Reflect 3 Attempt 02

Reflect-3 Asteria JSON Workflow

Use this skill for Asteria Investment Office tasks that ask for compact JSON decisions using the shared environment API plus a local request/template packet.

Source Precedence

• Treat the local prompt and payload as the contract: requested portfolio, task window, field names, enum choices, item order, and rounding precision.
• Treat the shared environment as the book of record for current portfolios, holdings, policies, instruments, issuers, index levels, opportunity sets, prior views, and macro signals.
• If a local memo contains stale snapshots, quantities, marks, or preferences, use it only as context. Current environment records override stale local values for calculations and eligibility.
• Use the environment as_of_date from the current record that drives the answer. For allocation lineage, prefer the top-level policy-set id when the template asks generically for policy_id; use a nested policy id only when the task/template specifically identifies that policy.

API Habits

• Start with /api/policies to get dates, thresholds, score mappings, duration bands, HY caps, and policy identifiers.
• Fetch the specific portfolio detail with /api/portfolios/<portfolio_id> instead of relying on portfolio list summaries.
• For credit tasks, join portfolio holdings to /api/instruments/bonds and /api/issuers by instrument_id and issuer_id.
• For correlation tasks, use /api/indices for the official level window and /api/index-levels or /api/index-levels/<index_id> for monthly levels.
• For allocation tasks, combine /api/allocation/opportunity-sets, /api/allocation/prior-views, and /api/macro-signals; filter by the requested quarter and opportunity sets.
• Equality filters are useful, but verify the shape returned by the API before assuming a filtered endpoint returns only one object.

JSON Contract Discipline

• Build the answer from the template, not from memory. Include required keys only unless the prompt asks for extras.
• Use exact enum strings from the template. Do not invent synonyms.
• Preserve the template's item order when it gives one: request order, explicit item_order, action ordering, or allowed-value order. For ambiguous "alphabetical" instructions, prefer the template/request ordering when it is explicit, and always sort pair ids within each pair as requested.
• Sort trade lists exactly as specified, commonly SELL rows before BUY rows and then ascending instrument_id within each action.
• Round only at the final output step, using the precision declared in the template. Keep internal calculations unrounded.

Credit And Bond Calculations

• Adjust current holdings by proposed trades, then calculate post-trade metrics from the adjusted book.
• Post-trade market value is current market value plus buys minus sells. For fully funded rotations, it should remain unchanged.
• HY allocation percent: 100 * sum(quantity_usd_m where rating_bucket == "HY") / post_trade_market_value.
• Weighted duration: sum(quantity_usd_m * modified_duration_years) / post_trade_market_value.
• Weighted yield: sum(quantity_usd_m * yield_to_maturity_pct) / post_trade_market_value.
• HY reduction in percentage points: pre-trade HY allocation percent minus post-trade HY allocation percent.
• Watchlist exposure: sum adjusted quantities for instruments whose issuer has watchlist: true.
• For BUY packages, honor all filters in the prompt: candidate, sector or energy linkage, allowed action, watchlist avoidance, issuer/subsector diversification, ticket count, and exact total notional split.
• For rebalance packages, sell current holdings rather than stale memo quantities, clear any required watchlist exposure, and fund buys with current eligible non-watchlist candidates when requested.
• High yield alone should not dominate selection. A client-facing carry package still needs to pass HY caps, duration bands, issuer/subsector diversification, and watchlist rules.

Correlation Calculations

• Use monthly simple returns from consecutive levels: return_t = level_t / level_(t-1) - 1.
• return_observations equals the number of level observations minus one.
• Compute Pearson correlation on the aligned simple-return arrays for each requested pair.
• highest_positive is the maximum correlation in the requested universe; lowest or best diversifier is the minimum correlation unless the prompt narrows the candidate universe.
• Pair ids inside each pair must be sorted by index id before output.
• Use the policy high-correlation threshold for concentration flags and the low-correlation threshold only when the task asks for threshold-qualified diversifiers.

Allocation View Mapping

• Pull the target-quarter macro signal for each requested opportunity set.
• Map signal score to view using the allocation policy thresholds: OW at or above the OW threshold, UW at or below the UW threshold, otherwise N.
• Map conviction from absolute signal score using the policy's LOW/MEDIUM/HIGH thresholds.
• Use the prior-view record for the requested target quarter as the prior state, then compare ranks (UW < N < OW) to emit UP, DOWN, or UNCHANGED.
• Use the opportunity-set taxonomy for asset_class and the macro signal's rationale_code for rationale fields.
• For overlays, choose the enum/action that matches the dominant active risk signal, and order rationale codes by business priority rather than alphabetically when the template says so.

Action Enum Conventions

• For sleeve actions, let the active view drive the direct action: UW usually maps to trim, OW to add, and N to hold or monitor.
• Reserve rotate for an explicit replacement from one sleeve or target to another; do not use it merely because a correlation finding is present.
• Use hedge for currency sleeves when the current view or memo specifically frames the position as a defensive currency offset.
• Use approve_with_monitoring when the package addresses the signal but a concentration/risk flag remains important for committee oversight; use rejection only when constraints remain breached.

Common Pitfalls

• Do not use stale local snapshots for current quantities, market value, HY allocation, or duration.
• Do not buy watchlist issuers to chase yield when the prompt asks for client-safe income or watchlist avoidance.
• Do not broaden diversification candidates beyond the requested universe or policy-qualified candidates.
• Do not round component inputs before computing weighted metrics or correlations.
• Do not output narrative text around the JSON when the prompt asks for JSON only.
• Do not include training answers, gold outputs, evaluation notes, or run artifacts in the working context or final response.