Governed portfolio intelligence for unstable regimes
ATLAS
A regime-aware decision system that measures structural state, tactical instability, confidence, disagreement, portfolio risk, and allocation constraints within one auditable framework.
Designed to know when to allocate, when to reduce risk, and when not to act.
Regime intelligence
Structural regime, transition geometry, velocity, acceleration, and persistence over macro-financial state space.
Tactical instability
Flip risk, confidence decay, disagreement, and fast reversible controls that restrict exposure without redefining regime.
SHELOB allocation
Governed portfolio and sleeve allocation under uncertainty, constraints, and explicit decision authority.
AI briefings
Machine-readable briefing selection and governed narrative generation from authoritative system state.
Audit and governance
Model-change audit, pipeline lineage, provenance, freshness controls, reason codes, and operator authority.
Abstention
Explicit transitional and indeterminate states with explainable triggers, episode lifecycle, and outcome assessment.
Self-Service Demo
Explore ATLAS with sample portfolio data in a sandboxed demo environment. No account required. Demo sessions use delayed market data and include no live portfolio information.
What ATLAS Is
ATLAS is decision-support infrastructure for portfolio governance and macro-regime monitoring. It is used to manage exposure to uncertainty in global markets.
Rather than forecasting short-term prices, ATLAS measures whether the current macro-financial environment is stable, transitional, or indeterminate; whether confidence in that assessment is rising or decaying; and whether the resulting posture should be to allocate, reduce, or abstain.
- Regime-aware. The structural regime is a slow, persistent classification of macro conditions, estimated on a declared calibration window and held out from live evaluation.
- Uncertainty-aware. Confidence, flip risk, score disagreement, and data freshness are explicit inputs to the admissibility map, not afterthoughts.
- Governed. Every artifact carries schema, keys, dates, provenance, and freshness state. Overrides require operator identity and reason codes.
- Abstention-first. When evidence is insufficient or unstable, the legitimate output is to reduce or abstain. Abstention is tracked, attributed, and evaluated after the fact.
What ATLAS Is Not
- ATLAS is not a trading bot or execution engine.
- ATLAS is not a short-term market-prediction system.
- ATLAS is not a signal-stacking or black-box alpha platform.
- ATLAS does not convert every analytical result into a trade.
The system provides a structured, auditable framework for evaluating macroeconomic conditions and portfolio alignment across regimes. Analytical strength does not automatically create operational authority.
SHELOB: Governed Allocation Under Uncertainty
SHELOB is ATLAS's governed allocation and portfolio-construction capability. It combines portfolio mathematics with structural regime state, tactical instability, sleeve constraints, optionality requirements, and explicit decision authority.
- More than an optimizer. The live allocation baseline (TILT) is governed and produced by the pipeline. Shadow optimizer families (MVO, Black-Litterman, equal-weight) run as diagnostics and counterfactuals, not as live instructions.
- Regime- and confidence-aware. Target changes are gated by abstention state, tactical instability, and portfolio constraints. An unstable or indeterminate state can block or scale a proposed adjustment.
- Constraint-bound. Sleeve limits, single-position caps, liquidity-tier scalers, and cash-absorption rules are enforced before any target is emitted.
- Descriptive outputs. SHELOB surfaces explain weights, availability, and readiness; the ANDURIL allocation overlay returns an explanatory permission and reason stack, not trade instructions.
- Fail-closed. Missing inputs, stale data, unresolved disagreement, or an inactive governance posture routes the output to a restricted or informational state.
Read the SHELOB formalization in the Technical Specification →
AI Briefings: Governed Narrative from Machine-Readable State
ATLAS exposes its governed state through a machine-readable briefing catalog. Structured requests are resolved to approved briefing types, populated from authoritative artifacts, checked against policy and entitlement constraints, and rendered as human-readable analysis.
- Catalog-driven. Briefing types, required payload fields, and policy boundaries are registered in a closed vocabulary before any prose is generated.
- Parser-based. A structured adapter resolves requests to the correct briefing builder, validates inputs, and enforces guardrails.
- Governed context. The AI layer reads pre-computed artifacts—regime state, tactical instability, portfolio health, optionality, governance status—not live market feeds or internal paths.
- Policy-checked. Forbidden intents, trade-generation language, and allocation authority are blocked by construction.
- Same facts, multiple surfaces. The same governed context can produce operator briefs, sidebar narratives, audit appendices, and research diagnostics.
The AI layer explains governed system state; it does not create the state.
Advanced Capabilities
Measure
- Structural and tactical horizon separation
- Cross-asset score trajectory and geometry
- State velocity, acceleration, and persistence
- Confidence decay and transition risk
- Cross-model disagreement
- Portfolio and sleeve health
Decide
- Allocate / Reduce / Abstain posture map
- Abstention episode lifecycle
- SHELOB governed allocation baseline
- Shadow optimizer counterfactuals
- Optionality and protection planning
- Stress and risk analytics
Govern
- Tenant-aware routing and operator scope
- Reason-coded overrides with expiry
- Promotion committee and candidate evaluation
- Model-change reader and changelog
- Research-to-production separation
- Production / beta / shadow boundary
Explain
- Machine-readable AI briefing catalog
- Structured parser and context contract
- Layered decision explanation
- Pipeline manifests and provenance
- Freshness and staleness labels
- Truthfulness contract per surface
Why Trust ATLAS
Every decision has a lineage
- Governed source data and versioned analytical artifacts
- Explicit freshness and observability labels
- Model, schema, and stage-definition versions
- Decomposed decision logic and reason-coded restrictions
- Operator-attributed overrides with provenance
- Reproducible pipeline runs and run manifests
Research does not equal authority
- Lab, shadow, validation, approval, and production are separate states
- Inconclusive and falsified results remain visible
- No new model receives decision authority by default
- Promotion is an explicit, operator-gated event
AI explains; governance decides
- Machine-readable briefing parser and catalog
- Governed context populated from authoritative artifacts
- Policy and entitlement checks before rendering
- No independent allocation or trading authority
Development Status
ATLAS runs in production for governed regime monitoring, portfolio state, and operator-isolated decision tooling. The beta environment hosts canary surfaces and experimental capabilities. Shadow and diagnostic outputs are labeled and do not carry decision authority.
Pricing
For pricing inquiries, institutional licensing, and operator onboarding, please visit the Contact page.
Updates
Release notes and development milestones
Production Update — Decision Ledger, Surface Honesty & Calibrated Caution
A production update is live at app.atlas-portal.ca. The release sharpens how ATLAS represents what it knows, what it does not know, and how the platform talks to operators about both.
- Unified decision ledger. Decisions, realized economics, regime context, and forward evidence now live behind a single auditable surface, so that the end-to-end story of any decision — from input artifact to outcome — can be reviewed in one place. The unified ledger is a measurement and audit reference; it has no allocation authority of its own.
- Surface-level truthfulness. Decision-facing panels now declare their evidence basis explicitly, indicating whether a view is backed by live governance, a shadow replay, a diagnostic model, or insufficient evidence. Missing or stale inputs surface as labeled states rather than silent fallbacks.
- Calibrated caution. The platform now maps measurement-uncertainty states to a graduated, minimal-response policy. Where the system cannot yet make a confident statement, it says so, and the appropriate degree of caution is recorded against the abstention ledger for outcome evaluation.
- Research discipline for new strategy contexts. Candidate trend and momentum strategy contexts are now tracked through a formal evidence dossier — forward evidence, reconstructed historical evidence, and live-portfolio context — before any promotion into governance. Authorization gates are explicit and time-bound; nothing promotes itself by default.
- Adversarial review of decision-facing surfaces. A broad audit was run across the platform’s advisory and shadow surfaces, looking specifically for ways a panel could imply a claim it could not back. Findings were closed before the release shipped.
- Operator-visible pipeline state. Pipeline telemetry, run health, and tenant-aware run identifiers are now consistently visible across observability surfaces, so operators can tell at a glance whether what they are reading is current, partial, or stale.
Production Update — Multi-Operator Governance, Pipeline Reliability & Structural Forecasting
A major production update is now live at app.atlas-portal.ca. The release brings ATLAS’s multitenant governance posture, pipeline reliability, and structural forecasting capabilities to production grade.
- Multi-operator governance. Operator and sleeve identity now flow end-to-end through the platform. Overrides, audit trails, entitlements, and portfolio visibility are isolated by operator context, so each operator sees and acts on only their own scope.
- Structural forecasting in production. Regime flip-probability estimation, persistence modeling, and hazard-rate diagnostics are now operational in production across multiple forward horizons. Forecasts are surfaced as diagnostic context, separated from exposure policy.
- Forecast calibration. A dedicated calibration artifact now measures how well structural forecasts have matched realized outcomes across horizons, so the platform can be honest about where its forecasts have been reliable and where they have not.
- Fail-fast pipeline. The pipeline now validates governance readiness, configuration, and stage prerequisites before expensive computation begins. Problems surface immediately at startup rather than late in a run.
- One source of truth for abstention. Abstention state was consolidated behind a single authoritative source feeding both the UI and portfolio gating, eliminating the possibility of disagreement between views.
- More robust portfolio ingest. Multi-file upload handling, timestamp safety, and collation were hardened, reducing operator friction during portfolio refreshes.
- Forecast observatory. Structural-transition surfaces, horizon-specific transition probabilities, and forecast diagnostics now render consistently across the operator UI.
- Abstention economics. Regime-conditioned breakdowns and the abstention economics panel were restored, including episode outcome classification and false-caution tracking by regime.
- Sleeve-scoped optionality. The optionality view is now strictly scoped to the selected sleeve, with no silent fallback to an aggregate view when sleeve context is missing.
- Production infrastructure. Entitlements infrastructure was modernized with automated migration and operator backfill, provenance guarding was added for containerized runs, and deployment tooling was extended to support the multitenant production surface.
Beta v2 — Regime-Conditioned Model, GitPortfolio & Forecasting
A new beta version is now available at beta.atlas-portal.ca incorporating significant new capabilities:
- Regime-Conditioned Model — Portfolio analytics and risk assessments are conditioned on the latest classified macro regime, with its as-of date.
- GitPortfolio — Live portfolio replay and counterfactual analysis. Track how portfolio decisions would have played out under alternative regime paths and allocation strategies.
- Forecasting — Forward-looking regime-transition probability modeling and macro-forecasting tools for scenario planning and exposure management.
Site Launch — Cross-Asset Regime Model
ATLAS Portal is live. The initial release introduces the core cross-asset regime classification model — a system for detecting structural regime shifts, systemic stress, and instability across global financial markets.
- Cross-asset regime classification and confidence scoring
- Multi-signal macro stress monitoring
- Regime flip-risk estimation and transition probability modeling
- Portfolio exposure alignment diagnostics
- Audit logging and governance tooling built toward an institutional standard
Technical Specification
Formal model objects, governance, and admissible decision rules
1. Decision-Theoretic Objective
ATLAS is specified as a tuple of measurable and governed objects. Every component is defined in the sections that follow with its domain, role, and admissible use.
- \(\mathcal{X}\) — observable state space (cross-asset, regime-scoring inputs)
- \(\mathcal{D}\) — governed artifact space (schema-validated, provenance-tracked outputs)
- \(\mathcal{G}\) — governance / data-quality state space (admissibility of \(\mathcal{D}\))
- \(\mathcal{S}\) — structural regime state space (finite, slow, persistent)
- \(\mathcal{T}\) — tactical instability state space \(\{\text{Stable},\, \text{Transitional},\, \text{Unstable}\}\)
- \(\mathcal{C}\) — confidence / boundary-distance space
- \(\mathcal{F}\) — flip-risk / transition-pressure space
- \(\Pi\) — admissible policy map (Section 14)
- \(\mathcal{V}\) — validation operator family (Section 21)
ATLAS is a measurement and decision-governance system. It is not a return forecaster, an unconstrained optimizer, or an execution engine.
The operative action is a function of structural state, tactical instability, portfolio state, uncertainty/validity, governance state, and authorized operator overlay:
The action space is \(\{\text{increase},\, \text{maintain},\, \text{reduce},\, \text{hedge},\, \text{abstain}\}\). In the current implementation the live posture set is \(\{\mathrm{Allocate},\, \mathrm{Reduce},\, \mathrm{Abstain}\}\). Decision authority follows the lane precedence
Protected-parameter convention. The mathematical structure is disclosed. Production coefficients, thresholds, and calibration constants are withheld to preserve model integrity, operational security, and proprietary implementation detail. Where coefficients or thresholds are sensitive, the symbol \(\theta_{\text{protected}}\) is used.
1.1 Standing assumptions
All random variables in this document are defined on a fixed probability space \((\Omega, \mathcal{H}, \mathbb{P})\) equipped with the discrete-time filtration \((\mathcal{F}_t)_{t \in \mathbb{Z}_{\ge 0}}\) introduced in Section 3. Statements of equality involving random variables are understood \(\mathbb{P}\)-almost surely; statements of equality involving deterministic objects are absolute.
- A1 (Finite latent state). The structural regime takes values in a finite set, \(|\mathcal{S}| < \infty\), and evolves as a Markov chain with transition matrix \(P\) (Section 9).
- A2 (Information adaptation). All decision-time objects — \(S_t,\, T_t,\, C_t,\, F_t,\, D_t^{\,\text{score}},\, \pi_t\) — are \(\mathcal{F}_t\)-measurable.
- A3 (Fail-closed tactical layer). Inputs that are NA-driven or whose governing artifacts are inadmissible cannot resolve to \(T_t = \text{Stable}\) (Section 5).
- A4 (Deterministic validation). Each validator \(V_j\) is a deterministic function on artifact contracts (Section 21).
- A5 (Closed-set authorization). The authorization predicate \(\operatorname{Auth}(o, s, r)\) is two-valued (Section 16).
2. Structural State Model
The slow structural state vector is
where \(G_t\) is growth, \(I_t\) is inflation, and \(L_t\) is liquidity / financial-conditions state. Let \(Z_t = (Z_{1,t}, \ldots, Z_{k,t})\) denote a low-dimensional vector of governed scores. Then the structural regime label is a projection
\(\Theta_S\) denotes the governed structural parameters, including normalizations, sign conventions, classification thresholds, and persistence rules. The categorical regime label \(R_t\) is a downstream projection of the continuous state \(S_t\); the full state is not collapsed to the label.
Structural-regime estimation is in the regime-switching tradition (Hamilton, 1989), in which the latent state evolves as a Markov chain with transition matrix
The diagonal \(P_{ii}\) is the one-step self-persistence of regime \(i\); the off-diagonal mass governs reachability between regimes. Structural regime is not tactical instability:
3. Alternative Embeddings
The structural state may be represented in several embeddings. Each has a declared status and decision authority.
| Embedding | Coordinates | Status | Authority |
|---|---|---|---|
| \(E_{\text{GIL}}\) | \((G_t, I_t, L_t)\) | PRODUCTION_DEFINITION | Live input to regime score and classification |
| \(E_{\text{GILR}}\) | \((G_t, I_t, L_t)\) plus visual time parameter \(\tau\) for helix plotting | PRODUCTION_STRUCTURE_WITH_PROTECTED_PARAMETERS | Topology diagnostics only |
| Higher-dimensional research embeddings | Macro + credit + liquidity + volatility axes | RESEARCH_MEASUREMENT | Research; not policy-wired |
4. State-Space Metric
All geometric quantities are defined with respect to a declared metric. The distance function is
\(M\) may define Euclidean, scaled-Euclidean, whitened, covariance-adjusted, or Mahalanobis geometry. The exact \(M\) matrix and axis scales are artifact-specific and protected. The metric must be declared before speed, curvature, path length, or basin distance are interpreted.
5. Kinematics
State velocity and acceleration are defined with respect to the visual time parameter \(\tau\):
Discrete forms used in production are
and the metric norms are
Smoothing windows, finite-difference weights, and minimum-history requirements are governed per artifact.
6. Geometry
Curvature, torsion, directional stability, and path-length quantities are computed on the declared embedding and metric.
Path geometry over a window \([a, b]\) is
These are geometric descriptors; not all have decision authority. Smoothing method, singular-case handling, and minimum-window length are protected.
7. Regime Centroids and Basin Structure
For each structural regime \(r\), the centroid is
The expanding centroid is
and the centroid distance is
Centroid calculation mode (retrospective, frozen, expanding, rolling) and basin thresholds are artifact-specific and protected. The public representation discloses the definition and the fit mode per artifact.
8. Boundaries and Transition Surfaces
The quadrant margin is
General boundary surfaces are written
Boundary distance, normal velocity, and normal acceleration are
Thresholds \(\theta_G, \theta_I\), boundary surface definitions, and normal vectors are protected. Boundary proximity is necessary but not sufficient; direction and acceleration matter.
9. Transition Matrices and Entropy
Transition probabilities are defined by
with empirical estimate
Four entropy forms are distinguished and are not collapsed into a single “transition entropy”:
Count minimums, smoothing pseudocounts, and horizon windows are protected.
10. Confidence
Confidence is proximity-to-boundary adjusted for persistence, uncertainty, disagreement, and data quality; it is not certainty. The live confidence object uses the score-margin form
A geometric distance formulation
is a SHADOW research diagnostic; it does not gate policy. In both formulations \(C_t \downarrow 0\) as \(Z_t \to \Gamma\).
Confidence is decomposed as
Confidence change and the confidence shock are
The functional form and the calibration of \(\mu_{\Delta C}\) and \(\sigma_{\Delta C}\) are protected.
11. Flip Risk and Transition Pressure
Flip risk is a transition-risk measure, not a directional market forecast. It is decomposed as
Flip-risk velocity and acceleration are
Three properties hold by construction: (i) static-level exceedance alone is not sufficient for elevated transition pressure; (ii) \(\Delta^{2} F_t\) can be informative even when \(F_t\) has not crossed a static threshold; (iii) short or incomplete histories are labeled rather than imputed to Stable.
12. Disagreement
Model disagreement (dispersion) across \(K\) scores or models is
Pairwise disagreement is
The set of models/scores \(K\) and the normalization method are protected. Score disagreement as the range across structural-score axes is the live form:
Surface disagreement, the indicator that two classification surfaces assign different structural labels, is measured but not wired into the live admissibility map. It remains a research track (WS7).
13. Meta-Regimes
The meta-regime space is
| Meta-regime | Definition |
|---|---|
TRANSITIONAL | Coherent movement (\(|V_t| > 0\)) but destination not yet persistent. |
INDETERMINATE | Caused by \(D_t > \theta_D\), \(Q_t \in \{\text{STALE}, \text{MISSING}, \text{SCHEMA_INVALID}\}\), or excessive estimator sensitivity. |
Threshold \(\theta_D\) and persistence requirements are protected. Meta-regime labels are explicit states, not error conditions.
14. Abstention
The policy map is
where \(\mathcal{U}\) is the set of admissible non-null exposure actions and \(\varnothing\) denotes abstention. In the current implementation \(\mathcal{U} = \{\mathrm{Allocate},\, \mathrm{Reduce}\}\).
The abstention gate is
Triggers include confidence shock, confidence decay, flip-risk acceleration, score/model disagreement, stale/invalid data, unresolved transition, governance restriction, and operator hold. Responses include gross reduction, capital preservation, optionality retention, delayed commitment, and operator review.
An abstention episode is a maximal interval over which \(\pi_t = \varnothing\), paired with a post-hoc outcome label at a pre-registered forward horizon:
- Justified Caution (JC). Forward drawdown exceeded tolerance; abstaining avoided realized loss.
- False Caution (FC). Forward drawdown remained within tolerance; abstaining was unnecessary.
- Inconclusive (IN). Insufficient forward data or intermediate outcome.
Abstention follows the reject-option framework of Chow (1970) and its extension to selective classification in El-Yaniv & Wiener (2010).
15. Stress and Risk Mathematics
Portfolio return and variance are
Marginal and component contributions to risk are
Tracking error, drawdown, VaR, and Expected Shortfall are
Variance, MCR, tracking error, and drawdown are live. VaR and ES are conceptual decompositions unless a live artifact proves otherwise.
The live stress-cooked pressure is
with activation \(\text{stress_active}_t = \text{pressure}_t > \text{stress}_z\) where \(\text{stress}_z = 1.0\).
16. Portfolio and Sleeve Health
Constraint headroom is
and normalized headroom is
Health dimensions include concentration, diversification, risk contribution, exposure drift, sleeve budget utilization, optionality coverage, hedge coverage, liquidity, and stale holdings. Limits \(b_j\), scales \(s_j\), sleeve definitions, and liquidity thresholds are protected. Health is a multi-dimensional constraint and coverage assessment, not a single score.
17. SHELOB
SHELOB is the governed allocation and portfolio-construction layer. The general optimization form is
subject to
and the regime/confidence feasible set
\(\lambda_t\), penalty weights, \(G_{\max}\), \(T_{\max}\), \(B_g\), \(\tau^{2}\), and sleeve limits are protected. SHELOB governs the feasible allocation space; candidate, governed, and operative weights are separated.
Infeasibility reason codes include INFEASIBLE_CONSTRAINT_SET, MISSING_RISK_MODEL, STALE_INPUT, UNAUTHORIZED_CONTEXT, INSUFFICIENT_HISTORY, and SOLVER_FAILURE.
18. Black-Litterman
Black-Litterman is one candidate within SHELOB, not the sole allocation method. The implied equilibrium return is
and the posterior mean is
\(\delta\), \(\tau\), \(\Omega\), view matrix \(P\), view vector \(q\), and view confidences are protected. Views are governed; BL is a shadow/governance candidate.
19. Optionality and Hedge Mathematics
A protection bundle at decision time \(t\) is a finite set of contracts
with strike grid \(\{k_j\}\), contract counts \(\{n_j\}\), and per-contract proxy premia \(\{c_j\}\). The coverage ratio is
the planning budget is
and the put payoff is
\(\rho_t\), strike ladders, live positions, budget limits, and tenor targets are protected. Optionality is planned, sized, and reconciled; proxies are labeled.
The protection-cost proxy is
labeled ESTIMATE_PROXY, where \(c_j^{(t_0)}\) is a static proxy premium at snapshot \(t_0\), not the real-time mid-market price. Real option-chain pricing is DEFERRED.
20. Model-Change Auditing
Model changes are governed events. Output difference is
Compared deltas include
Regime label change is
Allocation distance, risk difference, and constraint change are
Attribution is conceptually decomposed as
Change classes include NO_SEMANTIC_CHANGE, DISPLAY_ONLY, SCHEMA_CHANGE, FEATURE_CHANGE, PARAMETER_CHANGE, MODEL_CHANGE, DECISION_POLICY_CHANGE, and GOVERNANCE_CHANGE. Promotion states include PROPOSED, LAB, SHADOW, VALIDATED, INCONCLUSIVE, FALSIFIED, APPROVED, PROMOTED, and RETIRED.
21. Statistical Validation Framework
Validation metrics are defined on pre-registered horizons and held-out windows. False-alarm rate, miss rate, precision, and recall are
Brier and log scores are
The permutation p-value is
Sensitivity dimensions include embedding, metric, derivative method, calibration window, data vintage, and event definition. \(B\), significance thresholds, and minimum evaluable sample are protected. Negative, inconclusive, unstable, and falsified outcomes remain valid.
22. Data Revisions and Vintage Integrity
Vintage notation distinguishes values by observation date and as-of vintage:
The revision between vintages is
ATLAS distinguishes latest revised history, real-time vintage history, current incomplete observation, and last complete observation.
23. Freshness and Observability
Data age is
Freshness state is
States are \(\{\text{CURRENT},\; \text{CURRENT_DAY_INCOMPLETE},\; \text{STALE},\; \text{MISSING},\; \text{SCHEMA_INVALID},\; \text{UNOBSERVABLE}\}\). Truthfulness status is \(\{\text{AVAILABLE},\; \text{DEGRADED},\; \text{STALE},\; \text{UNAVAILABLE},\; \text{UNKNOWN}\}\) plus closed reason codes. Cadence thresholds and max stale days are protected. Missing data never implies stability; scientific status and data status are separate.
24. Pipeline Mathematics and Contracts
The transformation chain is
Each stage carries a contract: input schema, output schema, date basis, uniqueness keys, source hash, producer version, validation state, environment, tenant context, and artifact version. The run manifest is
Absolute paths, run IDs, git/image digests, and tenant identifiers are protected. Derived analytics do not depend on UI state; the UI consumes governed artifacts.
24.1 Observation and artifact layer
Let \(t\) denote decision time and \(d\) an observation date. Let \(X_t \in \mathcal{X}\) denote the observable state vector and \(\mathcal{D}_t\) the set of artifacts available at \(t\):
For each artifact define its maximum observation date and lag:
Artifact state is a closed-set label
and eligibility is
The natural artifact filtration is
Stale-state truthfulness. A current-day-complete artifact has \(\ell_i(t) = 0\) and \(q_i(t) = \mathrm{OK}\); a latest-available artifact has \(\ell_i(t) \ge 0\) and \(q_i(t) \in \{\mathrm{OK}, \mathrm{STALE}, \mathrm{DEGRADED}\}\). The two are not the same object; the latter must carry its label. Absence of data is not evidence of stability.
25. Governance Mathematics
Authority is a function of entitlement, validity, freshness, research/production status, operator authorization, and execution context:
Authority states are
Permission severity is ordered
Authorization is a closed-set predicate
A request is valid only if
Entitlement mappings, operator IDs, tenant IDs, and approval boundaries are protected. Analytical strength does not automatically create operational authority; the system fails closed with no silent fallbacks.
25.1 Multitenancy and operator scope
The per-operator-sleeve filtration is
Cross-sleeve leakage is structurally prohibited. Reports, exports, overrides, governed artifacts, and promotion gates are scoped to \((o, s)\). Overrides carry actor identity, reason code, and expiry; they are never reinterpreted as structural state.
25.2 Policy separation
Four separations are enforced architecturally:
- Measurement \(\neq\) policy. \(C_t, F_t, D_t^{\text{score}}\) are inputs to the admissibility map; they are not actions.
- Policy \(\neq\) execution. \(\pi_t\) is a posture, not a trade. Execution is a separately scoped function gated by operator authorization.
- Reporting \(\neq\) allocation authorization. Reportability does not imply \(\operatorname{Auth}(o,s,r)=1\).
- Research \(\neq\) live governance. Shadow artifacts are \(\mathcal{F}_t\)-measurable but do not enter the domain of the admissibility map.
- Summary \(\neq\) decision. Natural-language summaries are read-only projections; their range does not intersect \(\mathcal{U}_t\).
26. AI Briefing Parser Mathematics and Contracts
AI prose is downstream of governed facts. Briefing request resolution is
where \(q\) is a machine-readable request and \(b\) is the resolved briefing type. Context assembly and narrative generation are
AI authority is explanatory only unless the repository proves another bounded role. Parser status labels are BLOCKED_MISSING_SURFACE, PARTIAL_UNAVAILABLE_INPUTS, and AVAILABLE_CONTEXT_ONLY. Forbidden intents include buy, sell, trade now, target weight should be, execute, override gate, guaranteed, prediction, and alpha signal.
The AI layer explains governed state; it does not create, alter, or bypass governance.
27. Outcome Auditing
The abstention avoided-loss counterfactual is
Other measures include maximum adverse excursion, maximum favorable excursion, abstention duration, resolution time, false caution, justified caution, avoided drawdown, and opportunity cost. Outcome evaluation is a structured assessment, not causal proof.
28. Measurement Programs
The programs below are measurement and evaluation tracks. None are allocation engines; promotion to live policy requires explicit governance gates.
28.1 HEDGEHOG
STATUS: LIVE. Scope is the tuple of measurement axes
28.2 SENTINEL
STATUS: RESEARCH. Transition-forecasting evaluation:
28.3 LANTERN
STATUS: SHADOW. Caution-policy confusion matrix:
28.4 TICM
STATUS: LIVE labeling. Stress-event window labeling:
28.5 WS7
STATUS: RESEARCH. Surface-disagreement track:
29. Estimation, Calibration, and Uncertainty
Parameters are estimated on a declared calibration window held out from evaluation:
For a generic parameter \(\theta \in \Theta\):
Sample-size guardrail:
Parametric confidence intervals, bootstrap-based interval estimation, and spline-based smoothing are not claimed as components of the live estimation stack. Parameter uncertainty is real and is not collapsed into false precision.
30. Limitations and Model Risk
The following limitations are stated explicitly to prevent inferential drift:
- Regime classification is uncertain; historical relationships may change.
- Model outputs depend on data quality, revisions, and the completeness of cross-asset inputs.
- Regime topology describes geometry, not causality.
- Confidence is not certainty; a high-confidence day can still precede a regime change.
- An optimizer does not eliminate model risk or tail risk.
- AI briefings summarize governed state but may still require human review.
- Research and shadow results do not automatically become production controls.
- No model can remove tail risk or guarantee downside protection.
- Abstention can be costly as well as protective.
- Counterfactual outcomes are structured assessments, not causal proof.
- Nonstationarity, metric dependence, and derivative sensitivity affect geometric quantities.
- Stale data and governance latency can restrict authority.
- Model and score disagreement are real and are not averaged away.
31. Public Mathematical Notation
Symbols introduced in this specification and their live status. Implementation status is one of LIVE, LIVE labeling, SHADOW, RESEARCH, or DEFERRED.
| Symbol | Domain | Interpretation | Status |
|---|---|---|---|
| \(t\) | \(\mathbb{Z}_{\ge 0}\) | decision time index | LIVE |
| \(d\) | calendar dates | observation date | LIVE |
| \(X_t\) | \(\mathcal{X} \subseteq \mathbb{R}^p\) | observable state vector | LIVE |
| \(D_i\) | \(\mathcal{D}\) | governed artifact \(i\) | LIVE |
| \(G_t\) | \(\mathcal{G}\) | artifact-admissibility / governance state | LIVE |
| \(\mathcal{F}_t\) | \(\sigma\)-algebra | natural artifact filtration | LIVE |
| \(S_t\) | \(\mathbb{R}^3\) | structural state \((G, I, L)\) | LIVE |
| \(R_t\) | finite \(\mathcal{S}\) | structural regime label | LIVE |
| \(T_t\) | \(\{\text{Stable}, \text{Transitional}, \text{Unstable}\}\) | tactical instability | LIVE |
| \(Z_t\) | \(\mathbb{R}^k\) | structural-score vector | LIVE |
| \(P\) | row-stochastic matrix | structural transition matrix | LIVE |
| \(M\) | positive-definite matrix | metric tensor | LIVE |
| \(V_t, A_t\) | \(\mathbb{R}^3\) | state velocity / acceleration | LIVE |
| \(\kappa, \mathcal{T}\) | \(\mathbb{R}_{\ge 0}\), \(\mathbb{R}\) | curvature / torsion | RESEARCH |
| \(\mu_r\) | \(\mathbb{R}^3\) | regime centroid | LIVE |
| \(d_B, v^{\perp}, a^{\perp}\) | \(\mathbb{R}\) | boundary distance / normal velocity / normal acceleration | LIVE |
| \(p_{ij}\) | \([0,1]\) | transition probability | LIVE |
| \(H_{\text{state}}, H_{\text{pair}}, H_{\text{rate}}\) | \(\mathbb{R}_{\ge 0}\) | entropy forms | LIVE |
| \(C_t\) | \(\mathbb{R}_{\ge 0}\) | confidence (margin form) | LIVE |
| \(F_t\) | \([0, F^{\max}]\) | flip risk / transition pressure | LIVE |
| \(\Delta^{2} F_t\) | \(\mathbb{R}\) | flip-risk acceleration | LIVE |
| \(D_t^{\text{score}}\) | \(\mathbb{R}_{\ge 0}\) | score-axis disagreement | LIVE |
| \(D_t^{\text{surface}}\) | \(\{0,1\}\) | surface-label disagreement | RESEARCH |
| \(\pi_t\) | \(\mathcal{U} \cup \{\varnothing\}\) | policy posture | LIVE |
| \(w\) | \(\mathbb{R}^n\) | portfolio weights | LIVE (TILT); SHADOW (MVO/BL/EQW) |
| \(\Sigma, \mu\) | matrix / vector | covariance and expected return inputs | LIVE |
| \(\text{MCR}_i, \text{CR}_i\) | \(\mathbb{R}\) | marginal / component contribution to risk | LIVE |
| \(\text{TE}, \text{DD}, \text{VaR}, \text{ES}\) | \(\mathbb{R}_{\ge 0}\) | tracking error / drawdown / tail-risk concepts | LIVE / SHADOW |
| \(A_t^{\text{abstain}}\) | \(\{0,1\}\) | abstention gate | LIVE |
| \(\mathcal{W}_t\) | constraint set | regime/confidence feasible set | LIVE |
| \(\rho_t\) | \([0,1]\) | optionality budget rate | LIVE |
| \(x_{t|v}\) | scalar | vintage-aware observation | CONCEPTUAL |
| \(\mathcal{F}_t^{\text{data}}\) | closed set | freshness state | LIVE |
| \(A_t\) | closed set | authority state | LIVE |
| \(b, \mathcal{C}_b, N_b\) | various | AI briefing type, context, narrative | LIVE |
32. Non-Claims
Three statements, made explicitly to prevent inferential drift.
ATLAS does not assert
as a universal trading claim.
ATLAS does not define
as an unconstrained return-maximization problem.
ATLAS does define
The objective is exposure governance under regime instability, subject to validation, operator scope, and layer separation. It is not portfolio-return maximization under regime stability.
References
Ang, A. & Timmermann, A. (2012). Regime changes and financial markets. Annual Review of Financial Economics, 4, 313–337.
Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.
Chow, C. K. (1970). On optimum recognition error and reject tradeoff. IEEE Transactions on Information Theory, 16(1), 41–46.
Diebold, F. X. & Rudebusch, G. D. (1996). Measuring business cycles: A modern perspective. Review of Economics and Statistics, 78(1), 67–77.
El-Yaniv, R. & Wiener, Y. (2010). On the foundations of noise-free selective classification. Journal of Machine Learning Research, 11, 1605–1641.
Hamilton, J. D. (1989). A new approach to the economic analysis of nonstationary time series and the business cycle. Econometrica, 57(2), 357–384.
Harvey, C. R., Liu, Y. & Zhu, H. (2016). … and the cross-section of expected returns. Review of Financial Studies, 29(1), 5–68.
Popper, K. R. (1959). The Logic of Scientific Discovery. Hutchinson.
Taleb, N. N. (2007). The Black Swan: The Impact of the Highly Improbable. Random House.
Tetlock, P. E. & Gardner, D. (2015). Superforecasting: The Art and Science of Prediction. Crown.
Contact
Ask about ATLAS