Parallax Judge House View

Judge Active House View

When not to use

• Synthesizing a new house view from MCP signals → use /parallax-make-house-view
• Building a portfolio from a thesis → use /parallax-portfolio-builder
• Editing the saved view → use /parallax-load-house-view --edit
• One-off scenario / what-if → describe inline to the portfolio skill, don't run the judge
• Internal-consistency stress test (CIO contradictions inside the view itself) → use /parallax-stress-house-view

Gotchas

• This skill is READ-ONLY against the active house view. It never writes to view.yaml or provenance.yaml. Only audit.jsonl gets a new row (action="judge", applied=false ALWAYS) and the judge-reports bundle directory gets a new entry.
• JIT-load _parallax/house-view/loader.md for the audit-row schema (§6.1 / §6.2) and the conditional fields for action="judge".
• JIT-load _parallax/house-view/schema.yaml for the view structure + classification_taxonomy.judge_recommendation (judge does NOT write this class into provenance.yaml — it only emits it in the report; promotion happens via /parallax-load-house-view --edit).
• JIT-load _parallax/house-view/MCP_FIELD_INVENTORY.md to understand what the maker can actually extract — informs which cells the judge can confidently recommend on.
• JIT-load _parallax/parallax-conventions.md for MCP tool conventions and batch patterns.
• Phase 5 LLM recommendations go through recommendation.validate_citation post-call. Recommendations whose rationale does NOT contain a >=30-char verbatim substring of the source snippet are DROPPED and replaced with a "judge declined to recommend (citation check failed)" placeholder. This is the single biggest hallucination control — do NOT bypass it.
• The judge does NOT use gate_present.run_gate_loop — there is no confirmation gate. It's a read-only report.
• Maker shared modules (cross_country, pillar_compose, pillar_formulas) are imported lazily; if Phase B1 hasn't shipped, the orchestrator surfaces the gap in diagnostics and falls back to PARALLAX_SILENT for cells where the imputed view can't be computed.
• Server-side house_view_judge MCP tool: CANCELLED (2026-05-24). The judge is client-side permanently — bank clients run this skill / CLI on their own side. See v2 plan §3.2 for rationale (transparency for model-validation review, zero cross-tenant blast radius). Do NOT resurrect the server-side framing without a new architectural decision.
• Auto-on-load triggers (portfolio-builder, rebalance, thematic-screen) suppress the run when view_age_days is independent — it replaces the live MCP fan-out with a canned JSON payload keyed by tool:arg1:arg2:... summary strings (for tests or CI). Either flag can be used alone or together.

Cost: ~282 tokens at the default market set (~14 markets × 4 components; see _parallax/token-costs.md). --dry skips the Phase 5 LLM step but NOT the Phase 1 macro fan-out — the full ~282-token cost is still incurred.

Workflow (Phases 0-8 per v2 plan §3.1)

Phase 0 — Load active view

Invoke judge.phase_0_load_view(), which wraps stress.load_active_view(). This:

• Reads ~/.parallax/active-house-view/view.yaml.
• Verifies the audit.jsonl hash chain (raises on broken chain).
• Computes view_hash.

Halt conditions: view directory missing → exit with the standard "no active view" message. Chain broken → propagate the audit_chain error.

Phase 1 — MCP fan-out

Same recipe as the maker: 14 markets × 4 components (macro_indicators, tactical, sectors, news) + 1 get_telemetry call. Concurrency capped at 8. (fixed_income is deferred — no v0 consumer; re-add in lockstep with the maker when a rates leg lands.)

Per-market timeout: 45s. UNREACHABLE markets are classified via stress.classify_mcp_meta_state and weighted to 0 by the maker's aggregator. If unreachableshare > 30% across markets, the maker raises; the judge surfaces this in diagnostics and falls back to PARALLAXSILENT.

PARTIAL responses (success=True with prose like "data unavailable"): treat as silent — not UNREACHABLE.

When mock_mcp_responses is provided (via --mock-mcp or programmatic injection), it replaces the live MCP fan-out verbatim regardless of --dry; otherwise the runtime delegates to the injected mcp_call_fn. --dry is orthogonal — it only suppresses the Phase 5 LLM recommendation call, not the MCP source.

Phase 2 — Per-cell diff

For each non-zero cell in the active view (enumerated via stress.enumerate_dimensions), call stress.resolve_cell_state(cio_tilt, parallax_view, age_delta, market=..., covered_markets=...). States:

• ALIGNED
• DIVERGENT_FRESH (sign mismatch; views are not stale)
• DIVERGENT_STALE (sign mismatch; view is stale relative to Parallax)
• CIO_SILENT (uploader has no view; Parallax does)
• PARALLAX_SILENT (uploader has a view; Parallax silent)
• UNCOVERED (market not in Parallax coverage)

parallax_view per cell comes from the imputed view — cross_country.aggregate + pillar_compose.compute_pillars applied to the MCP fan-out. If the maker modules aren't installed yet, the orchestrator returns PARALLAX_SILENT for every cell and notes the gap in diagnostics.

Phase 3 — Severity classification

Run drift_classify.classify_severity(resolutions, view_age_days, denominator). Tiers:

• drift_minor: 50%, OR any DIVERGENT_* on a pillar, OR sign-flip on tilts.macro_regime.*

Magnitude escalation: any |cio_value - parallax_value| >= 3 bumps severity one tier (minor → moderate → material). Capped at drift_material (the drift_breaking slot is reserved for future tightening).

Phase 4 — Build recommended deltas

Call stress.build_recommended_deltas(resolutions, cio_age, parallax_age, include_fresh=True). Per A1 work:

• DIVERGENT_FRESH cells get kind="informational_fresh"
• DIVERGENT_STALE cells get kind="informational" (unchanged)
• All pass stress.validate_recommended_deltas (allowlist already extended).

Phase 5 — LLM-as-judge recommendations (material+ severity only)

For the cells at drift_material (or stricter), the orchestrator:

1. Calls recommendation.build_recommendation_prompt(...) for EVERY material cell first — each prompt is built only from its own cell's snippet, so the cells are mutually independent.
2. Dispatches those cells to Claude via the runtime-injected llm_call_fn in parallel where the runtime supports concurrent dispatch; ordering within the batch is immaterial because no cell reads another's output and Phase 6/7 consume the completed set. Each response is validated by recommendation.validate_citation (step 3 below) as it returns — parallelism does NOT touch or weaken the per-cell >=30-char verbatim-substring check.
3. Validates the citation via recommendation.validate_citation. The LLM's rationale (or any evidence_refs entry) MUST contain a >=30-char verbatim substring of the source snippet. Honest declines (recommended_value=null + "insufficient evidence" rationale) bypass the substring check.
4. On validation failure, the recommendation is dropped and replaced with recommendation.make_decline_placeholder(...). The decline marker ("judge declined to recommend (citation check failed)") is visible in both the report and the audit row.

Prompt template: see recommendation.SYSTEM_PROMPT + build_recommendation_prompt. The structured-output schema constrains recommended_value ∈ {-2,-1,0,1,2,null}, confidence ∈ [0,1], rationale ≤ 300 chars, suggested_basis_statement_addendum ≤ 100 chars.

Do not weaken the citation validator. No bypass flag, no debug knob. Loosening to "semantic match" would re-introduce the hallucination surface the validator exists to close.

Phase 6 — Render report bundle

Write to ~/.parallax/judge-reports/-/:

• report.md — client-facing markdown (severity verdict, per-cell table, recommendations)
• report.json — structured form for cron / automation consumption
• mcp_responses.jsonl — one line per MCP call: {call: "tool:args", response: ...}
• reasoning_chain.yaml — written by chain_emit (Phase 8), lives under ~/.parallax/reasoning-chains/
• audit_entry.json — copy of the appended audit row for offline auditors

Phase 7 — Append single audit row

Per loader.md §6.1/§6.2, append exactly one row with:

{
  "schema_version": 1,
  "ts": "...",
  "view_id": "...",
  "version_id": "...",
  "view_hash": "...",
  "skill": "parallax-judge-house-view",
  "action": "judge",
  "applied": false,
  "judged_view_id": "...",
  "judged_version_id": "...",
  "view_age_days": 42,
  "parallax_age_days": 3,
  "drift_summary": {
    "aligned_count": 7,
    "drift_minor_count": 3,
    "drift_material_count": 0,
    "drift_breaking_count": 0,
    "parallax_silent_count": 2,
    "uncovered_count": 0
  },
  "recommendations": [/* deltas from Phase 4 */]
}

applied is ALWAYS false for action="judge". The judge never modifies the view. Acceptance happens later, manually, via /parallax-load-house-view --edit citing the judge's audit hash in the basis_statement (a future --apply-judge flag will automate this).

Phase 8 — Emit reasoning chain

Call chain_emit.emit_phase_0_chain with:

• skill_version="parallax-judge-house-view@1.0.0"
• skill="parallax-judge-house-view"
• base_scores={"response_inline": , "response_hash": }
• final_portfolio={"weights": {}} (judge produces no portfolio; chain spec §3.5 allows empty weights)

The chain artifact lands in ~/.parallax/reasoning-chains//.yaml. Replayable: the response_hash is the determinism anchor.

Trigger sources (cadence.py)

| Trigger | When | Notes | |---|---|---| | on_demand | explicit /parallax-judge-house-view | full pipeline always runs | | auto_on_load | consumer skill (portfolio-builder, rebalance, thematic-screen) auto-invokes | gated on view_age_days >= 30; banner only at drift_material; cache TTL 1h | | scheduled | bank cron via --json | same pipeline, JSON-only output |

Output

Primary output is the markdown report at ~/.parallax/judge-reports//report.md. In --json mode, the JSON sidecar is also written to stdout for downstream consumption. The audit row is the canonical machine surface.

AI-interaction disclosure: report.md ends with the banner rendered per parallax-conventions.md §9.2. The per-cell judgments and Phase 5 recommendations are LLM-generated and the report is read directly by a natural person (the CIO/operator), so §9.2 applies even though this is an operator-facing artifact — the skill does NOT qualify for the config-artifact exemption (it emits an analysis report, not a gated configuration bundle; see conventions §9.2 exemption rationale). The --json sidecar is machine-consumed and carries no banner.

Not on the roadmap

• Server-side house_view_judge MCP tool — CANCELLED, not deferred. The judge runs client-side only. Bank clients schedule their own cron against /parallax-judge-house-view --json on their side. Rationale: methodology transparency for model-validation review (same reason maker is client-side), and zero cross-tenant blast radius from a broken judge run. See v2 plan §3.2.

Deferred (still on roadmap)

• /parallax-load-house-view --apply-judge — the future flag that auto-populates an edit draft from the audit row's recommendations field. Schema is already in place (recommendations carries the same shape as stress's recommended_deltas).

Calibration disclosure

This skill operates against views in calibration_status: heuristic_phase0. The per-cell recommendations and severity thresholds are heuristic; intended for directional research only — do not use for regulatory capital, fiduciary-grade portfolio construction, or client-facing recommendations without further validation.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bencharoenwong
• Source: bencharoenwong/parallax-workflows
• License: MIT

Install and usage instructions live in the source repository linked above.