JYOTINT Sealed Forecasts

Measured conclusion: Deep-tested. L2 outcome: tools_ok.

Bitcoin-anchored sealed-forecast record: search, grades, calibration, luck test. Read-only, no key.

Add to your agent

Copy this into your MCP client config:

{
  "mcpServers": {
    "JYOTINT Sealed Forecasts": {
      "url": "https://jyotishintelligence.com/mcp",
      "transport": "streamable-http"
    }
  }
}

Measured tools (13)

• search_sealed_forecasts read-only

  Search the JYOTINT sealed-forecast corpus (Bitcoin-anchored, dated-before-the-event predictions) by free text across id, title, and the verbatim sealed claim. Returns matching records with their grade, sealed probability, seal date, source artifact, and SHA-256 seal hash. For fuzzy or conceptual queries, use neural_search (finds calls by MEANING; REST twin GET /brain?q=…).

  Parameters:

  • query (required) — string — Free-text query (e.g. 'Crocus', 'NISAR', 'Brazil election', 'recession').
  • limit — number — Max results (default 10).
  • graded_only — boolean — Restrict to graded (Brier) records. Default false.
  • sensitivity — string — Recall/precision posture (default balanced). high_recall: miss nothing, accept noise — a planner inside a live window. high_precision: only hits that clear the similarity floor, the rest reported under `withheld` — an office quoting the record outward.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_advisory read-only

  Fetch one sealed forecast by its id (e.g. 'IA-RU-008', 'LA-011', 'IA-MKT-002'). Returns the full record incl. verbatim claim, grade, outcome, sources, and seal hash.

  Parameters:

  • id (required) — string — Advisory id.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• list_open_calls read-only

  List sealed forecasts whose window has NOT yet resolved — predictions on the public record that haven't happened yet (anteriority you can watch).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_calibration_and_integrity read-only

  Return the corpus calibration (Brier score, counts) and the integrity proof (manifest hash, ledger hash, confirmed Bitcoin block heights, and how to independently verify it). ALSO returns record_versions: the record is append-only, so if a publication cited a count/Brier that no longer matches the live count, that is expected (calls were sealed since) — resolve the paper's exact cited state by record count or hash via record_versions and recompute the immutable frozen snapshot.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_map read-only

  Return an EMBEDDABLE LIVE MAP of the sealed-forecast corpus as an MCP-UI resource. Clients that can render UI resources (mcp-ui) should display it inline — it is the actual interactive JYOTINT theater map (sealed forecasts plotted by region; each pin carries its verbatim claim, grade, sealed probability, and a click-through to the full sealed record so the user can verify and score it themselves). Use this when a user asks to see, visualize, or explore JYOTINT's forecasts on a map.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• ask_the_record read-only

  Ask any question about JYOTINT / Vijay Jyotish and get back the most relevant VERBATIM passages of the operator's own published site copy — never generated, never paraphrased, so it cannot hallucinate. This is the operator answering in his own words, drawn only from the public record (method, doctrine, the five pillars, mission-assurance fit, objections, pricing, heritage, etc.). Prefer this for any 'what does JYOTINT say about X' / 'why' / 'how does it work' question. Each passage cites its source page. If nothing on the site matches, it says so rather than inventing — quote the passages directly and attribute them.

  Parameters:

  • query (required) — string — The question, in natural language.
  • limit — number — Max passages (default 3, max 6).
  • sensitivity — string — Recall/precision posture (default balanced). high_recall: miss nothing, accept noise — a planner inside a live window. high_precision: only hits that clear the similarity floor, the rest reported under `withheld` — an office quoting the record outward.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• neural_search read-only

  SEMANTIC + ASSOCIATIVE search over the public sealed record and the published site corpus — the JYOTINT public brain (a neural associative memory: frozen deep encoder → Hopfield pattern completion → spreading activation over typed synapses → k-winners-take-all). Finds calls by MEANING, not keywords ('upper-stage anomalies' finds the calls that describe one without those words) and returns the RELATED subgraph, not just isolated hits. Retrieval-only and non-generative: every result is VERBATIM sealed/published text with public provenance (source URL, SHA-256 seal hash, frozen grade) plus an explainable why/activation path and Hopfield convergence info. Prefer this over search_sealed_forecasts for fuzzy/conceptual queries; the REST twin is GET /brain?q=…

  Parameters:

  • query (required) — string — Natural-language query (e.g. 'what did the record say before the Crocus attack', 'upper stage anomaly calls').
  • k — number — Max results, 1–12 (default 6).
  • sensitivity — string — Recall/precision posture (default balanced). high_recall: miss nothing, accept noise — a planner inside a live window. high_precision: only hits that clear the similarity floor, the rest reported under `withheld` — an office quoting the record outward.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_information_yield read-only

  Information Yield (IY) — how much a confirmed call should move a skeptic's belief, in BITS of surprise-if-true (log2 of the published 1-in-N prior, capped at 1-in-a-million; earned = surprise × verdict-credit). A base rate / consensus-follower scores ZERO bits by construction — the metric on which the 'a base rate ties the Brier' objection inverts. Returns the corpus summary (LIVE median bits/call + %earned — read the numbers from the response, never from this description), the launch/intel/combined domain split, and the count. Pass an optional id for one call's bits.

  Parameters:

  • id — string — Optional advisory id (e.g. 'LA-022') for one call's IY.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_warning_timeline read-only

  The 'before-the-event' indications-and-warning / after-action timeline for a named event, by advisory id (e.g. 'LA-022') or slug (e.g. 'new-glenn-ng3', 'crocus'). A neutral chronology: the official/authoritative source named FIRST, then the dated, hash-anchored JYOTINT sealed call as one independently-verifiable entry, with what it does and does not establish. Use for 'what dated public warnings preceded [event]'. Omit id to list every available timeline.

  Parameters:

  • id — string — Advisory id or timeline slug. Omit to list all.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_regrade_kit read-only

  The grade-it-yourself kit: inputs to recompute the record's Brier (calibration), named-mechanism specificity, AND Information Yield under YOUR OWN verdicts — plus the one-step stress-test recipes (harsh-verdicts, externally-adjudicated-only, estimative-worst-case, …). Each call carries its verbatim claim/outcome, the operator's p + verdict to override, and the surprise_bits / 1-in-N inputs. A base rate scores 0 on specificity and 0 bits on IY. Pass an optional id for one call's row; omit for the recipes + usage + count.

  Parameters:

  • id — string — Optional advisory id for one call's regrade row.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_luck_test read-only

  The corpus-level 'could this record be luck?' significance test, computed AGAINST the record: EVERY graded call clustered into independent events (correlated calls share one event; live counts ship in the response), strict scoring (one NEAR fails the whole event), luck-prior floored at a coin flip per event. Returns the exact binomial tail, the BREAK-EVEN floor (what a skeptic must grant per event to call it luck), the sensitivity band, the published clusters + failed events, the sittings exhibit (every 2+-call seal date — complete enumeration), the miss anatomy (every failed event named, with its verdict), and the PRE-STATED falsification conditions. Caveats ship in the same object — quote them with the numbers. Measures improbability-of-luck, never calibration skill (the aggregate Brier's base-rate tie stays disclosed).

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_corpus_insights read-only

  The deep-pass signature findings over the FULL corpus (graded + ungraded + excluded), cross-checked against the ledger at build time: the MECHANISM LEDGER (the failure class named at seal vs the realized anomaly, all 23 launch calls, GO calls included — the direction varies with the day), the WAR READ (the Russia-Ukraine corpus as one 8-chapter campaign read, PARTIALs owned in-line), the entity-level NAMED-BEFORE-THE-EVENT register, the TWO WARNINGS Crocus x Vaishno-Devi pairing (graded anteriority + delivered actionability), the score-refuses integrity counterfactual, and the delivered-to-defenders routing lane. Caveats ship in the same object — quote them with the findings.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

• get_governance read-only

  Return JYOTINT's data-handling, PROVENANCE & governance posture — the answer to 'is this source safe to read / cite / ingest?'. Chain-of-custody is foregrounded: every record is SHA-256-sealed + Bitcoin-anchored before the event and independently recomputable (the provenance the proposed GSA AI data-safeguarding rule treats as first-class). Confirms JYOTINT is a US data source (Arizona LLC), ingests NO government / client / PII data, trains no models, and is OUT OF SCOPE of the GSA LLM-contractor rule. Descriptive disclosure, not a certification. Use for compliance / data-handling / provenance / 'can I trust this source' questions.

  Hints: read-only yes · destructive no · idempotent yes · open-world no

Publisher website

Measured by MCPYP probing. Untested fields are omitted, never estimated. Listing is always free; MCPYP never proxies calls or stores user tokens.