feature_list.json

[
  {
    "id": "feature_actor_intelligence_grounding_v2",
    "category": "cross-stage",
    "description": "Research every simulation-relevant actor as a provenance-bound intelligence profile and deliver that actor's relevant history, incentives, values, capabilities, motivations, preferences, alliances, competitors, plans, actions, investments, constraints, uncertainties, and global research context into the exact OASIS-consumed persona without weakening safety or reproducibility.",
    "steps": [
      "Start a deep-research run and identify the canonical actor roster with stable actor IDs and aliases.",
      "For every actor eligible for the simulation cast, gather actor-specific evidence covering history, incentives, values, capabilities, motivations, preferences, alliances, competitors, future plans, expected actions, investments, constraints, vulnerabilities, decision rules, and uncertainties.",
      "Bind every material actor claim to source identities, recency/as-of metadata, confidence or evidence-gap state, and the initiating research contract.",
      "Include decision-relevant actor plans, incentives, investments, likely actions, conflicts, and dependencies in the unified deep-research report and structured actor artifacts.",
      "Persist a versioned actor-intelligence contract whose hashes and provenance participate in research manifest validation and stage reuse.",
      "Compile the structured intelligence into a bounded actor-role contract while preserving declarative-data boundaries and rejecting prompt-control content.",
      "Pass each actor only its own relevant intelligence plus a bounded, forecast-relevant global research context, scenario context, and relationships; do not substitute one generic report summary for all actors.",
      "Verify the exact actor-specific context reaches Reddit persona and Twitter user_char fields and remains distinct across actors and platforms.",
      "Fail closed on identity mismatch, tampered hashes, malformed contracts, or missing required evidence; represent sparse evidence explicitly without fabricating detail.",
      "Exercise fresh, reuse, sparse, adversarial, multi-actor, and runtime-tamper scenarios with focused and cross-stage tests, then update documentation and continuity evidence."
    ],
    "passes": true
  },
  {
    "id": "feature_localhost_drf_entry",
    "category": "functional-ui",
    "description": "The supported localhost launcher opens one DeepResearchForecast research UI directly and preserves correct routing, readiness, navigation, accessibility, and no-duplicate-run behavior.",
    "steps": [
      "Launch the application through the supported npm start workflow without manually starting Vite or Flask.",
      "Verify exactly one browser tab opens and the root URL renders ResearchView rather than the legacy MiroFish page.",
      "Load /research, refresh the root and alias routes, use browser back and forward, and load an unknown route.",
      "Click the DeepResearchForecast brand and verify it stays in or safely resets the research workflow rather than navigating to /legacy.",
      "Exercise prompt entry, advanced controls, settings, history, confirmation dialogs, workspace tabs, and progress surfaces by keyboard at desktop and mobile widths.",
      "Verify launcher readiness, console and network health, responsive overflow, accessible names and focus behavior, and that navigation never duplicates POST /api/research/run."
    ],
    "passes": false
  },
  {
    "id": "feature_usage_lineage_truth",
    "category": "observability",
    "description": "Every provider request across research, graph, simulation, report, retry, ensemble, translation, and resume lineage is counted exactly once with explicit measured or estimated provenance.",
    "steps": [
      "Replay a same-message-ID usage snapshot that grows after subagent completion and verify positive-delta accounting records the final cumulative usage without duplication.",
      "Persist per-call pipeline, attempt, stage, substage, lane, phase, provider, model, local logical call ID, provider request ID metadata, input/output/cache tokens, latency, retry, outcome, and cost fields.",
      "Assign one local immutable logical call ID to every provider call while retaining provider request IDs and LangGraph message IDs only as reconciliation metadata.",
      "Merge research, Graphiti, OASIS, report, translation, extraction, audit, and ensemble child-process records into one immutable lineage ledger.",
      "Distinguish current-attempt, cumulative-lineage, measured-lower-bound, and estimated totals after retries and resumes.",
      "Reconcile each stage and the complete pipeline as the sum of unique local logical call IDs and reject double-counted synthetic aggregates."
    ],
    "passes": false
  },
  {
    "id": "feature_live_budget_and_health",
    "category": "reliability",
    "description": "Users can see truthful live stage health, time, spend, and remaining budget, and cross-process token/cost budgets stop new calls without corrupting resumable work.",
    "steps": [
      "Expose elapsed stage time, approximate ETA, heartbeat age, owner health, stale state, measured or estimated tokens, cost, and budget remaining through the status API.",
      "Render a compact live health and cost panel above the pipeline timeline with explicit unknown and stale states.",
      "Apply one durable pipeline-and-attempt budget across research, graph, simulation, and report subprocesses.",
      "Race parallel provider fixtures at the remaining-budget boundary and verify atomic reservations prevent every worker from independently passing the same pre-call check.",
      "Cross a budget in an offline provider fixture and verify the next request is rejected with a typed durable receipt while completed artifacts remain resumable.",
      "Verify a provider outage or stale heartbeat stops boundedly and cannot issue a duplicate run or unbounded retry storm."
    ],
    "passes": false
  },
  {
    "id": "feature_shared_handoff_api_authority",
    "category": "cross-stage",
    "description": "Base pipelines, scenario forks, and resumes resolve every dossier, translation, PDF, editing, progress, visualization, and report API through the state-authoritative shared handoff directory.",
    "steps": [
      "Create offline base-pipeline and scenario-fork states where the fork intentionally points at the base handoff directory.",
      "Read dossier, translation, PDF, edit, progress, visualization, and report surfaces through both base and fork pipeline IDs.",
      "Verify every route resolves state.handoff_dir with a safe pipeline-local fallback only for compatible legacy state.",
      "Tamper with handoff identity, hashes, or attempt lineage and verify reuse fails closed rather than reading unrelated bytes.",
      "Verify resume and fork operations preserve the same artifact authority in both internal orchestration and external APIs."
    ],
    "passes": false
  },
  {
    "id": "feature_polymarket_provenance_history",
    "category": "forecast-quality",
    "description": "Polymarket evidence can influence a forecast only when equivalence and confidence are retained, and selected markets carry exact provenance, current price, liquidity, spread, CLOB identity, and bounded history enrichment.",
    "steps": [
      "Feed exact or near market matches at confidence 0.49 and 0.50 through anchoring, divergence, reconciliation, and publication.",
      "Verify the rejected 0.49 match causes no revision call and leaves probability and rationale byte-for-byte unchanged.",
      "Require accepted market rationales to bind the exact market ID or question and quoted price, not merely contain the word market.",
      "Align agent-tool, deterministic, report, and UI market schemas for IDs, URLs, prices, bid and ask, liquidity, volume, change, timestamps, relevance, equivalence, and confidence.",
      "Batch-enrich only selected exact market IDs through one bounded Gamma lookup and fetch CLOB history when available, recording an explicit unavailable reason otherwise.",
      "Map each CLOB token through its declared outcome label and use only the token for literal Yes, never an assumed first array position.",
      "Poll unresolved, ambiguous, and resolved markets repeatedly and verify idempotent ledger updates with only unambiguous resolution becoming calibration authority.",
      "Render research-versus-current price, freshness, equivalence, confidence, spread, history, and capture provenance in the dossier and report UI."
    ],
    "passes": false
  },
  {
    "id": "feature_forecast_visual_quality",
    "category": "visualization-ui",
    "description": "The report UI presents readable, source-bound binary, scenario, market, uncertainty, diagnostic simulation, revision, timeline, and quantitative evidence with explicit coverage limitations.",
    "steps": [
      "Render a first-class sortable binary forecast table with resolution criteria, horizon, confidence, and source provenance.",
      "Render scenario probabilities and ensemble uncertainty without silently inventing residual or extra scenarios.",
      "Show market comparisons with exact contract links, equivalence, confidence, freshness, and price history.",
      "Label world-state and other simulation trajectories as diagnostic elicited-model projections that do not update forecast probabilities.",
      "Version visualization manifests with producer policy, generation time, input hashes, source metadata, and explicit skip reasons.",
      "Replay saved reports and verify chronological axes, comparable units, readable labels, uncertainty, responsive layout, alt text or table alternatives, and no excessive inline Plotly duplication."
    ],
    "passes": false
  },
  {
    "id": "feature_graph_timeout_efficiency",
    "category": "performance-reliability",
    "description": "Graph construction avoids multi-hour timeout tails and redundant extraction while preserving a sealed, explicit degradation contract for partial graphs.",
    "steps": [
      "Benchmark current chunk size, retry limits, community-detection defaults, resolution guards, structured seeds, and executor attribution against historical fixtures.",
      "Inject repeated 900-second-equivalent timeouts and verify a short consecutive-timeout or provider circuit stops the stage boundedly.",
      "Retain successful chunks, rejected-chunk counts, timeout causes, and completeness limits in a signed degradation receipt.",
      "Use compact structured actor and relationship artifacts to avoid re-deriving already-sealed facts where equivalence is proven.",
      "Verify downstream prepare and report stages either accept the declared partial graph under policy or fail closed deterministically.",
      "Compare wall time, model calls, tokens, retained facts, and forecast-quality fixtures against the historical graph baseline."
    ],
    "passes": false
  },
  {
    "id": "feature_research_marginal_yield",
    "category": "performance-quality",
    "description": "Research continues only while a bounded pass adds independent decision-relevant evidence, closes a named gap, resolves a contradiction, or changes a forecast input without weakening the dossier judge contract.",
    "steps": [
      "Replay historical late-gap and duplicate-query logs through the current isolated-thread, compact-prior, gap-replacement, plateau, and zero-yield controls.",
      "Normalize and deduplicate exact search queries and stop repeated terminal transport failures within one run.",
      "Measure tokens, wall time, tool calls, fetch failures, independent retained sources, KIQ closures, contradiction resolutions, and forecast-input changes per pass.",
      "Stop a pass when marginal evidence yield is exhausted, while retaining explicit uncovered gaps rather than inventing completion.",
      "Verify actor, source, citation, judge, extraction, and publication contracts are no weaker than the historical accepted baseline."
    ],
    "passes": false
  },
  {
    "id": "feature_report_ensemble_efficiency",
    "category": "performance-quality",
    "description": "Report retries and optional forecast ensembles reuse invariant work and never regenerate full prose, translation, visualization, and audit artifacts when only a structured probability vector is required.",
    "steps": [
      "Split report telemetry into section writing, extraction, repair, translation, audit, visualization, and retry operations.",
      "Keep the default forecast seed count at one until a forecast-only ensemble path is verified.",
      "Generate an extra ensemble seed as a structured forecast vector without a full prose report, translation, visual render, or publication audit.",
      "Bound language-purity and section repair to the affected content instead of escalating to whole-report regeneration or unbounded sequential calls.",
      "Resume a report-only pipeline after a completed simulation and verify the sealed simulation is reused without any repeated simulation calls.",
      "Resume a partial simulation from its last sealed round and verify completed rounds and their provider calls are not replayed.",
      "Verify published report bytes and audit identities remain immutable while sidecar ensemble evidence is separately sealed.",
      "Compare calls, tokens, latency, probability output, and publication integrity against the historical full-report ensemble baseline."
    ],
    "passes": false
  },
  {
    "id": "feature_controlled_optimization_validation",
    "category": "acceptance",
    "description": "After separate paid-run authorization, one fresh controlled pipeline proves the optimized current source reduces cost and latency without degrading evidence, forecast, market, simulation, artifact, or UI quality.",
    "steps": [
      "Obtain explicit authorization for exactly one new paid run and record the prompt, model, provider, configuration, budget, and immutable baseline comparison contract.",
      "Launch exactly one pipeline and verify no duplicate POST, healthy heartbeats, provider identity, and per-request ledger activity from the first call.",
      "Capture per-stage and per-substage calls, reconciled measured tokens, explicitly labeled estimates and lower bounds, cache usage, cost, wall time, retries, errors, evidence yield, and degradation receipts.",
      "Verify all six stage artifacts, cross-stage hashes, market provenance, diagnostic simulation policy, report audit, visualization manifest, translation or PDF if requested, and localhost presentation.",
      "Compare against the historical reference runs and reject any optimization that saves resources by weakening source coverage, judge outcomes, provenance, or publication quality.",
      "Record exact results and decide the next single improvement slice from measured residual bottlenecks."
    ],
    "passes": false
  },
  {
    "id": "feature_research_engine_v3",
    "category": "performance-quality",
    "description": "Deep-research engine v3 (RESEARCH_ENGINE=v3, default): bounded agentic KIQ research with provider prompt caching, verified citations/numbers, canonical scenario frame, resumable phases and honest exit codes, accepted unchanged by the parent pipeline.",
    "steps": [
      "Run the engine offline with the test fakes (English and Chinese, quick/standard/deep) and verify rc 0, a report >= 400 chars, positional [S#] == sources.json == References, and the canonical scenario block parsed by forecast_inputs_from_report_markdown.",
      "Verify byte-identical [ENGINE_CORE][RUN BRIEF][shared context] prefixes across sibling calls and agent steps, prime() before each fan-out, one [usage] line per provider call with cached= tokens, and unchanged GLM request bodies.",
      "Fault-inject provider outage, quota, content filter, deadline, budget, empty/unparseable replies and tool outages in every phase; verify deterministic degradation with research_quality.degraded events, or a resumable exit 2 for outages/no model output/no evidence, and correct resume.",
      "Drive the real PipelineOrchestrator research stage with a child process and verify acceptance, reception of the unsealed actor plane, audit-only lint, telemetry and reuse on a second run.",
      "Owner-gated: one live GLM-5.3 run confirming cache hits and provider acceptance of the thinking/reasoning_effort/max_tokens parameters."
    ],
    "passes": true
  }
]