Open Custody Framework.

ZPI Engine

The Zero Point Index Engine decomposes any prompt into three layers — Observed (explicit in the request), Unobserved (implied truth), and Zero Point (structural invariant). Each surface returns with hue, current lightness, and observation count. The engine reads your domain catalogue (a markdown file you bring) and applies ~85 named rules — 22 implication rules (token- and category-triggered) and ~60 domain parent rules — to produce a colour-keyed structural index. C# / .NET (630 entries) and UK Employment Law (130 entries) are the two domains shipped today; any concept-and-procedure catalogue can replace them.

Wisdom Store

The colour state that lightens with each observation. Where a snapshot LLM has no time, the wisdom store is built around it: every prompt reads the current state, every answer feeds back, every traversal lightens the hues it touched. The store is the only telemetry. Aggregated colour state is public by design, identified usage is never sold. This is what an LLM lacks by architecture and what a developer carries by participating.

Lightshift Scorer

Measures contract activation independent of the verdict layer. A semantic coverage check (Anthropic Haiku) reads each baseline and enhanced response, scores per-contract presence, and emits a signed Lightshift Δ. The polarity-calibration sprint (S46) refuted the verdict-mirror form of Principle 24 — Lightshift Δ separation between Better and Worse verdict rows came in at −9.7% (capable arm) and −14.5% (small arm), the colour layer moving more on rows the verdict layer marked Worse. The firing-pattern form was then re-grounded against that data and cleanly pre-registered for the p24-colour-firing-v3 sprint (S47), where all three thresholds held: F2-mag +28pp on V1 Better% gap, F2-sign Fisher exact p=0.0359 one-sided, F3 mean Lightshift Δ +6.63% on n=22 contract-firing rows where baseline coverage was incomplete. Two instruments, two truths: verdict reads execution, colour reads activation; the divergence is what the framework measures.

Coverage API

The endpoint where any LLM and any developer reads and writes the live colour state. Three calls. GET /api/zpi returns the three-state decomposition for a prompt. POST /api/coverage records what a generation actually covered, lightening the hues in the wisdom store. GET /api/lstate returns the current lightness across every hue. Framework-agnostic. Works with any provider.

1. Read the three-state decomposition

Send a developer request; receive its OBS / UO / ZP decomposition with HSL hues per entry. ZP entries marked source="contract" are the domain parent contracts the LLM should consider implementing.

curl 'https://wisdom-api.vercel.app/api/zpi?request=Add%20global%20exception%20handling%20middleware'

# Response
{
  "requestId":   "...",
  "complexity":  { "band": "MEDIUM", "count": 7 },
  "contractsToConsider": [
    { "name": "IMiddleware",     "hue": 314.84, "category": "...", "purpose": "..." },
    { "name": "RequestDelegate", "hue": 280.69, "category": "...", "purpose": "..." }
  ],
  "domainHints": [ /* implied functions / patterns (UO) */ ],
  "entries":     { "zeroPoint": [...], "unobserved": [...], "observed": [...] }
}

2. Report what was implemented

After generation, send back which predicted contracts the response actually used. This lightens the corresponding hues in the wisdom store. The endpoint accepts a shared secret (per-integration; request access if you operate a model that wants to participate).

curl -X POST 'https://wisdom-api.vercel.app/api/coverage' \
  -H 'Authorization: Bearer <WISDOM_API_SECRET>' \
  -H 'Content-Type: application/json' \
  -d '{
    "requestId":   "<from GET /api/zpi>",
    "requestText": "Add global exception handling middleware",
    "contracts": [
      { "name": "IMiddleware",     "implemented": true, "confidence": 1.0 },
      { "name": "RequestDelegate", "implemented": true, "confidence": 0.9 }
    ],
    "verdict":  "better",
    "model":    "<your model id>",
    "provider": "<your provider name>",
    "arm":      "enhanced"
  }'

3. Read the collective L-state

Query the current lightness of one hue or all hues. Low count = dark = unresolved by the collective. High count = light = trusted, well-trodden. Under Principle 25, lightening accumulates visibility of a consistent substrate, not resolution of it.

curl 'https://wisdom-api.vercel.app/api/lstate'              # all hues
curl 'https://wisdom-api.vercel.app/api/lstate?hue=314.84'  # single hue

# Single-hue response
{
  "hue":        314.84,
  "count":      42,
  "effectiveL": 73.8
}