CATALYST/EDGE

Catalyst Edge Research is source-linked market intelligence for AI agents. The standalone, read-only MCP server returns compact, source-linked catalyst evidence dossiers for public-company tickers. It collects independent evidence families concurrently and applies a documented deterministic score. It does not import the sibling analysis-api application or its random-weight model.

The public surface has two read-only tools: catalyst_edge_score returns the compact dossier, and catalyst_edge_claim_sources pages through every immutable source record counted by a grouped claim.

Audience and distribution status

Catalyst Edge is for analysts and builders. Builders includes retail traders and technically capable individual investors who assemble their own Claude, Codex, Cursor, or other MCP research workflows. It is not limited to institutional teams.

The product promise is bounded:

Ask your AI research agent what changed for a ticker, why it matters, what contradicts it, and show every source and missing-data warning.

The scorer is deterministic and unbacktested. Catalyst Edge does not promise alpha, investment performance, buy/sell signals, personalized advice, or execution.

Version 0.1.4 is the approved free Local Beta for GitHub, PyPI, Codex, and the MCP Registry. Claude Desktop can use the same PyPI package through a manual stdio configuration, but the one-click .mcpb is intentionally withheld until its publisher can be verified by a production-trusted signing path. The current direction also includes a Hosted Pro — $29/month paid-intent test. Research does not support building Hosted Pro now: only recent activation-linked, verified, price-aware signups count toward staged 350 compatibility-spike, 1,350 scoped-review, and 11,100 safeguarded full-build reconsideration gates. None grants automatic implementation, payment, or deployment authority. Paid launch remains separately gated in GTM_PLAN.md; the economics are documented in the Hosted Pro architecture research.

Current status and completion boundary

As of 2026-07-21, the zero-subscription local runtime has the free dependencies needed to deliver the revised product: direct SEC filings and insider records plus implemented GDELT Web NGrams discovery and opt-in issuer-feed and Bluesky capabilities. The MCP transports, schemas, provenance, typed missingness, event graph, and deterministic scorer are implemented and tested.

The first product-completion slice landed on 2026-07-15: evidence now carries event type, materiality, novelty, correction lineage, source-record support, and factual why-it-matters context. SEC 8-K item codes and insider transaction facts drive event-specific headlines, invalidation criteria, and next checks. The slice is covered by fixed real RKLB SEC metadata and was rechecked against live RKLB Form 144 and 8-K records.

The default local GDELT lifecycle schedules a bounded catch-up when persisted state is due, while periodic refresh remains outside the MCP request path, shutdown cancels cleanly, and catalyst-edge-health reports last-success age plus fresh, stale, failed, never-refreshed, or unregistered state. Every dossier and paginated claim containing GDELT-derived data includes a machine-readable The GDELT Project citation and official link even when detailed sources are suppressed. Issuer feeds, discovery rules, social aliases, and publisher-domain quality tiers now load from one strictly validated local JSON registry. Registry v2 applies reviewed per-alias kind, match mode, required/negative context, validity interval, canonical CIK, rule version, and review provenance before a GDELT candidate enters the event graph. Custom registry v1 files remain compatible through deterministic legacy-rule translation. The packaged reviewed defaults preserve the existing cohort; a custom path replaces that cohort rather than silently merging unreviewed aliases.

Grouped issuer-feed and GDELT evidence now carries an immutable claim_id, exact source_record_count, and up to 20 supporting_source_ids. Call catalyst_edge_claim_sources with the claim ID and its integer cursor to recover every source ID, accession/record ID, canonical URL, timestamp, hash, parser, and policy decision in pages of at most 20. data_quality.reason_records separately retains ordered scoped reasons for unsupported or unavailable sources, no observations, rejected entities, discovery-only evidence, and evaluated non-material candidates. Display precedence does not discard coexisting reasons, and truncation is explicit through reason_record_count and reason_records_truncated. GDELT rejection disposition is source-scoped and aggregate in the dossier; individual candidate decisions remain in the bounded append-only entity audit.

The zero-subscription runtime meets its current local acceptance corpus as of 2026-07-16. A 25-case real SEC evaluation complements the 28 synthetic Phase 6 contract scenarios. It exposed and closed event-priority, merger-delisting, and recorded Item 8.01 specificity defects; all recorded classification, provenance, freshness, distinct-event, research-value, and dossier-direction checks pass.

The SEC-backed fund lane is implemented for SPY, QQQ, DIA, IWM, XLE, XLK, GLD, and GDX. QQQ, IWM, XLE, XLK, and GDX resolve reviewed SEC registrant CIK plus series/class IDs and produce neutral N-CEN/NPORT context with separate report, period-end, filing, and acceptance chronology. SPY and DIA have official CIKs but no SEC mutual-fund series/class mapping; GLD is outside the N-CEN/NPORT investment-company lane. Those three cases return explicit source_unsupported reasons instead of invented identifiers. Reviewed sponsor-primary URLs are retained as metadata only, and every reviewed fund bypasses corporate filing and insider semantics.

Item 8.01 primary-document enrichment is intentionally bounded rather than a general SEC semantic parser. The versioned sec-primary-document-v1 rules cover explicit completed debt offerings, entered or amended equity distribution agreements, actual share-repurchase activity, and filed prospectus supplements. They record the selected rule and version, reject proposed or negated near matches, preserve amendment semantics, and leave multiple specific events generic rather than choosing an arbitrary first match. Representative HTML, table, and inline-XBRL fixtures exercise those boundaries. Unsupported wording or filing structures remain other_material_event and require human review.

No numeric scorer change was justified because the real corpus contains no forward-return labels. The scorer remains deterministic and explicitly unbacktested. Paid options flow, licensed OHLC, sentiment, hosted deployment, and broader SEC semantic extraction are future capabilities. Version 0.1.4 ships only the local, self-serve surface.

In older implementation notes and runtime messages, “production” means the real non-fixture composition path used by the local MCP. It does not imply a hosted service, third-party provider role, paid account, or consumer rollout.

Install

Python 3.10+ and uv are required. Register the pinned PyPI package with Codex:

codex mcp add catalyst-edge \
  --env 'CATALYST_EDGE_SEC_USER_AGENT=YOUR_ORGANIZATION YOUR_EMAIL' \
  --env 'CATALYST_EDGE_EVIDENCE_STORE=/absolute/local/path/evidence.sqlite3' \
  -- uvx --from 'catalyst-edge-mcp==0.1.4' catalyst-edge-mcp

Open a fresh task and confirm discovery of exactly catalyst_edge_score and catalyst_edge_claim_sources. The complete Codex, Claude Desktop, privacy, and rollback procedure is in docs/demo/customer-installation-runbook.md.

Build and verify from source

Maintainers can verify the source checkout with the locked toolchain:

uv sync --frozen --extra dev
uv run --frozen pytest
uv run --frozen ruff check .
uv build --no-sources --out-dir dist

All default tests are offline. Provider tests use sanitized fixtures and httpx.MockTransport; live credentials are not required. Direct SEC event, ownership, Form 144, N-CEN, and NPORT parsing plus issuer RSS/Atom collection, event-graph behavior, and GDELT Web NGrams discovery are implemented with fixed fixtures. Official-host Bluesky partial-attention collection is also fixture-covered. Phase 5 sentiment/options gates and 28 dated Phase 6 synthetic contract cases are also executable offline.

Pull requests and release tags run the required Python 3.10/3.14 offline matrix, stdio/loopback contract suite, clean build, and installed-artifact verifier in .github/workflows/validation.yml. The workflow has read-only repository permissions and contains no package publish, release creation, registry, credential, or artifact-upload step.

The same workflow validates the MCPB release candidate with pinned Node/npm and @anthropic-ai/mcpb versions, a locked dependency integrity, zero known npm audit findings, a fail-closed file inventory, and deterministic ZIP metadata. Maintainers can run the identical packaging lane from a source checkout:

npm ci --ignore-scripts
npm audit --audit-level=low
npm run mcpb:validate
npm run mcpb:pack
uv run --frozen python scripts/verify_mcpb.py \
  --bundle dist/catalyst-edge-mcp-0.1.4.mcpb

This produces an unsigned compatibility artifact for maintainer QA. It is not distributed as a production-trusted Claude one-click package.

The live Web NGrams replacement evidence is recorded in docs/validation/gdelt-web-ngrams-live-2026-07-14.md.

Source and provider configuration

Provider credentials are read only from this process environment. They are never loaded from ../analysis-api, logged, or returned in raw signals.

Copy the checked-in template, populate the SEC identity, then export it into the current shell. The template enables attributed GDELT discovery and leaves issuer feeds and Bluesky disabled:

cp .env.example .env
# Edit .env without committing it, then:
set -a
source .env
set +a
uv run catalyst-edge-smoke NVDA --lookback-days 14

The smoke command performs a secret-free configuration preflight before making any provider request. Exit status 2 means configuration is missing or invalid; exit status 1 means providers ran but live evidence did not satisfy readiness; exit status 0 means the live provenance gate passed.

Runtime settings:

The packaged registry currently contains Apple Newsroom and NVIDIA press-release feeds on issuer-controlled hosts, plus the existing reviewed discovery/social cohort. Tesla and Rocket Lab remain unregistered for issuer feeds because no official RSS/Atom endpoint was confirmed. Copy catalyst_edge_mcp/data/reviewed_registries.json, edit the copy, and set CATALYST_EDGE_REGISTRY_PATH to use a local cohort. The loader rejects unknown fields, duplicate ticker ownership, case-insensitive cross-issuer alias collisions, noncanonical tickers, malformed CIKs, non-HTTPS feed URLs, and feed hosts absent from the exact official-host allowlist. The adapter retains titles, timestamps, identifiers, hashes, and links—not publisher bodies. Conditional ETag and Last-Modified state is refreshed at ten-minute intervals. Exact fingerprints run before same-issuer, 48-hour RapidFuzz matching at a token-set score of 92; corrections and materially changed numbers become linked event versions. The local store also snapshots the reviewed source-policy decisions used by collectors.

GDELT uses reviewed company entity rules and the official downloadable Web NGrams feed at storage.googleapis.com/data.gdeltproject.org/gdeltv5/weblegacy/ngrams. The legacy DOC 2.0 endpoint now directs high-traffic callers to these files, so local MCP requests remain cache-only while catalyst-edge-refresh-gdelt scans the newest bounded minute index/TOC pairs out of band. Each pair is downloaded once and matched against all requested issuers. A deterministic ruleset-versioned decision accepts or rejects each valid TOC candidate before ingestion. Accepted body-context matches must also name a reviewed issuer alias or non-single-letter ticker in the surfaced publisher title; otherwise they are audited as title_not_aligned. The cache-read path applies the same check so legacy tangential titles stop surfacing without deleting audit history. Exact HTTPS host/path validation, compressed and decompressed byte ceilings, a five-file run limit, a 200-candidate per-issuer audit cap, and a separate 50-accepted-document ingestion cap bound the work without allowing false positives to starve later valid matches. Only publisher titles, timestamps, domains, hashes, HTTPS links, and derived entity-decision metadata are retained; article bodies and ngram context are never stored. Accepted and rejected decisions are appended idempotently to entity_match_audit; a ruleset or context change creates a new audit record rather than overwriting history. HTTP, timeout, malformed schema, and missing-file states remain typed and preserve cached evidence. A reject-only refresh is a fresh successful collection with no observations, not a provider failure. Discovery observations merge into the same 48-hour canonical graph but remain below SEC and issuer-primary sources in global ranking. Previously cached v1 observations are not destructively purged and age out under the configured lookback.

GDELT source quality is also registry-driven. Reviewed wire-service domains receive 0.70, reviewed financial-press domains 0.68, reviewed release-distribution domains 0.62, and every unlisted domain the conservative 0.60 fallback. Matching uses an exact hostname or dot-delimited subdomain boundary, so lookalike suffixes do not inherit a tier. These remain discovery-quality heuristics to be checked during the real-case evaluation; none changes GDELT's neutral direction or readiness credit.

When the MCP server starts, its lifespan coordinator reads persisted collector state. Never-refreshed or due issuers receive one immediate bounded catch-up; a recent prior attempt delays until the configured interval, preventing restart storms. Refreshes then run serially at the configured cadence. The coordinator never runs inside catalyst_edge_score, and shutdown cancels its task before closing the SQLite handle. catalyst-edge-health is read-only and exits nonzero unless every requested reviewed ticker is fresh.

Bluesky searches reviewed exact cashtags and company aliases through public.api.bsky.app, falling back only to api.bsky.app. Both are documented, unauthenticated AppView hosts; no proxy, scraper, account, or credential path exists. The six-hour out-of-band lifecycle queries one bounded first page for the previous completed UTC day and never follows a cursor. Each daily bucket deduplicates AT URIs, counts unique authors, and retains counts, pseudonymous SHA-256 identifiers, response hash, coverage/outage state, and at most three representative links—not post bodies. Buckets auto-prune to the 14 completed days needed for two equal seven-day windows; catalyst-edge-purge-bluesky-cache TICKER... deletes the local buckets and collector state on request. Missing days, reported hit overflow, upstream failures, stale state, or disappeared URI hashes fail closed. Until 14 consecutive adequate forward buckets exist, output explicitly reports warm_up. Even after warm-up, each seven-day window must contain at least five posts from three unique authors. Search is ranked and incomplete, so output is always labeled partial public attention, remains neutral, and never claims sentiment or market-wide coverage. MCP requests read cache only. Mastodon remains uncomposed because no reviewed instance set was available to establish representative cross-instance coverage; no instance is treated as a global index.

Phase 5 reviewed Finnhub sentiment, TextBlob, VADER, DistilBERT SST-2, and ProsusAI FinBERT. None passes every rights, input-data, compatibility, preprocessing, labeled-quality, rounding, and threshold gate, so no sentiment adapter is composed. FlowAlgo and CheddarFlow also remain uncomposed: public terms do not grant the required automated extraction, storage, redistribution, and derived-output rights. The local composition root therefore never calls an options provider before policy evaluation. See docs/audits/phase5-capability-gates-2026-07-13.md.

Phase 6 validates 28 dated, sanitized synthetic product-contract cases covering strong and weak bullish evidence, bearish material events, neutral issuer/discovery items, insider clusters, Form 144 intent, missing/stale/provider failures, Bluesky historical warm-up/outage/sample regressions, contradictions, rejected sentiment, and unlicensed options/technical missingness. All 28 expected-versus-produced assertions pass. A separate 25-case real SEC product evaluation now covers the primary-source gate. See docs/validation/phase6-historical-validation-2026-07-13.md.

The real-case evaluation, semantic corrections, offline suite, live target cohort, fresh GDELT health, and final RKLB launch_ready=true smoke are recorded in docs/validation/real-catalyst-evaluation-2026-07-15.md and docs/validation/local-product-completion-2026-07-15.md.

The live evidence-semantic launch gate passed for RKLB on 2026-07-14 from merged main; the other four acceptance tickers correctly remained fail-closed. See docs/validation/live-launch-acceptance-2026-07-14.md.

Run

Stdio:

uv run catalyst-edge-mcp

Local stateless streamable HTTP:

CATALYST_EDGE_TRANSPORT=streamable-http \
CATALYST_EDGE_PORT=8000 \
uv run catalyst-edge-mcp

Direct invocation uses the same real, non-fixture composition root:

uv run catalyst-edge-score NVDA --lookback-days 14 --risk-mode research

The opt-in live readiness check prints provider names, coverage, provenance status, and sanitized warnings—never response payloads or credentials:

uv run catalyst-edge-smoke NVDA --lookback-days 14

Readiness requires SEC provenance plus a fresh directional direct-insider observation, allowlisted material event, or explicitly authorized provider observation. Provider names and keys alone do not qualify. yfinance, GDELT discovery, attention-only social data, and missing/stale data receive no readiness credit. A schema-valid no-data response demonstrates graceful failure only.

Tool input

{
  "ticker": "NVDA",
  "lookback_days": 14,
  "include_sources": true,
  "include_raw_signals": false,
  "risk_mode": "research"
}

Ticker validation runs before service composition. Unknown fields, unsafe strings, non-boolean flags, and lookbacks outside 1–90 are rejected. risk_mode accepts research, alert_triage, and thesis_review.

Response example

The following shortened multi-provider example includes conditional vendor items that assume explicit deployed policy approval. Production responses can contain up to three items per family and 15 total.

{
  "ticker": "NVDA",
  "as_of": "2026-07-12T16:00:00Z",
  "lookback_days": 14,
  "edge": {
    "score": 62,
    "direction": "bullish",
    "confidence": 0.69,
    "horizon_days": 5,
    "scoring_method": "deterministic_v1",
    "model_status": "not_trained"
  },
  "summary": {
    "headline": "Insider activity has the largest recent bullish contribution.",
    "what_changed": [
      "Net reported open-market insider value was 1500000 across 3 distinct insiders versus 100000.",
      "A new SEC 8-K was observed with item codes 2.02,9.01.",
      "Social mentions changed from 920 to 1410; sentiment changed by +0.180."
    ],
    "why_it_matters": "Source independence and agreement can strengthen confidence; missing families and lower-quality observations constrain it.",
    "what_would_invalidate": [
      "A primary filing or company record contradicts the normalized event.",
      "The reported insider cluster reverses in a subsequent equal window.",
      "Options activity normalizes during the requested horizon.",
      "A sector-wide event fully explains the ticker-specific observations."
    ]
  },
  "evidence": [
    {
      "family": "filings_news",
      "signal": "sec_form_8_k",
      "direction": "neutral",
      "strength": 0.8,
      "confidence": 0.98,
      "timestamp": "2026-07-11T00:00:00Z",
      "source_quality": 1.0,
      "change": {
        "description": "A new SEC 8-K was observed with item codes 2.02,9.01.",
        "current_value": null,
        "baseline_value": null,
        "delta": null,
        "unit": null,
        "comparison_window": null
      },
      "sources": [
        {
          "name": "SEC EDGAR",
          "url": "https://www.sec.gov/Archives/edgar/data/1045810/000104581026000001/nvda-8k.htm",
          "observed_at": "2026-07-11T00:00:00Z"
        }
      ],
      "notes": "SEC 8-K filing reports item codes 2.02,9.01.",
      "contribution": 0.0,
      "raw_signal": null,
      "source_count": 1
    },
    {
      "family": "insider_trading",
      "signal": "insider_purchase_strong_cluster",
      "direction": "bullish",
      "strength": 0.9,
      "confidence": 0.82,
      "timestamp": "2026-07-10T00:00:00Z",
      "source_quality": 1.0,
      "change": {
        "description": "Net reported open-market insider value was 1500000 across 3 distinct insiders versus 100000.",
        "current_value": 1500000.0,
        "baseline_value": 100000.0,
        "delta": 1400000.0,
        "unit": "USD reported value",
        "comparison_window": "current half vs preceding equal window"
      },
      "sources": [{"name": "SEC EDGAR", "source_id": "sec", "source_tier": "primary_regulator", "url": "https://www.sec.gov/Archives/edgar/data/1045810/000104581026000101/form4.xml", "canonical_url": "https://www.sec.gov/Archives/edgar/data/1045810/000104581026000101/form4.xml", "accession_or_record_id": "0001045810-26-000101", "published_at": "2026-07-10T00:00:00Z", "observed_at": "2026-07-10T00:00:00Z", "retrieved_at": "2026-07-12T16:00:00Z", "raw_sha256": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa", "parser_version": "sec-ownership-v1", "policy_decision": "approved", "model_or_lexicon_revision": null, "related_sources": []}],
      "notes": "Only non-derivative open-market P/S transactions with disclosed shares and price are directional.",
      "contribution": 7.21,
      "raw_signal": null,
      "source_count": 1
    },
    {
      "family": "options_flow",
      "signal": "call_flow_increase",
      "direction": "bullish",
      "strength": 0.7,
      "confidence": 0.78,
      "timestamp": "2026-07-12T14:00:00Z",
      "source_quality": 0.85,
      "change": {
        "description": "Directional options premium changed from 400000 to 1250000; 6 sweep/block events were observed.",
        "current_value": 1250000.0,
        "baseline_value": 400000.0,
        "delta": 850000.0,
        "unit": "USD directional premium",
        "comparison_window": "current half vs preceding equal window"
      },
      "sources": [{"name": "flowalgo", "url": null, "observed_at": "2026-07-12T14:00:00Z"}],
      "notes": null,
      "contribution": 4.55,
      "raw_signal": null,
      "source_count": 1
    },
    {
      "family": "technical",
      "signal": "ema_12_crossed_above_26",
      "direction": "bullish",
      "strength": 0.65,
      "confidence": 0.75,
      "timestamp": "2026-07-12T00:00:00Z",
      "source_quality": 0.85,
      "change": {
        "description": "EMA spread transitioned from -0.20 to 0.35.",
        "current_value": 0.35,
        "baseline_value": -0.20,
        "delta": 0.55,
        "unit": "EMA spread",
        "comparison_window": "latest daily value vs prior daily value"
      },
      "sources": [{"name": "FMP technical indicator", "url": null, "observed_at": "2026-07-12T00:00:00Z"}],
      "notes": null,
      "contribution": 2.38,
      "raw_signal": null,
      "source_count": 1
    },
    {
      "family": "social",
      "signal": "social_sentiment_increase",
      "direction": "bullish",
      "strength": 0.55,
      "confidence": 0.72,
      "timestamp": "2026-07-12T12:00:00Z",
      "source_quality": 0.80,
      "change": {
        "description": "Social mentions changed from 920 to 1410; sentiment changed by +0.180.",
        "current_value": 1410.0,
        "baseline_value": 920.0,
        "delta": 490.0,
        "unit": "mentions",
        "comparison_window": "current half vs preceding equal window"
      },
      "sources": [{"name": "Finnhub social sentiment", "url": null, "observed_at": "2026-07-12T12:00:00Z"}],
      "notes": null,
      "contribution": 1.27,
      "raw_signal": null,
      "source_count": 1
    }
  ],
  "attributions": [],
  "data_quality": {
    "coverage": "complete",
    "missing_families": [],
    "stale_families": [],
    "warnings": [
      "Deterministic v1 scoring is not backtested.",
      "This dossier does not provide an investment recommendation."
    ]
  },
  "next_checks": [
    "Verify the closest primary filing or company announcement.",
    "Check whether a sector-wide event explains the observation.",
    "Review timestamps, source independence, and comparison baselines."
  ]
}

Private diagnostic and no-data behavior

yfinance may be selected explicitly for development/private diagnostics, but its upstream data rights are not commercially cleared by the library. The default local composition root does not call it. Isolated diagnostic code and fixtures grant no evidence, canonical coverage, or readiness credit. It is never described as smart-money or trade flow. Install diagnostic dependencies explicitly with uv sync --extra diagnostic.

Private-yfinance diagnostic response example:

{
  "ticker": "NVDA",
  "as_of": "2026-07-12T16:00:00Z",
  "lookback_days": 14,
  "edge": {
    "score": 50,
    "direction": "neutral",
    "confidence": 0.0,
    "horizon_days": 5,
    "scoring_method": "deterministic_v1",
    "model_status": "not_trained"
  },
  "summary": {
    "headline": "No fresh canonical catalyst evidence was available.",
    "what_changed": [],
    "why_it_matters": "Private diagnostic observations are excluded from production scoring.",
    "what_would_invalidate": [
      "A transaction-level options source contradicts the snapshot activity."
    ]
  },
  "evidence": [],
  "attributions": [],
  "data_quality": {
    "coverage": "none",
    "missing_families": [
      "filings_news",
      "insider_trading",
      "options_flow",
      "social",
      "technical"
    ],
    "stale_families": [],
    "family_statuses": [
      {
        "family": "options_flow",
        "available": false,
        "status": "licensed_feed_required",
        "reason": "licensed_transaction_feed_required",
        "observed_at": null,
        "coverage_ratio": 0.0
      }
    ],
    "warnings": [
      "options_flow provider yfinance is private diagnostic only; no production evidence or coverage credit was granted.",
      "Overall confidence is below 0.50.",
      "Deterministic v1 scoring is not backtested.",
      "This dossier does not provide an investment recommendation."
    ]
  },
  "next_checks": [
    "Verify the closest primary filing or company announcement.",
    "Check whether a sector-wide event explains the observation.",
    "Review timestamps, source independence, and comparison baselines."
  ]
}

If every canonical family is absent, the response remains schema-valid with score 50, neutral direction, confidence 0, coverage="none", and all five canonical families in missing_families. Provider failures, timeouts, stale evidence, missing baselines, and low confidence are warnings; missingness never becomes bearish evidence.

Full canonical no-data response example (with every evidence source absent or disabled):

{
  "ticker": "NVDA",
  "as_of": "2026-07-12T16:00:00Z",
  "lookback_days": 14,
  "edge": {
    "score": 50,
    "direction": "neutral",
    "confidence": 0.0,
    "horizon_days": 5,
    "scoring_method": "deterministic_v1",
    "model_status": "not_trained"
  },
  "summary": {
    "headline": "No fresh canonical catalyst evidence was available.",
    "what_changed": [],
    "why_it_matters": "Missing evidence is uncertainty, not bearish evidence.",
    "what_would_invalidate": [
      "Fresh source-linked evidence from a canonical family becomes available."
    ]
  },
  "evidence": [],
  "attributions": [],
  "data_quality": {
    "coverage": "none",
    "missing_families": [
      "filings_news",
      "insider_trading",
      "options_flow",
      "social",
      "technical"
    ],
    "stale_families": [],
    "warnings": [
      "No live evidence adapters are configured.",
      "Overall confidence is below 0.50.",
      "Deterministic v1 scoring is not backtested.",
      "This dossier does not provide an investment recommendation."
    ]
  },
  "next_checks": [
    "Verify the closest primary filing or company announcement.",
    "Check whether a sector-wide event explains the observation.",
    "Review timestamps, source independence, and comparison baselines."
  ]
}

Scoring and data limitations

deterministic_v1 weights normalized direction, strength, confidence, documented source quality, and recency within fixed family budgets. Missing coverage reduces directional conviction toward neutral. The score is not trained, calibrated, or backtested and makes no expected-return claim. Every response exposes model_status="not_trained", states the backtesting limitation, and states that no investment recommendation is provided.

Raw signals are excluded by default. When explicitly requested, a recursive redactor removes credential and identity-like fields, bounds containers and strings, and caps each serialized raw item at 8 KiB.

See TDD.md for exact formulas, provider contracts, compactness rules, and the Flask migration design that intentionally remains outside this repository.