Reflect 3 Attempt 02

Asteria Investment Office — Institutional Portfolio Risk Solver Skill

Reusable workflow rules for solving Asteria Investment Office portfolio-risk tasks. Base environment: <remote-env-url> (Asteria Investment Office API). All current data comes from this shared service; local task payloads are intake context only and may be stale. Three task families share this service: (1) energy/fixed-income credit trade strategy, (2) international equity correlation review, (3) cross-asset active allocation view updates.

0. Universal rules

• ALWAYS treat the shared environment service as the book of record. Local payloads (desk_request, risk_meeting_memo, committee_request, review_request, allocation_request) are intake context that may carry stale marks, stale quantities, or prior-week shortlists. When a payload value conflicts with the current environment, use the environment. The data_precedence answer for such cases is current_environment_over_stale_payload.
• The environment emits a stale_data_warning and payloads carry snapshot dates; any worksheet dated before the environment as_of_date must be reconciled to the service before deciding. Use current portfolio quantities (not stale exception-board quantities).
• as_of_date in answers = the environment as_of_date (e.g. 2026-05-29), which is consistent across /api/portfolios, /api/policies, /api/market/energy, etc.
• Keep every numeric field rounded to the precision declared in the task's answer_template.json. Do not add fields; do not include narrative outside the JSON.
• Ordering rules are load-bearing: honor every ordering / Sort ... instruction in the template exactly (ascending by instrument_id, SELL before BUY, alphabetical pair ids, focus_opportunity_sets request order, sleeve order, etc.).

1. Environment endpoints that matter

• GET /api/catalog — all valid ids (portfolios, bonds, issuers, indices, opportunity sets, policies). Use to confirm membership before emitting any id.
• GET /api/policies — the policy set. Key policies:
  • credit_default: duration_band_years [3.0,5.0], max_hy_allocation_pct 20.0, issuer_concentration_limit_pct 12.0, subsector_min_count_for_diversified 2, target_hy_reduction_pct 0.0.
  • credit_risk_reduction: same bands/caps but target_hy_reduction_pct 4.0 (the minimum preferred HY reduction for risk-reduction rotations).
  • correlation: correlation_high_threshold 0.8, correlation_low_threshold 0.2, review window (e.g. 2025-05-30 to 2026-04-30).
  • allocation_mapping (POL_ALLOCATION_MAPPING): view_score_thresholds (OW_min 0.35, UW_max -0.35, neutral between), conviction thresholds (HIGH_abs_min 0.7, MEDIUM_abs_min 0.35, LOW_abs_below 0.35), view_rank (OW 1, N 0, UW -1).
  • multi_asset (POL_MULTI_ASSET_DEFAULT): master multi-asset policy that uses_allocation_mapping + correlation + credit defaults.
  • multi_asset_risk (POL_MULTI_ASSET_RISK): uses credit_risk_reduction + correlation; committee_escalation_threshold "two_or_more_material_exceptions".
• GET /api/portfolios and /api/portfolios/<id> — portfolio summary, current holdings (instrument_id, quantity_usd_m, sleeve, notes), constraints, market_value_usd_m.
• GET /api/instruments/bonds (and ?candidate=true) — bond master. Fields: instrument_id, issuer_id, rating, rating_bucket (IG/HY), sector, subsector, energy_linked, modified_duration_years, yield_to_maturity_pct, coupon_pct, spread_bps, maturity, recommended_theme_tags, candidate flag.
• GET /api/issuers — issuer research: rating_bucket, sector, subsector, credit_outlook, research_tags, and a boolean watchlist (the watchlist flag is the authoritative watchlist-source; do not infer watchlist status from theme tags alone).
• GET /api/market/energy — energy signals: pitch_themes and per-commodity signals with score and direction. Use these to pick the sales/positioning theme.
• GET /api/indices, /api/index-levels, /api/index-levels/<index_id> — index metadata and monthly levels (date, level). Correlations are computed from these levels; the API gives NO precomputed correlations.
• GET /api/allocation/opportunity-sets — taxonomy mapping each opportunity_set to an asset_class (Equities/Duration/Credit/Currency) and display_order.
• GET /api/allocation/prior-views — prior-quarter views. Each entry has quarter, previous_quarter, opportunity_set, view, conviction. For a target-quarter refresh, the PRIOR view = the entry whose previous_quarter equals the prior quarter (e.g. for a Q2_2026 refresh with prior_quarter Q1_2026, use entries where previous_quarter=Q1_2026). Do not treat the endpoint's quarter field as the new answer; it labels the transition, and the view field is the prior-quarter input.
• GET /api/macro-signals — per-opportunity_set signal score, rationale_code, drivers, quarter. For a target quarter, use the entries with that quarter.

2. Workflow A — Energy / fixed-income credit trade strategy

Templates require: portfolio_id, as_of_date, trade_package (sorted ascending by instrument_id), post_trade_metrics, constraint_checks, sales_positioning, data_precedence. A fixed-income rotation variant adds rotation.trades, risk_metrics, exception_flags, watchlist_handling, risk_note_code.

Selection logic (CONFIRMED by judge feedback)

• Capacity/spec: honor the exact ticket count, total notional, and split (e.g. exactly two BUY tickets totaling USD 8.0M split evenly = 4.0M each).
• Eligibility filters, all required:
  1. energy_linked = true when the desk preference names energy exposures (LNG exporters, gas demand, energy-linked carry).
  2. Issuer watchlist = false — NEVER select a watchlisted issuer. Watchlisted issuers in this environment include the shale/E&P, telecom, and refining names; confirm via /api/issuers each run.
  3. Modified duration INSIDE the policy band [3.0, 5.0]. Duration-ineligible distractors are any bond with modified_duration_years > 5.0 or < 3.0 — exclude them even if their carry looks attractive. Long-dated LNG/oil bonds and very short refiner bonds are classic distractors here.
  4. The two selected buys must be from DISTINCT issuers AND DISTINCT subsectors (selected_issuer_diversification_pass and selected_subsector_diversification_pass both require the selected set to span >=2 issuers and >=2 subsectors).
• Carry objective = MAXIMIZE expected carry (weighted YTM uplift) subject to the above. Among eligible non-watchlist energy-linked HY carry bonds inside the duration band, pick the HIGHEST yield_to_maturity. A lower-carry renewables HY bond is NOT an acceptable substitute when a higher-carry non-watchlist energy HY bond (e.g. merchant power) is eligible and satisfies diversification. This carry-maximization was the decisive pass criterion.
• The anchor buy is the LNG/natural-gas IG name that matches LNG_EXPORTS / GAS_DEMAND theme tags and the strongest positive energy signal (LNG export pull). Pair it with the highest-YTM eligible non-watchlist energy HY carry bond from a DIFFERENT subsector.

Post-trade metric computation

Using current portfolio holdings (quantity_usd_m) + the new BUY quantities:

• total_market_value_usd_m = sum of all post-trade holding quantities. Precision 2.
• hy_allocation_pct = (sum of post-trade quantities whose rating_bucket == "HY") / total_market_value_usd_m * 100. Precision 2.
• weighted_modified_duration_years = sum(quantity * modified_duration_years) / total_market_value_usd_m. Precision 2.
• weighted_yield_to_maturity_pct = sum(quantity * yield_to_maturity_pct) / total_market_value_usd_m. Precision 2.
• notional_usd_m per ticket: precision 1. Weights are MARKET-VALUE (quantity) weights, not notional-equal weights.

Constraint checks (booleans)

• hy_cap_pass: post-trade hy_allocation_pct <= max_hy_allocation_pct (20.0).
• duration_band_pass: weighted_modified_duration_years within [3.0, 5.0].
• selected_issuer_diversification_pass: the selected buys span >=2 distinct issuers.
• selected_subsector_diversification_pass: the selected buys span >=2 distinct subsectors.
• watchlist_avoidance_pass: no selected buy is from a watchlisted issuer.

sales_positioning

• target_segment: match the desk client_context (e.g. "multi-asset income update" -> multi_asset_income).
• theme: match the dominant POSITIVE energy signal / pitch theme. For an LNG-anchored package the theme is lng_export_tailwind. (Other allowed themes: oil_oversupply_caution, midstream_stability, transition_bond_selectivity, avoid_watchlist_yield_trap.)

data_precedence

• current_environment_over_stale_payload whenever the local worksheet pre-dates the service and operations has not reconciled it.

Fixed-income rotation variant (PF-FI-* risk rebalance)

• The rotation MUST bring HY to <= the 20% cap (hy_cap_pass = true). Submitting a rotation that leaves HY above the cap (hy_cap_pass false) scores badly. This is a hard pass criterion.
• HOWEVER the rotation must also PRESERVE CARRY: do NOT eliminate all HY. Sell the watchlisted bond(s) fully (watchlist_exposure_cleared = true, post-trade watchlist exposure 0) PLUS enough non-watchlist HY to get under the 20% cap, while KEEPING the highest-carry non-watchlist HY bond for carry. Fully rotating HY to 0 is penalized; keeping one carry HY while clearing the cap is rewarded.
• target_hy_reduction_met: hy_reduction_pct_points (pre-trade hy% - post-trade hy%)

  = the policy target_hy_reduction_pct (4.0 for POL_CREDIT_RISK_REDUCTION). The 4pp is a FLOOR; the 20% cap is the BINDING constraint that usually forces a larger cut.

• Sells = the watchlisted holding(s) + the lower-carry non-watchlist HY needed to clear the cap. Buys = the eligible non-watchlist IG candidates from the desk shortlist (exclude any candidate whose issuer is watchlisted — "risk team concerned about issuer status" means watchlist). buys_avoid_watchlist = true.
• watchlist_sell_ids: instrument_ids of watchlisted bonds sold, ascending order.
• Trade ordering: SELL before BUY, then instrument_id ascending within each action.
• risk_note_code MUST be consistent with the flags. If hy_cap_pass = true (cap met), do NOT use hy_cap_pressure — use carry_tradeoff (gave up HY carry for IG safety) or watchlist_concentration/duration_preservation as fits. Reserve hy_cap_pressure for cases where the cap remains breached.
• Quantity precision 1; post_trade_hy_allocation_pct / duration / reduction precision 2; post_trade_watchlist_exposure_usd_m precision 1.

3. Workflow B — International equity correlation review

Templates require: portfolio_id, review_window {level_start_date, level_end_date, return_observations}, index_set, extreme_pairs {highest_positive, lowest}, concentration {china_asia_dependence_flag, primary_code, high_threshold_breached}, diversification_candidates, sleeve_actions.

Correlation computation (CONFIRMED)

• Pull monthly levels for each index in the request universe over [level_start_date, level_end_date].
• Compute monthly SIMPLE returns: r[i] = level[i+1]/level[i] - 1.
• return_observations = number of returns = (number of levels) - 1. For a 12-month window (12 monthly levels) this is 11. Do NOT report the level count.
• Pearson correlation between each pair of return series: r = cov(a,b) / (stdev(a)*stdev(b)). Round to THREE decimals.
• index_set = the request universe, sorted ascending. ASCII sort: uppercase letters (A-Z, 65-90) sort before underscore _ (95), so IDX_ACWI_IMI precedes IDX_AC_ASIA_PAC_EX_JP (the W before _). Verify each run with a real sort.
• extreme_pairs.highest_positive = the off-diagonal pair with the MAX correlation; extreme_pairs.lowest = the pair with the MIN correlation (most negative is allowed and expected). Each pair_id is a 2-element list sorted alphabetically; correlation to 3 decimals.
• Do NOT use any precomputed correlation field — compute from levels.

Concentration

• china_asia_dependence_flag = true when the China and/or Asia-Pacific sleeves are present and their pair correlations breach the high threshold (0.8).
• high_threshold_breached = true when ANY pair correlation (absolute or positive per the policy) exceeds correlation_high_threshold (0.8). In a tightly-overlapping universe many pairs breach; report true.
• primary_code: the dominant concentration pattern. Use CHINA_ASIA_DEPENDENCE when the memo concern codes are Asia/China-focused and the China<->Asia/EM correlations are above 0.8. Use GLOBAL_DEVELOPED_OVERLAP when the single highest pair is a global/developed overlap (e.g. EM<->World) that dominates. Choose based on which concentration the data + memo jointly emphasize; when unsure, the memo's stated concern codes are the tiebreaker. Allowed: CHINA_ASIA_DEPENDENCE, GLOBAL_DEVELOPED_OVERLAP, NO_MATERIAL_CONCENTRATION.

Diversification candidates

• Candidates = indices whose correlations to the concentrated block fall below the low threshold (0.2) — i.e. the genuine low-correlation diversifiers. In this environment the Latin America index is the consistent diversifier (negative correlations to the China/Asia/EM block). Emit ascending by index id.

Sleeve actions (CONFIRMED)

• Exactly two actions, ordered ascending by sleeve. Use trim on the concentrated sleeve (e.g. trim the dedicated China sleeve) and add on the low-correlation diversifier (e.g. add Latin America).
• DO NOT use rotate for these review actions — using rotate (rotating the concentrated sleeve into an EM-ex-China index) was CONFIRMED to lower the score. The review action set is trim-the-concentration + add-the-diversifier.

4. Workflow C — Cross-asset active allocation view updates

Two variants: (C1) a pure allocation-view memo (task 003 type) and (C2) a committee JSON linking correlation findings to allocation views (task 005 type).

Deriving views (CONFIRMED correct)

• For each requested opportunity_set, read the prior view from /api/allocation/prior-views (entry whose previous_quarter = prior_quarter) and the current signal from /api/macro-signals (entry whose quarter = target_quarter).
• New view from signal score via POL_ALLOCATION_MAPPING thresholds: score >= 0.35 -> OW; score <= -0.35 -> UW; otherwise N.
• conviction from abs(score): >= 0.7 -> HIGH; >= 0.35 -> MEDIUM; < 0.35 -> LOW.
• rationale_code = the macro-signal's rationale_code directly (do not invent).
• change vs prior view using view_rank (OW=1, N=0, UW=-1): new_rank > prior_rank -> UP; < -> DOWN; equal -> UNCHANGED.
• asset_class from /api/allocation/opportunity-sets taxonomy (Equities / Duration / Credit / Currency).
• signal_score (C2 only) = the macro signal score, precision 3.
• prior_view (C2 only) = the prior-quarter view code.
• Order rows by the request payload's focus_opportunity_sets order (NOT alphabetical).
• target_quarter / prior_quarter are required fixed values from the request.

Risk overlay (C1)

• overlay_code, primary_action, rationale_codes (business priority, highest first). The overlay should reflect the dominant signal theme across the focus set (e.g. duration-OW + HY-UW + IG-OW => a quality/duration tilt; HY valuation risk => credit risk reduction). Allowed overlays: DURATION_QUALITY_TILT, CREDIT_RISK_REDUCTION, EQUITY_BETA_EXTENSION, CURRENCY_DEFENSIVE_HEDGE, NO_OVERLAY. Pick the one matching the strongest risk/repositioning signal and keep primary_action paired with its overlay code.
• policy_id: report the GOVERNING multi-asset policy (POL_MULTI_ASSET_DEFAULT) as the lineage policy_id for a multi-asset reference model, not the sub-policy POL_ALLOCATION_MAPPING (which is the threshold source, not the portfolio policy).

Committee JSON (C2 — PF-MA-HELIO type)

Required keys: portfolio_id, as_of_date, review_quarter, correlation_summary, target_sleeve_actions, allocation_views, rebalance_trigger, portfolio_risk_concentration_flag, next_step.

• correlation_summary: a 2-item list, item_order [highest_concentration, best_diversifier]. Each item: {pair_role, pair (2 index ids alphabetical), correlation (Pearson of monthly simple returns, 3 decimals)}. highest_concentration = the pair with the MAX correlation in the requested index set; best_diversifier = the pair with the MIN (most negative) correlation. Use the same monthly-return Pearson method as Workflow B over the stated 12-month window.
• allocation_views: one row per requested opportunity_set (order = request order), with prior_view, signal_score (3 dec), view, change, conviction, rationale_code derived as above.
• target_sleeve_actions: one per requested opportunity_set (order = request order). Map the NEW VIEW to the action: OW -> add, UW -> trim, N -> hold. (For a sleeve whose new view is N after being OW, use hold — do not trim a neutral currency view.) This view-based mapping is preferred.
• rebalance_trigger: the substantive breach that drove the review. When a pair correlation exceeds correlation_high_threshold (0.8), use correlation_cap_breach. (Other allowed: hy_cap_pressure, duration_drift, watchlist_concentration, committee_review.)
• portfolio_risk_concentration_flag: true when any reviewed pair breaches the high threshold (correlation > 0.8) or the portfolio carries a concentrated high-correlation sleeve pair.
• next_step: approve_with_monitoring when a rotation is proposed that addresses a flagged concentration. CONFIRMED: approve_rotation (clean approve, no monitoring) is WRONG when a concentration is flagged — it lowers the score. With a flagged risk concentration, approve WITH monitoring. (Use defer_pending_risk_review only for two-or-more material exceptions; reject_constraint_breach only when the rotation cannot resolve the breach.)

5. Precision & formatting quick reference

6. Confirmed misjudgment / exclusion rules (do not violate)

1. Watchlist avoidance: never BUY a watchlisted issuer; never keep a watchlisted holding in a risk-reduction rotation. Verify watchlist on /api/issuers.
2. Duration-ineligible distractors: exclude any bond outside [3.0, 5.0] modified duration even if carry is high; classic traps are long-dated LNG/oil (>5.0) and short refiner/shale (<3.0) paper.
3. Carry maximization: among eligible non-watchlist energy HY carry bonds inside the band, choose the HIGHEST YTM that preserves issuer+subsector diversification. Do not substitute a lower-carry renewables HY bond when a higher-carry eligible name exists.
4. HY cap is a hard pass: a fixed-income rotation MUST end with hy_allocation_pct <= 20.0 (hy_cap_pass true). Leaving HY above the cap is a severe failure.
5. Preserve carry: a risk-reduction rotation keeps the highest-carry non-watchlist HY bond; do not drive HY to 0.
6. Stale-worksheet precedence: use current environment quantities/marks over the local payload snapshot (data_precedence = current_environment_over_stale_payload).
7. Correlation review sleeve actions: trim the concentrated sleeve + add the diversifier. Do NOT use rotate (confirmed score loss).
8. Committee next_step: approve_with_monitoring (not approve_rotation) when a risk concentration is flagged.
9. risk_note_code consistency: do not report hy_cap_pressure when hy_cap_pass is true.
10. Correlations: compute Pearson from monthly simple returns on index levels; never use a precomputed correlation field; return_observations = levels - 1.

7. SOP per workflow

SOP A — energy trade strategy

1. GET /api/portfolios/, /api/instruments/bonds, /api/issuers, /api/market/energy, /api/policies.
2. Filter bonds: energy_linked=true, watchlist=false, modified_duration in [3,5], candidate=true (or held). Mark duration-ineligible distractors.
3. Pick the LNG/gas IG anchor (LNG_EXPORTS/GAS_DEMAND tags, strongest LNG signal).
4. Pick the highest-YTM eligible non-watchlist energy HY carry bond from a DIFFERENT issuer and subsector.
5. Size tickets to the spec (e.g. 4.0M each), sort ascending by instrument_id.
6. Compute market-value-weighted post-trade metrics at the required precision.
7. Set constraint_checks (hy_cap, duration_band, selected issuer/subsector diversification, watchlist avoidance).
8. Set sales_positioning (target_segment from client_context; theme from dominant energy signal) and data_precedence = current_environment_over_stale_payload.

SOP B — correlation review

1. GET /api/portfolios/, /api/index-levels, /api/policies, /api/indices.
2. For the request universe, compute monthly simple returns and the full Pearson matrix.
3. Set review_window (dates from request; return_observations = levels - 1).
4. Set index_set sorted ASCII ascending.
5. Set extreme_pairs (highest_positive = max pair; lowest = min pair; pair ids alphabetical; 3 decimals).
6. Set concentration (china_asia_dependence_flag, primary_code per memo+data, high_threshold_breached).
7. Set diversification_candidates = indices with correlations below the low threshold vs the concentrated block.
8. Set sleeve_actions = [trim concentrated sleeve, add diversifier], ascending by sleeve.

SOP C — allocation view refresh

1. GET /api/allocation/opportunity-sets, /api/allocation/prior-views, /api/macro-signals, /api/policies.
2. For each requested opportunity_set: prior view (prior-views entry with previous_quarter = prior_quarter); signal (macro-signals entry with quarter = target_quarter).
3. Derive view (>=0.35 OW, <=-0.35 UW, else N), conviction (>=0.7 HIGH, >=0.35 MEDIUM, else LOW), rationale_code (from signal), change (vs prior view_rank), asset_class (taxonomy).
4. Order rows by the request focus_opportunity_sets order.
5. Set risk_overlay (overlay_code + paired primary_action + priority rationale_codes) and policy_id (governing multi-asset policy).
6. For committee JSON (C2): add correlation_summary (highest_concentration / best_diversifier pairs from monthly-return Pearson), target_sleeve_actions (OW->add, UW->trim, N->hold in request order), rebalance_trigger (correlation_cap_breach when a pair >0.8), portfolio_risk_concentration_flag (true when >0.8), next_step (approve_with_monitoring when a rotation addresses a flagged concentration).