count, rate, and ordinal metrics
- Observational comparison and KPI-monitor studies
- In-game variant assignment
- Exposure and outcome ingestion
- Hybrid studies combining stated preference with live outcomes
- Portfolio-level KPI monitoring
- Experiment lifecycle events: register, tick, stop, and decision reached
- Idempotent event handling
Adaptive allocation & stopping
- Thompson Sampling allocation
- Allocation rails that preserve exploration
- Adaptive traffic allocation
- Automated tick cycles for live experiments
- Posterior-driven stopping
- Probability-of-superiority rules
- Expected-regret thresholds
- Credible-interval-width goals
- Superiority, uplift, regret, and guardrail decision rules
- Non-inferiority rules
- Peeking mitigation
Experiment quality & safety
- Guardrail metrics and thresholds
- Sample-ratio mismatch detection
- Drift detection
- Drift-triggered follow-up studies
- Probability-of-best and expected-regret readouts
- Holdout validation
- Statistical release gates
- Web SDK and Unity integration
- Remote-config and feature-flag integrations including PlayFab, Statsig, Firebase, and LaunchDarkly
- 02
Use the right research method
Choose the method that fits the decision—from feature prioritization and pricing to bundles, concepts, messages, and behavioral preference.
Conjoint & discrete choice
- Hierarchical Bayes conjoint using HB-MNL
- Standard and adaptive conjoint design
- ACBC with screening, build-your-own, tournament, and calibration phases
- Partial-profile designs
- Prohibited pairs
- Dual-response none
- Balanced overlap
- Individual- and population-level part-worths
- Attribute importance modeling
- Willingness-to-pay modeling
- Segment and cohort preference cuts
Ranking, preference & concept methods
- MaxDiff / best-worst scaling
- Standard, sparse, express, and tournament-wave MaxDiff designs
- Pairwise preference studies using BTL, Thurstone, and hierarchical Bayes models
- Gaussian-process preference learning
- Menu-based choice studies
- Rating studies
- Constant-sum studies
- Concept testing
- Message testing
- Feature tradeoff studies
- Behavioral and revealed preference from telemetry
Adaptive study design & validation
- Bayesian optimal experimental design
- Expected-information-gain task selection
- Adaptive conjoint design batches
- Adaptive and bandit task selection for MaxDiff
- Adaptive task allocation
- Living posterior per study
- Respondent quality checks
- Holdout validation
- Convergence diagnostics
- Statistical release gates
- Learned and amortized design policies where configured
- 03
Reach players through the right channel
Field research through panels, communities, owned audiences, surveys, playtests, and live products without rebuilding the study for each channel.
Hosted & external fielding
- Hosted Subtle Beacon surveys
- Rating, menu-based choice, constant-sum, and ACBC respondent sessions
- Tenant-branded surveys and custom domains
- Survey-vendor fielding including Qualtrics, Typeform, and related providers
- Gamer-panel fielding including Helpfull and related providers
- Discord and community studies
- CRM hybrid fielding through HubSpot and Salesforce
- Email-list studies with customizable invitations and branding
- LLM interviewer fielding
- Manual CSV import and export
Behavioral & product evidence
- Live-game exposure through SDKs and remote config
- Product analytics outcome connectors
- Warehouse and event-pipeline connectors
- Store, marketing, and attribution outcome sources
- Playtest builds with behavioral follow-up
- Hybrid A/B and stated-preference studies
- Multi-binding orchestration across several evidence sources in one study
Fielding operations & review
- Respondent-facing task previews
- Copy and media previews
- Durable stage → review → launch approval workflow
- Dynamic fielding-provider registration
- Tenant approval of external providers
- Copy and media upload
- Leak and sensitive-content review
- Composite approval covering copy, channel, spend, and public exposure
- 04
Keep every study grounded in the game
Carry the title, audience, decision, catalog, and operating context through design, fielding, analysis, and reuse.
Game & decision context
- Persistent workspace per game or product
- Subject, population, and catalog anchors
- Decision frame for every business question
- Study question, candidates, and rationale captured at intake
- Subject registry and game disambiguation
- Audience profiles
- Fielding-audience recommendations
- Product claims with inferred vs human-reported provenance
Iridae context enrichment
- Latent Spark mechanic beliefs, embeddings, and competitive context
- Patient Cartographer pathway and tactic metadata at study intake
- Context threaded through study design, fielding, federation, automation, and decision outputs
- Operational context attached to compiled study specifications and result payloads
- 05
Turn evidence into decisions
Move from model output to a defensible production, product, marketing, pricing, or portfolio decision.
Research readouts
- Winning option and confidence
- Probability each option is best
- Expected regret of acting now
- Credible intervals on outcomes
- Segment-level preference differences
- Attribute importance estimates
- Part-worth estimates
- Willingness-to-pay ranges
- Guardrail status
- Drift and anomaly alerts
- Rendered charts and VizKit visual artifacts
- Plain-language posterior readouts
- Plain-language decision briefs
- Recommended next study
Market & commercial simulation
- Choice-share simulation using softmax and first-choice rules
- Segment-level scenario comparison
- Pricing sensitivity analysis
- Bundle optimization
- Source-of-volume analysis
- TURF-style reach analysis
- Revenue and profit tradeoff analysis
- Counterfactual scenarios
- Dynamic multi-period adoption and churn scenarios
- Competitive-response scenarios
- Roadmap what-if simulations
- Sensitivity grids for price, content, and feature levels
- Industry-standard export formats where configured
Diagnostics & follow-up
- Diagnostic proposals with hypothesis, bindings, cost, and success criteria
- Decision briefs tied to the originating alert or business question
- Recommended next study on investigation threads
- Federation prior-selection explanations
- Explain why previous studies were selected to inform a new prior
- 06
Detect what the team did not ask
Watch for drift, anomalies, and disagreement between evidence channels, then propose the research needed to resolve them.
Portfolio sensing
- Scheduled portfolio health scans
- External KPI registry and ingest through HTTP, NDJSON, and webhooks
- KPI bounds, z-score, and divergence rules
- Signal divergence detection across community, playtest, experiment, and KPI evidence
- Portfolio KPI and comparison visuals
- Investigation inbox and timeline
- Drift-triggered and portfolio-triggered follow-up studies
Automated investigation
- Automated investigation workflows
- Anomaly hypothesis generation with game and operating context
- Anomaly classification including channel bias, drift, and seasonality
- Diagnostic study templates
- Multi-channel and multi-binding diagnostic designs
- Decision briefs connected to the original alert
- Human approval before fielding, spend, or public exposure
- Trusted-automation presets for notify vs confirm behavior
- 07
Automate research without losing control
Let Subtle Beacon guide or run the repetitive work while researchers retain authority over methods, interpretation, spend, and exposure.
Study design assistance
- Natural-language situation parsing
- Study-type recommendations with rationale
- Guided setup from question → type → design → register
- Incremental configuration builder
- LLM-seeded default field values
- Field and schema validation
- Bayesian sample, cost, and duration estimates
- Study templates and experiment playbooks
- Copy generation and editing assistance
- Study preview before launch
Autopilot to copilot workflows
- Full design → register → field → fit → brief loop
- Agent-assisted fielding from prepare through launch and polling
- Agent-assisted follow-up design
- Manual-studio, trusted-automation, and agent-heavy interaction presets
- Human approval for high-risk fields, spend, and public exposure
- Pending-approval discovery for researchers, coordinators, and agents
- 08
Reuse learning safely
Let prior studies inform new ones without blindly pooling results or exposing raw respondent data.
Within-studio learning
- Same-tenant prior transfer
- Causal transfer assessment before applying historical evidence
- Historical and live-market evidence composed into new study priors
- Mechanism-aware discounting when studies are related but not identical
- Holdout validation of transferred priors
Cross-studio federation
- Contribute anonymized preference summaries rather than raw respondent rows
- Consume population priors with similarity retrieval
- Causal prior selection
- Ranked shortlist of which past studies should inform the new study
- Mechanism-aware transfer assessment and discount factors
- Explain-before-apply workflow
- Mechanism memory from aggregate validation feedback
- Tenant opt-out and policy gates on export and consumption
- No contributor identity in tenant-facing artifacts
- 09
Fit the studio’s existing stack
Use Subtle Beacon as the research system of record while keeping existing analytics, data, fielding, and deployment tools in place.
Research, data & product integrations
- Product analytics integrations
- Warehouse integrations
- Event-pipeline integrations
- Remote-config and feature-flag integrations
- Survey and panel integrations
- Store and marketing integrations
- Revenue and CRM integrations
- Remote fielding-provider handshake
- Patient Cartographer experiment triggers
- Latent Spark competitive context and synthetic priors
- Posterior push to Anchored Horizon for forecasting and planning
APIs, agents & eventing
- Headless API
- OpenAPI specification
- Published TypeScript client
- MCP tools, resources, and prompts
- A2A-compatible agent surfaces
- Agent registration and claim flow
- Scoped delegation
- Webhooks
- AsyncAPI event streams
- Federated agent discovery across Iridae peers
- One-shot and scheduled operating-intelligence chains
- Durable workflow checkpoints, resume, and cancel
- 10
Meet research and enterprise governance requirements
Control access, approvals, confidentiality, automation, and auditability across humans, agents, vendors, and research channels.
Research governance
- Tenant isolation
- Role-based access control
- Approval policies by action type and fielding participant
- Spend and fielding approval gates
- Public-exposure review
- Copy and media approval
- Leak and sensitive-content checks
- Study confidentiality tags: public, internal, and embargoed
- Audit trails and lineage artifacts
- Policy-controlled federation and cross-system export
Agent & automation governance
- Agent identity
- Scoped agent permissions
- Manual-studio, trusted-automation, and agent-heavy policy presets
- Delegated permissions with scope, caps, and expiry
- Human and agent access through the same governed contracts
Enterprise platform
- Hosted multi-tenant SaaS
- Secure tenant-scoped data handling
- Idempotent operations
- Rate limits and health probes
- Observability metrics and runbooks
- Usage reporting
- Entitlement-aware premium capabilities
- Commerce checkout, rate cards, and billing portal where enabled
- Design-partner and portfolio deployments
Experiment types & measurement
- A/B and multi-arm experiments
- Conversion, count, rate, and ordinal metrics
- Observational comparison and KPI-monitor studies
- In-game variant assignment
- Exposure and outcome ingestion
- Hybrid studies combining stated preference with live outcomes
- Portfolio-level KPI monitoring
- Experiment lifecycle events: register, tick, stop, and decision reached
- Idempotent event handling
Adaptive allocation & stopping
- Thompson Sampling allocation
- Allocation rails that preserve exploration
- Adaptive traffic allocation
- Automated tick cycles for live experiments
- Posterior-driven stopping
- Probability-of-superiority rules
- Expected-regret thresholds
- Credible-interval-width goals
- Superiority, uplift, regret, and guardrail decision rules
- Non-inferiority rules
- Peeking mitigation
Experiment quality & safety
- Guardrail metrics and thresholds
- Sample-ratio mismatch detection
- Drift detection
- Drift-triggered follow-up studies
- Probability-of-best and expected-regret readouts
- Holdout validation
- Statistical release gates
- Web SDK and Unity integration
- Remote-config and feature-flag integrations including PlayFab, Statsig, Firebase, and LaunchDarkly
Conjoint & discrete choice
- Hierarchical Bayes conjoint using HB-MNL
- Standard and adaptive conjoint design
- ACBC with screening, build-your-own, tournament, and calibration phases
- Partial-profile designs
- Prohibited pairs
- Dual-response none
- Balanced overlap
- Individual- and population-level part-worths
- Attribute importance modeling
- Willingness-to-pay modeling
- Segment and cohort preference cuts
Ranking, preference & concept methods
- MaxDiff / best-worst scaling
- Standard, sparse, express, and tournament-wave MaxDiff designs
- Pairwise preference studies using BTL, Thurstone, and hierarchical Bayes models
- Gaussian-process preference learning
- Menu-based choice studies
- Rating studies
- Constant-sum studies
- Concept testing
- Message testing
- Feature tradeoff studies
- Behavioral and revealed preference from telemetry
Adaptive study design & validation
- Bayesian optimal experimental design
- Expected-information-gain task selection
- Adaptive conjoint design batches
- Adaptive and bandit task selection for MaxDiff
- Adaptive task allocation
- Living posterior per study
- Respondent quality checks
- Holdout validation
- Convergence diagnostics
- Statistical release gates
- Learned and amortized design policies where configured
Hosted & external fielding
- Hosted Subtle Beacon surveys
- Rating, menu-based choice, constant-sum, and ACBC respondent sessions
- Tenant-branded surveys and custom domains
- Survey-vendor fielding including Qualtrics, Typeform, and related providers
- Gamer-panel fielding including Helpfull and related providers
- Discord and community studies
- CRM hybrid fielding through HubSpot and Salesforce
- Email-list studies with customizable invitations and branding
- LLM interviewer fielding
- Manual CSV import and export
Behavioral & product evidence
- Live-game exposure through SDKs and remote config
- Product analytics outcome connectors
- Warehouse and event-pipeline connectors
- Store, marketing, and attribution outcome sources
- Playtest builds with behavioral follow-up
- Hybrid A/B and stated-preference studies
- Multi-binding orchestration across several evidence sources in one study
Fielding operations & review
- Respondent-facing task previews
- Copy and media previews
- Durable stage → review → launch approval workflow
- Dynamic fielding-provider registration
- Tenant approval of external providers
- Copy and media upload
- Leak and sensitive-content review
- Composite approval covering copy, channel, spend, and public exposure
Game & decision context
- Persistent workspace per game or product
- Subject, population, and catalog anchors
- Decision frame for every business question
- Study question, candidates, and rationale captured at intake
- Subject registry and game disambiguation
- Audience profiles
- Fielding-audience recommendations
- Product claims with inferred vs human-reported provenance
Iridae context enrichment
- Latent Spark mechanic beliefs, embeddings, and competitive context
- Patient Cartographer pathway and tactic metadata at study intake
- Context threaded through study design, fielding, federation, automation, and decision outputs
- Operational context attached to compiled study specifications and result payloads
Research readouts
- Winning option and confidence
- Probability each option is best
- Expected regret of acting now
- Credible intervals on outcomes
- Segment-level preference differences
- Attribute importance estimates
- Part-worth estimates
- Willingness-to-pay ranges
- Guardrail status
- Drift and anomaly alerts
- Rendered charts and VizKit visual artifacts
- Plain-language posterior readouts
- Plain-language decision briefs
- Recommended next study
Market & commercial simulation
- Choice-share simulation using softmax and first-choice rules
- Segment-level scenario comparison
- Pricing sensitivity analysis
- Bundle optimization
- Source-of-volume analysis
- TURF-style reach analysis
- Revenue and profit tradeoff analysis
- Counterfactual scenarios
- Dynamic multi-period adoption and churn scenarios
- Competitive-response scenarios
- Roadmap what-if simulations
- Sensitivity grids for price, content, and feature levels
- Industry-standard export formats where configured
Diagnostics & follow-up
- Diagnostic proposals with hypothesis, bindings, cost, and success criteria
- Decision briefs tied to the originating alert or business question
- Recommended next study on investigation threads
- Federation prior-selection explanations
- Explain why previous studies were selected to inform a new prior
Portfolio sensing
- Scheduled portfolio health scans
- External KPI registry and ingest through HTTP, NDJSON, and webhooks
- KPI bounds, z-score, and divergence rules
- Signal divergence detection across community, playtest, experiment, and KPI evidence
- Portfolio KPI and comparison visuals
- Investigation inbox and timeline
- Drift-triggered and portfolio-triggered follow-up studies
Automated investigation
- Automated investigation workflows
- Anomaly hypothesis generation with game and operating context
- Anomaly classification including channel bias, drift, and seasonality
- Diagnostic study templates
- Multi-channel and multi-binding diagnostic designs
- Decision briefs connected to the original alert
- Human approval before fielding, spend, or public exposure
- Trusted-automation presets for notify vs confirm behavior
Study design assistance
- Natural-language situation parsing
- Study-type recommendations with rationale
- Guided setup from question → type → design → register
- Incremental configuration builder
- LLM-seeded default field values
- Field and schema validation
- Bayesian sample, cost, and duration estimates
- Study templates and experiment playbooks
- Copy generation and editing assistance
- Study preview before launch
Autopilot to copilot workflows
- Full design → register → field → fit → brief loop
- Agent-assisted fielding from prepare through launch and polling
- Agent-assisted follow-up design
- Manual-studio, trusted-automation, and agent-heavy interaction presets
- Human approval for high-risk fields, spend, and public exposure
- Pending-approval discovery for researchers, coordinators, and agents
Within-studio learning
- Same-tenant prior transfer
- Causal transfer assessment before applying historical evidence
- Historical and live-market evidence composed into new study priors
- Mechanism-aware discounting when studies are related but not identical
- Holdout validation of transferred priors
Cross-studio federation
- Contribute anonymized preference summaries rather than raw respondent rows
- Consume population priors with similarity retrieval
- Causal prior selection
- Ranked shortlist of which past studies should inform the new study
- Mechanism-aware transfer assessment and discount factors
- Explain-before-apply workflow
- Mechanism memory from aggregate validation feedback
- Tenant opt-out and policy gates on export and consumption
- No contributor identity in tenant-facing artifacts
Research, data & product integrations
- Product analytics integrations
- Warehouse integrations
- Event-pipeline integrations
- Remote-config and feature-flag integrations
- Survey and panel integrations
- Store and marketing integrations
- Revenue and CRM integrations
- Remote fielding-provider handshake
- Patient Cartographer experiment triggers
- Latent Spark competitive context and synthetic priors
- Posterior push to Anchored Horizon for forecasting and planning
APIs, agents & eventing
- Headless API
- OpenAPI specification
- Published TypeScript client
- MCP tools, resources, and prompts
- A2A-compatible agent surfaces
- Agent registration and claim flow
- Scoped delegation
- Webhooks
- AsyncAPI event streams
- Federated agent discovery across Iridae peers
- One-shot and scheduled operating-intelligence chains
- Durable workflow checkpoints, resume, and cancel
Research governance
- Tenant isolation
- Role-based access control
- Approval policies by action type and fielding participant
- Spend and fielding approval gates
- Public-exposure review
- Copy and media approval
- Leak and sensitive-content checks
- Study confidentiality tags: public, internal, and embargoed
- Audit trails and lineage artifacts
- Policy-controlled federation and cross-system export
Agent & automation governance
- Agent identity
- Scoped agent permissions
- Manual-studio, trusted-automation, and agent-heavy policy presets
- Delegated permissions with scope, caps, and expiry
- Human and agent access through the same governed contracts
Enterprise platform
- Hosted multi-tenant SaaS
- Secure tenant-scoped data handling
- Idempotent operations
- Rate limits and health probes
- Observability metrics and runbooks
- Usage reporting
- Entitlement-aware premium capabilities
- Commerce checkout, rate cards, and billing portal where enabled
- Design-partner and portfolio deployments
See Subtle Beacon in action.
Explore how your studio can design studies, connect evidence, and move from uncertainty to better decisions.