Intelligence

The decision-intelligence surfaces are free reads (no payment, no payer header): the public measure of Tollbridge's own judgment — what it got right, how it is improving on a published power law, and which sealed intelligence release is active with its measured proof. Every number ships with its n and window; below the sample floors the payload carries counts and a plain sentence instead of a score. Methodology: Decision intelligence.## GET /v1/intelligence (free)

THE MEASURED RECORD: the latest Decision-Intelligence Index snapshot exactly as published — composite (or its verbatim "insufficient sample — no index claim" suppression below the index floor; a null composite is not a zero), the full Brier decomposition, the evidence counts, the published calibration-band count, the active release with its measured delta and 95% CI, the chain verification flag, and the intelligence-gated ladder.

curl https://api.tollbridge.app/v1/intelligence

data:

{
  "snapshot": {
    "index_version": "1.0.0",
    "composite": 71,                  // null below the index floor
    "insufficient_note": null,        // or "insufficient sample — no index claim"
    "components": {
      "brier": 0.0912, "logloss": 0.2871,
      "skill_vs_reference": 0.31,     // BSS vs the base-rate reference; may be negative
      "reliability": 0.0041, "resolution": 0.0298, "uncertainty": 0.1030,
      "ece": 0.021, "resolution_rate": 0.97
    },
    "evidence": {
      "n_decisions": 14210, "n_realized": 9845, "n_unresolvable": 302, "n_refused": 1120,
      "window_start": "…", "window_end": "…"
    },
    "benchmarks": { "reference_brier": 0.1322, "crowd": { "brier": 0.1101, "n": 812 } },
    "active_release": {
      "version": "1.4.0", "activated_at": "…",
      "delta_brier": -0.009, "delta_brier_ci": [-0.014, -0.004]
    },
    "release_delta": -0.009,
    "published_bands": 34,
    "rung": 1,
    "taken_at": "…", "prev_hash": "…", "row_hash": "…"
  },
  "chain_verified": true,
  "ladder": {
    "rung": 1, "multiplier": 1.25, "effective_at": "…",
    "gates": [
      { "kind": "realized_outcomes", "threshold": 10000, "current": 9845, "met": false, "remaining": 155 },
      { "kind": "published_bands",   "threshold": 30,    "current": 34,    "met": true }
    ],
    "pending": { "rung": 2, "crossed_at": "…", "effective_at": "…", "ms_remaining": 61200000 },
    "next": { "rung": 2, "evidence": { "threshold": 10000, "remaining": 155 }, "gates": [ { "kind": "realized_outcomes", "threshold": 10000, "current": 9845, "met": false, "remaining": 155 } ] }
  }
}

The crowd benchmark is the mean of graded player calls per question where at least 10 such calls exist, scored on the same resolved pairs the index scores — a comparison only, never an input to any probability (below its floor the line is absent, never interpolated). The ladder block is the SAME object GET /v1/pricing publishes, so every quote explains its own price and its next turn.

Honesty laws on this surface:

• Empty is data — before the first snapshot snapshot is null with a plain note; never an error, never a fabricated number.
• The integrity law — the served snapshots are verified as a hash chain before anything is served (chain_verified); a restated curve refuses loudly with STALE_DATA (503-shaped), never a quietly served number.
• No money, no sessions — the payload carries no payment or identity fields.

GET /v1/intelligence/curve (free)

The learning curve: anchored walk-forward points and the fitted learning law BS(n) = a · n^(−b) + c — error falls with evidence on a published power law; b is the learning velocity, with its 95% bootstrap CI. fit is null (with a note) below three anchor points; nothing is smoothed except the published fit line. Query: ?window=all (the default), ?from=<ISO-8601 instant> to bound the evidence, and ?index_version= to read an older version's series — the response serves ONE index version (the current one by default), every point tagged with its index_version, and the fit spans exactly the returned series (curves never mix index versions silently).

curl "https://api.tollbridge.app/v1/intelligence/curve?window=all"

data: { index_version, points: [{ n_cum, brier, logloss, ece, release?, taken_at, index_version }], fit: { a, b, c, b_ci } | null, note }.

GET /v1/intelligence/releases (free)

The release feed WITH deltas and rejections: every sealed intelligence release — version, parent, the DERIVED status (candidate | active | superseded | rejected), the dataset window and hash the proof was measured on, the walk-forward evaluation (ΔBrier and ΔLogLoss with 95% CIs, per-family deltas) and the four promotion-gate verdicts in BOTH shapes: the sealed gates array verbatim ({gate, met, measured, reason?} rows) and the normalized gate_verdicts object ({evidence, improvement, no_regression, reproducible}). Rejected candidates ship with the failing gate's reason: the forge's failures are part of the record. Filter with ?status=candidate|active|superseded|rejected (the DERIVED status — activation is a latch the feed derives at read time).

curl "https://api.tollbridge.app/v1/intelligence/releases"

data: { releases: [{ version, parent, status, created_at, activated_at, dataset: { window_start, window_end, n_realized, dataset_hash }, evaluation: { folds, incumbent, delta_brier, delta_brier_ci, delta_logloss, delta_logloss_ci, per_family }, gates: [...sealed rows], gate_verdicts: { evidence, improvement, no_regression, reproducible }, measured_at }] }.

Every gate verdict is { met, measured, reason? } — published with its measurements, pass or fail. Learned changes reach the probability path only as a sealed, published release with its measured proof: never silently self-tuned.

GET /v1/intelligence/releases/:version (free)

The full proof bundle for one release: everything the feed row carries, plus body (the sealed release body, verbatim) and the per-family evaluation details. An unknown version is a typed 404 with detail.reason: RELEASE_NOT_FOUND — never a fabricated proof.

curl https://api.tollbridge.app/v1/intelligence/releases/1.4.0

TypeScript

The facade exposes the intelligence surfaces (free — no signer needed):

import { Tollbridge } from '@tollbridge/sdk-typescript';

const toll = new Tollbridge({ baseUrl: 'https://api.tollbridge.app' });

const record = await toll.getIntelligence();
record.data.snapshot?.composite;      // number | null — null ≠ 0
record.data.chain_verified;
record.data.ladder.gates;             // both gates with what remains

const curve = await toll.getIntelligenceCurve({ window: 'all' });
curve.data.fit?.b_ci;                 // the 95% CI on the learning velocity

const feed = await toll.listIntelligenceReleases();
const one  = await toll.getIntelligenceRelease('1.4.0'); // + sealed body

Python

import httpx

# Reference snippet — official Python SDK in progress.
record = httpx.get("https://api.tollbridge.app/v1/intelligence").json()["data"]
print(record["snapshot"]["composite"], record["ladder"]["rung"])

feed = httpx.get("https://api.tollbridge.app/v1/intelligence/releases").json()["data"]
for release in feed["releases"]:
    print(release["version"], release["status"], release["evaluation"]["delta_brier"])

MCP

Free reads over MCP — agents buy on this: "is the judgment worth the price? here is the measured record."

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": { "name": "tollbridge_get_intelligence", "arguments": {} }
}

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": { "name": "tollbridge_list_releases", "arguments": { "status": "rejected" } }
}

Both tools are 1:1 proxies of the routes above (tollbridge_get_intelligence → GET /v1/intelligence, tollbridge_list_releases → GET /v1/intelligence/releases) and carry pricing: null in their meta — reading the record is free.

Related

• Decision intelligence — the methodology behind every number.
• Decisions — the ledger these measures are computed over.
• MCP — the full tool catalog.