> ## Documentation Index
> Fetch the complete documentation index at: https://docs.epode.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Task intent evidence

> Keep intent provenance, journey behavior, decisions, and outcomes distinct inside one task-scoped evidence graph.

The agent experience graph can explain more than a page view, but only if the evidence remains honest.
Task intent evidence is an append-only, company-owned record of one agent-mediated customer task. It preserves
what was stated or inferred, what the agent explored, what the product evaluated, and what the company observed
afterward without turning sequence into attribution.

## Evidence types stay separate

Each ledger is scoped to exactly one product-issued `journeyId` and uses four node types:

* `intent`: a goal, requirement, preference, constraint, interest, or other task context;
* `journey`: a question, path, option, disposition, or stage observed during the task;
* `decision`: the company's exact-match, near-miss, no-match, or unknown evaluation; and
* `outcome`: a conversion, completion, or other stable company outcome label observed later.

An outcome is connected with an `observed_after` edge. The graph always carries
`attribution.status: "not_established"`; there is no `caused_by` edge or causal score.

## Provenance is required

Intent facts must declare one of three sources:

* `current_user_request`;
* `user_confirmed`; or
* `agent_inference_from_current_task`.

Inference requires bounded confidence. It cannot overwrite direct user evidence. A changed value for the same
dimension must explicitly supersede the prior fact, so contradictory values never disappear through last-write
wins behavior.

The ledger's closed schema has no customer, account, cookie, prompt, or transcript field. Values are bounded,
but the company must still pass only the minimum task-relevant value. Cross-session memory and personalization
continue to use the separate permissioned customer-context flow.

`journeyId` scopes and correlates navigation; it is never authorization. A production evidence reader must
authenticate the company, resolve the journey inside that company's product boundary, and reject cross-product
or cross-tenant references before returning a snapshot. Raw prompts, personal data, payment details, and order
details do not belong in the ledger.

## Node SDK

```ts theme={null}
import {
  createTaskIntentEvidenceLedger,
  summarizeTaskIntentEvidence,
} from "@epode/node/intent-evidence";

const evidence = createTaskIntentEvidenceLedger(journeyId);

const budget = evidence.recordIntent({
  id: "intent-budget-150",
  observedAt: new Date(),
  dimension: "budget",
  kind: "constraint",
  known: true,
  value: 150,
  strength: "hard",
  provenance: "current_user_request",
});

const decision = evidence.recordDecision({
  id: "decision-search-42",
  observedAt: new Date(),
  result: "exact_match",
  optionIds: ["focus-grid-desk-lamp"],
  basedOnIntentIds: [budget.id],
});

evidence.recordOutcome({
  id: "outcome-order-42",
  observedAt: new Date(),
  outcome: "conversion",
  source: "business_system",
  observedAfterNodeIds: [decision.id],
});

const snapshot = evidence.snapshot();
const currentTask = summarizeTaskIntentEvidence(snapshot);
```

Node IDs make identical retries idempotent. Reusing an ID for different evidence fails closed. Serialized
snapshots can be validated and rehydrated; dangling, backward, cyclic, unknown-field, and false-attribution
graphs are rejected.

## Privacy-thresholded cohort analysis

Use `@epode/node/intent-evidence/cohorts` to describe how task evidence and separately observed outcomes
co-occur across many validated ledgers. The analyzer publishes explicit eligible, observed, and missing task
denominators. It never publishes journey IDs, intent values, option IDs, outcome labels, rankings,
recommendations, or causal claims.

```ts theme={null}
import { analyzeTaskIntentEvidenceCohorts } from "@epode/node/intent-evidence/cohorts";

const analysis = analyzeTaskIntentEvidenceCohorts(taskSnapshots, {
  minimumCohortSize: 10,
});
```

Analysis is deliberately limited to univariate intent-dimension, latest-decision-result, and terminal-status
cohorts. Cohort labels and counts below the threshold are omitted. When either the observed or missing side of
an outcome split is a non-zero small cell, both values are suppressed so the cohort total cannot reconstruct
the protected cell. Intent-dimension cohorts can overlap; use each cohort's task denominator and never sum
cohort totals. The default threshold is 10 and cannot be lowered below 5.

Threshold suppression reduces direct small-cell disclosure; it is not differential privacy, anonymization, or
permission to publish customer data. Treat repeated, overlapping, or externally joined analyses as a new
privacy review. Keep access tenant-bound, use stable analysis populations, and do not turn descriptive cohort
differences into attribution or automated action. A difference is only a hypothesis for further measurement,
never an effect estimate or action recommendation.

## Browser-proven reference journey

Fieldnote preserves the original journey through negotiation, decision, item detail, and human handoff. Its
purchase route records an accepted option followed by a separately observed conversion. The evidence endpoint
requires an ephemeral server capability and returns the task snapshot and current summary, never the first-party
browser cookie. The capability is a reference-product test seam; production uses normal tenant-bound company
authorization.

Run the real Chromium checks:

```bash theme={null}
make intent-evidence-e2e
```

The first check uses isolated browser contexts for two realistic customers: one exact-match journey that
crosses the agent-to-human handoff and purchases, and one hard-constraint no-match journey with no outcome. A
second check runs 14 isolated agent tasks, completes five purchases in separate human browser contexts, and
derives thresholded cohorts from the captured evidence. It verifies exact denominators, rare-label suppression,
task isolation, and the fixed non-attribution disclosure.