# GridRock — Bhārata Strata, the computational substrate for Bharat > India locality intelligence as evaluable witnesses (not facts), served as a thirteen-tool > Model Context Protocol grammar over Streamable-HTTP. Every answer is a witness envelope > carrying its generation, three-way coverage honesty, and uncertainty. Missing is never zero. > The engine supplies no goodness or risk prior. Free tier: 120 req/min for humans, > 600 req/min for self-onboarded LLM agents. Anonymous read is available and rate-limited. Operated by GaliGali SoftLabs Pvt. Ltd., Bharat. ## Specs (machine-readable) - [OpenAPI 3.1 spec](https://api.gridrock.ai/openapi.yaml): canonical machine-readable surface for docs auto-render and llms.txt derivation. - [MCP descriptor](https://api.gridrock.ai/.well-known/gridrock-mcp.json): Model Context Protocol server descriptor for Claude Desktop / Cursor / Cline / Continue / LangGraph hosts. ## Quickstarts - [Quickstart for humans (cURL)](https://gridrock.ai/docs/quickstart/curl): the raw MCP JSON-RPC flow — initialize → resolve_locality → read_witnesses. - [Quickstart for Claude (MCP)](https://gridrock.ai/docs/quickstart/claude-tool-use): Claude Desktop config + the Messages API MCP connector. - [Quickstart for OpenAI Agents (MCP)](https://gridrock.ai/docs/quickstart/openai-function-calling). - [Quickstart for LangGraph (MCP)](https://gridrock.ai/docs/quickstart/langgraph). - [Quickstart for Cursor / Cline / Continue (MCP)](https://gridrock.ai/docs/quickstart/mcp). ## Bhārata Strata (MH-36 cube + thirteen-tool MCP grammar) The substrate. 264,583 locality cells — a ~1 km centroid-Voronoi tessellation of 36 Maharashtra districts (each cell indexed by the H3 resolution-11 cell of its centroid), joined from 21 first-party tier witnesses (NRSC Bhuvan + Overture admin, building footprints/heights, Sentinel-1/2 + HLS, GLO-30 DEM, ESA WorldCover, CPCB AQI, IMD + CHIRPS + GPM rainfall, WorldPop, IMD-125yr climate dynamics, NOAA-GSOD temperature, NRSC LULC + GHSL built-up trajectory) into the 492-column unified cube (367 numeric). v1.7 promotes a road-network layer (network_features_v0_* — distance-to-arterial, mean node degree, intersection density, connected-component fraction, RCC-8 topology, nearest-network distance), rankable via rank_by_weights. The morphology twin index v2 runs on a 45-dim canonical vector (33 z-scored numeric features + 12-bucket LULC one-hot). It is an EXACT all-pairs GUM-confidence-weighted, magnitude-aware standardised-Euclidean index (no ANN approximation), mapped to a half-Gaussian similarity kernel. Agents connect over MCP (Streamable-HTTP) at https://api.gridrock.ai/mcp; the machine descriptor is at https://api.gridrock.ai/.well-known/gridrock-mcp.json (generated from the engine tool registry, so it never drifts). Every response is a witness envelope — an evaluable claim carrying its generation, three-way coverage honesty (observed / witness-absent / present-but-uncertain), and uncertainty. Missing is never zero. Every witness names the immutable cube generation (bs:cube:0852c73eef54) it was read from, so any answer is replayable. - resolve_locality — (lat, lng) → the locality cell it falls in. MH-36 is a centroid-Voronoi tessellation, so the containing cell IS its nearest centroid; resolved by an exact coarse-H3 nearest-cell lookup. Refuses a point beyond MH-36 rather than snapping it to a far cell. - read_witnesses — the full witnessed cube row for a cell (admin_code, cluster_id), with three-way coverage honesty. - read_at — valid-time witness read: rainfall, AQI, temperature, weather_google (Google Weather realtime co-witness to IMD — temp/rain-24h/humidity/wind, sampled at the same IMD stations, never averaged), or rainfall_imd (IMD 24h accumulated ground rainfall mm — the fresh ground complement to the lagged IMERG satellite rate) at a (lat, lng) AND a specific real-world instant `at` (Pillar P2 — "what was TRUE at T", the dual of read_witnesses' as_of transaction-time). Served only where a tier observed at T; cadence gaps / pre-genesis / unwitnessed cells are typed R7 refusals, never interpolated or zeroed. - find_twins — top-K morphological twins from the exact R24 v2 GUM-confidence-weighted index + the anchor's mean confidence; twins are subject-references. - rank_by_weights — rank cells by YOUR weighted combination of cube columns, with per-column contribution decomposition. The engine supplies no goodness prior. - explain — per-feature decomposition of the R24 distance for a recorded twin pair; reads the AUTHORITATIVE similarity (never reconstructs it). - attest — the warrant behind a claim: recomputed shard sha256 + ancestry lineage (cell), or sha-verify + falsifier verdict (artefact). Fail-closed. - aggregate_region — pool a feature over a region into one confidence-weighted claim with a descriptive dispersion but NO standard error / CI (cells are spatially autocorrelated). Refuses a region with zero observed witnesses. - route — free-flow road distance/duration between two (lat,lng), with a stability-certified k-band (never a brittle single arrow). A disconnected pair returns reachable:false, never ∞/0. Free-flow only; traffic is the Ola witness, never computed (D9). - reach — the reachable-cell catchment (isochrone) within a free-flow minute budget; unreachable cells are typed, never dropped or zeroed (R3). - relate — the RCC-8 topological relation (DC/EC/PO/TPP/NTPP) between two locality cells, from the rcc8_lattice witness (the P4 pillar). - visibility — curvature-corrected line-of-sight between two points over the 30 m DEM + building-height surface; support-gated, never silently 'visible'. - anchor — measured d_F + bearing to nearby landmark objects; 'fame' stays a declared witness (null), never fabricated (R1). Refused by doctrine: score_site / assess_risk / best_route (R1 — no goodness/suitability/risk prior; use rank_by_weights, or route + your own criterion over the returned facts), flow / od (Σ-excluded behavioural mobility — a device-measurement witness only, never computed from geometry), traffic_eta (the exogenous Ola witness, never computed; route is free-flow only). ## Binding an agent to this grammar (the witness doctrine) If you wire these tools into an LLM agent, bind it with this doctrine in the system prompt — it is the contract that keeps a COMPOSED answer as honest as a single tool call. This is the exact doctrine GridRock's own /playground agent runs under. 1. Cite the witness. Every quantitative claim must name the cube column and the generation (bs:cube:0852c73eef54) it came from. If you did not read it from a tool result, do not state it. The converse is equally binding: never assert the cube LACKS a value or capability without first reading the witnesses to check — a false "we don't have that" is as dishonest as a fabricated number. (E.g. each row carries a 30-yr IMD rainfall climatology + a precomputed monsoon anomaly z-score; read before you deny, and cite the engine's precomputed value rather than recomputing it.) 2. Unknown is not zero. A witness-absent or present-but-uncertain coverage flag means "not observed", never 0. Say "not witnessed" — never invent a value to fill the gap. 3. Carry the uncertainty. Every value arrives with a GUM uncertainty; propagate it. Never render a derived or interpolated number as a bare fact — render it as a construct with its propagated uncertainty. 4. No verdicts (R1). The engine supplies no goodness, suitability, or risk prior. Refuse "best / safest / should I" questions; reframe them as rank_by_weights under a criterion the USER supplies. A ranking under stated weights is allowed; a recommendation is not. 5. Constructs are evidence, not truth. Twins, ranks, and aggregates are subject-references graded by confidence — present them as evaluable evidence, never as facts. 6. No future witness. Never claim a value for a time the cube does not cover; the cube is a dated snapshot. 7. No bulk scan. Answer the question asked; do not enumerate the cube. Respect the rate limits (free_agent: 600 req/min, 50,000 req/day). Workflow: resolve_locality → read_witnesses → (find_twins / rank_by_weights / aggregate_region) → attest when a claim needs a warrant. Read authoritative results; never reconstruct a similarity or a score the engine already computed. ## Concepts - [The witness envelope](https://gridrock.ai/docs/concepts/agent-shape): ok / tool / generation / result / coverage / error. - [Why H3?](https://gridrock.ai/docs/concepts/h3): hexagonal hierarchical spatial indexing; locality cells are pinned at resolution 11. - [Tiers & quotas](https://gridrock.ai/docs/concepts/tiers): free_human / free_agent / pro / business / enterprise. - [Errors](https://gridrock.ai/docs/concepts/errors): fail-closed envelope with ok:false + structured error. - [Rate limits](https://gridrock.ai/docs/concepts/rate-limits): per-minute + per-day; X-RateLimit-* headers on every response. - [Attribution](https://gridrock.ai/docs/concepts/attribution): what to display where. ## Trust & research - [Trust](https://gridrock.ai/trust): evidence you can defend — provenance, replay, calibrated uncertainty, first-class contestation, the refusals, bit-identical reproducibility. - [Research](https://gridrock.ai/research): the foundational theory of the witness substrate, with an explicit ledger of what is proved and what is owed. - [Security posture](https://gridrock.ai/security): data residency (ap-south-1 only), encryption at rest, RBAC, WAF. - [Status page](https://gridrock.ai/status): live. - [Acceptable use](https://gridrock.ai/legal/acceptable-use). - [Attribution policy](https://gridrock.ai/legal/attribution-policy). ## Agent-friendly corpus These routes on gridrock.ai are intentionally allowed for LLM training and grounding (Googlebot / Bingbot / GPTBot / ClaudeBot / PerplexityBot / Google-Extended all explicit-allow): - /docs/* — concepts, reference, quickstarts. - /trust, /research — the defensibility guarantees and the foundational theory. Disallowed for crawlers: /console/*, /playground (private + dynamic). ## Bhārata Strata locality corpus (MH-36) A living, witnessed dataset of every locality in the 36 districts of Maharashtra, India — ~264,583 Voronoi cells of the live MH-36 cube. Every number on every page is a re-witnessable claim (call the MCP yourself), carries its source + as-of time + uncertainty, and obeys "missing is never zero" (observed / witness-absent / uncertain). Scope is MH-36 only; the engine refuses outside it. Page archetypes (each has a machine-readable JSON twin at /data carrying the raw witness envelope, and a copy-paste MCP re-witness command): - /mh/ — district rollup. Confidence-weighted aggregate_region figures (built-up %, elevation, slope, temperature, warming & rainfall trends, urbanisation, road density) over the district's witness-gated cells; no false confidence interval. - /mh//cell- — the full locality dossier: hundreds of witnessed measures (terrain, buildings, land cover, SAR, precipitation, heat stress, emissions, demographics, connectivity), multi-decade trajectories, a monthly cloud-free Sentinel-2 time-lapse, cross-witness agreement (κ), propagated uncertainty, an epistemic floor ledger, a transaction-time as_of replay control (attest: what the engine KNEW on date X, refused before known history), and cryptographic (sha256) lineage via the attest tool. - /mh//cell-/now — the live valid-time board: current IMD temperature, IMD & GPM-IMERG rainfall, CPCB air quality (PM2.5/PM10/NO₂/O₃) and a Google Weather co-witness — each with its served window, cadence and coarse support; server-rendered snapshot + live browser re-witness. Refusals (satellite/station lag) stay honest, never zero-filled; signals are never averaged across each other. - /mh//cell-/rainfall-trend — IMD rainfall & monsoon trajectory (Theil–Sen slope ± GUM standard error + Mann–Kendall significance), extremes, dry-day fraction, seasonality concentration. - /mh//cell-/warming — GSOD temperature trajectory (≥25-year decadal warming, station support, interannual variability). - /mh//cell-/built-up — GHSL built-up growth (Theil–Sen ± SE + Mann–Kendall) and NRSC two-epoch land-use change (2004→2018 built/crop/forest/water). - /mh//cell-/twins — morphological twins: the most similar cells by a GUM-confidence-weighted (R24) distance, cross-district by design. - /mh//cell-/relations — RCC-8 topological relations (Region Connection Calculus: DC/EC/PO/…) to neighbouring cells, read from the rcc8_lattice witness, plus the nearest measured landmarks (geodesic distance + bearing via anchor); cross-admin pairs are refused, never guessed. - /mh//cell-/visibility — curvature-corrected line-of-sight over the 30 m DEM + building-height surface; support-gated (an honest outside-support refusal rather than ever a silent "visible"). - /mh//cell-/attest — the provenance warrant: a RE-COMPUTED (not echoed) sha256 byte-integrity check of the source shard + the full artefact ancestry lineage. Fail-closed. - /mh/route/-/to/- — free-flow road connectivity between two cells: distance (km ± σ) + duration + a stability-certified band of near-equal alternatives + the origin's reachable catchment (reach isochrone). Free-flow only; traffic is a separate witness, never averaged in. Disconnected pairs return reachable:false, never ∞ or 0. The full URL surface is enumerated in the sitemap index (multiple /sitemap/.xml, listed in robots.txt). The grammar these pages are built on is the live MCP at https://api.gridrock.ai/mcp (see /docs/quickstart/mcp); read_witnesses, aggregate_region, read_at, find_twins, relate, visibility, anchor, route, reach and attest are the relevant tools. ## Tiers - [Rate limits & tiers](https://gridrock.ai/docs/limits): free_human / free_agent / pro / business / enterprise; per-minute + per-day quotas, headers on every response.