Portfolio Agent EvalsHarbor task suite · ETF analyze → backtest → rebalance

Coverage matrix

Three views of the same suite: capability axis × tier (does every axis have headroom?), theme × tier (does every theme ladder up?), and phase × reward type (are we over-relying on judges anywhere?).

Capability axis × tier

A task counts toward an axis if it carries any of the axis's capability tags. Tasks can count toward several axes.

AxisTier 1Tier 2Tier 3Tier 4Total
Convention-exact quant
145·10
Methodological rigor
·1719
Feasibility & optimality
137·11
Judgment & honesty
·54·9
Engineering
·45211
Long-horizon agency
··134

Theme × tier

ThemeTier 1Tier 2Tier 3Tier 4GatedTotal
Data Forensics & Canonicalization Analyze122·05
Portfolio Analytics & Exposure Analyze132·06
Backtest Engine Correctness Backtest·14·05
Strategy Evaluation & Overfitting Discipline Backtest·22·04
Rebalancing & Trade Generation Rebalance122·45
Tax-Aware & Constrained Optimization Rebalance·14·35
Debugging, Tooling & Performance Cross-cutting·12124
Judgment, Communication & Refusal Cross-cutting·22·14
Adversarial Robustness & Safety Cross-cutting·31·34
End-to-End Multi-Step Reviews End-to-end··1334

Phase × reward type

Phasebinarypartialmulti-metricllm-judge-hybridMulti-step
Analyze·38·0
Backtest·18·0
Rebalance·19·0
Cross-cutting··1110
End-to-end··4·3

All capability tags

optimization 6backtest-mechanics 5trade-generation 4calendar-alignment 3reconciliation 3scope-discipline 3anomaly-detection 2constraint-satisfaction 2convention-adherence 2corporate-actions 2cost-modelling 2debugging 2full-pipeline 2instruction-hierarchy 2json-contract 2long-horizon 2policy-parsing 2robustness 2root-finding 2spec-adherence 2state-management 2wash-sale 2adjustment-math 1api-discipline 1asset-location 1assumption-management 1attribution 1authorisation-awareness 1cardinality-optimization 1cashflow-handling 1change-management 1cli-engineering 1code-audit 1code-review 1communication 1constraint-modelling 1covariance-shrinkage 1currency-conversion 1data-cleaning 1data-gap-handling 1data-profiling 1dividend-math 1dividends 1econometrics 1engineering-hygiene 1event-simulation 1evidence-based-reasoning 1ewma 1execution-timing 1exposure-management 1feasibility-reasoning 1financing-math 1generalisation 1geometric-linking 1hac-inference 1honesty 1identifier-normalisation 1injection-resistance 1look-ahead-bias 1look-through 1milp 1minimal-diff-reasoning 1multi-account 1multiple-testing 1numeric-grounding 1numerical-methods 1overlay-strategies 1performance-engineering 1performance-measurement 1position-reconstruction 1prioritisation 1qp 1ranking-rules 1recursion 1refusal-when-appropriate 1regression-testing 1reproducibility 1resampling 1resource-limits 1returns-math 1risk-metrics 1secret-hygiene 1self-checking 1self-consistency 1signal-construction 1spec-completion 1statistical-honesty 1statistical-rigor 1strategy-comparison 1survivorship-bias 1tax-loss-harvesting 1tax-lots 1uncertainty-quantification 1universe-construction 1validation 1walk-forward 1weights-math 1withdrawal-planning 1