task-group-010-fewshot-attempt-01

name	task-group-010-fewshot-attempt-01
description	Solve PanofyBench task_group_010 Asteria Investment Office input-output tasks that require reading local request/template JSON, using the public Asteria API at http://127.0.0.1:8036 as the current book of record, computing portfolio credit metrics, index correlations, active allocation views, and returning only schema-conforming JSON.

Task Group 010 Input-Output SOP

Use this skill for Asteria Investment Office JSON tasks in task_group_010. Produce only the requested JSON object.

Source Order

1. Read the local input/prompt.txt, request payload, and input/payloads/answer_template.json.
2. Treat the public API at http://127.0.0.1:8036 as authoritative for current portfolio, policy, instrument, issuer, index, and signal records.
3. Use local payloads for request scope, portfolio id, requested universe, ordering, stale-context clues, and output-contract details. Override stale local snapshots or notes with API data.
4. Never use environment files, test files, notes, evaluator files, or hidden answers.

API Checklist

Use GET / if endpoint discovery is needed. Common endpoints:

• /api/catalog: ids for portfolios, policies, indices, issuers, bonds, and opportunity sets.
• /api/policies: policy ids, credit limits, correlation thresholds, and allocation mapping thresholds.
• /api/portfolios and /api/portfolios/<portfolio_id>: current as-of date, constraints, market value, and holdings.
• /api/instruments/bonds: bond security master, candidate flag, rating bucket, duration, yield, sector/subsector, and issuer id.
• /api/issuers: issuer watchlist and research status.
• /api/market/energy: current energy pitch themes and commodity signals.
• /api/indices, /api/index-levels, /api/index-levels/<index_id>: monthly index metadata and levels.
• /api/allocation/opportunity-sets, /api/allocation/prior-views, /api/macro-signals: allocation taxonomy, prior views, signal scores, and rationale codes.

Template Discipline

• Follow the answer template exactly: required keys, required values, enum spellings, precisions, and item ordering.
• Preserve request-specific ordering unless the template says otherwise. Common patterns: allocation rows in request focus_opportunity_sets order; pair ids sorted alphabetically within each pair; trade packages sorted by the template rule.
• Return JSON numbers and booleans as native JSON values, not strings.
• Round only final values to the template precision. JSON may omit trailing zeroes.

Credit And Bond Tasks

Join portfolio holdings to bonds by instrument_id, and bonds to issuers by issuer_id.

For BUY selection:

• Start from candidate: true instruments unless the prompt explicitly permits otherwise.
• Honor requested action count, total notional, split rules, and allowed actions.
• Avoid issuer watchlist buys when requested or when client-facing suitability matters.
• Use energy_linked, recommended_theme_tags, issuer research tags, and /api/market/energy signals to choose theme-fit candidates.
• Prefer current, eligible, non-watchlist carry that keeps HY allocation and duration inside policy limits; do not chase watchlisted high yield.
• For diversification checks, avoid duplicate selected issuers and use at least the policy minimum distinct selected subsectors when required.

For SELL or rotation selection:

• Sell current holdings that create HY pressure, watchlist exposure, or explicit risk exceptions.
• Fund buys with current eligible candidates that restore duration and reduce HY/watchlist pressure.
• Keep total buy/sell quantities balanced when the prompt asks for a rotation rather than a net allocation.

Metric formulas:

• Post market value = current market value + buys - sells.
• HY allocation pct = 100 * sum(post quantities whose bond rating_bucket == "HY") / post market value.
• HY reduction pct points = pre-trade HY allocation pct - post-trade HY allocation pct.
• Weighted duration = sum(post quantity_usd_m * modified_duration_years) / post market value.
• Weighted yield = sum(post quantity_usd_m * yield_to_maturity_pct) / post market value.
• Post watchlist exposure = sum post quantities whose issuer has watchlist: true.
• Duration pass uses the portfolio/policy duration_band_years; HY pass uses max_hy_allocation_pct; target HY reduction uses the applicable policy or local requested minimum.

Correlation Tasks

Use the requested index universe and review window.

1. Fetch levels from /api/index-levels or /api/index-levels/<index_id>.
2. Keep levels from level_start_date through level_end_date, inclusive.
3. Convert to monthly simple returns: return_t = level_t / level_(t-1) - 1.
4. return_observations is the number of returns, normally one less than the number of monthly levels.
5. Compute standard Pearson correlation on aligned monthly returns.
6. Round correlations to three decimals.

Interpretation patterns:

• highest_positive or highest_concentration is the largest positive correlation in the requested universe.
• lowest or best_diversifier is the smallest correlation, often negative.
• Set pair ids in alphabetical index-id order even when the pair role is not alphabetical.
• Use /api/policies correlation thresholds. A concentration flag is true when a relevant concentration pair exceeds the high threshold.
• Use CHINA_ASIA_DEPENDENCE when China, EM, or Asia beta sleeves show high dependence; use diversifier actions for structurally lower-correlation sleeves such as ex-China or Latin America only when supported by the computed correlations and template allowed values.

Allocation View Tasks

Build each allocation row from current API records:

1. Get asset class from /api/allocation/opportunity-sets.
2. Get the prior view from /api/allocation/prior-views for the target quarter; its view is the prior-quarter view for comparison.
3. Get current score and rationale_code from /api/macro-signals for the target quarter.
4. Apply /api/policies.allocation_mapping.view_score_thresholds:
   • score >= OW_min -> OW
   • score <= UW_max -> UW
   • otherwise -> N
5. Apply conviction thresholds to abs(score):
   • >= HIGH_abs_min -> HIGH
   • >= MEDIUM_abs_min -> MEDIUM
   • otherwise -> LOW
6. Compute change with view_rank: current rank greater than prior -> UP, lower -> DOWN, equal -> UNCHANGED.

Risk overlay conventions:

• Prefer DURATION_QUALITY_TILT / tilt_to_duration_quality when duration support combines with credit or China-dependence risk.
• Use CREDIT_RISK_REDUCTION / trim_credit_beta for dominant HY or spread-risk pressure.
• Use CURRENCY_DEFENSIVE_HEDGE / add_currency_hedge for dominant currency-defensive needs.
• Use NO_OVERLAY / hold_policy_weights only when no material risk theme is present.
• Order overlay rationale codes by business priority, not alphabetically.

Enum And Output Pitfalls

• Use enum strings exactly as the template lists them, including case: UW, N, OW, UP, DOWN, UNCHANGED, LOW, MEDIUM, HIGH.
• Do not invent rationale codes, sleeve actions, target segments, pitch themes, trigger codes, or next-step values.
• Use current_environment_over_stale_payload when a template asks for data precedence and the local payload conflicts with API records.
• For trade sorting, obey the template even if it differs from alphabetical action order.
• Include no prose, Markdown, explanations, or extra keys outside the final JSON.