# PRD-Lite — Confidence Gate v1

**Status:** Build-ready  
**Owner:** Pooja Hegde  
**Partners:** Support Operations, Applied ML, Knowledge, Design, Data Science, Legal/Privacy

> Portfolio note: This is a simulated product requirement document. It is paired with a synthetic [experiment readout](EXPERIMENT_READOUT.md).

## Problem

AI-generated support drafts optimize response speed but present unsupported claims with the same fluent tone as verified ones. Agents must leave the reply composer to check knowledge, account state, and policy. Under queue pressure, inconsistent verification creates avoidable customer harm.

## User

**Primary:** Support agent handling AI-eligible conversations.  
**Secondary:** Quality lead accountable for policy compliance and correction rate.  
**Affected:** Customer relying on the response to make a financial, privacy, security, or account decision.

## Job to be done

> When an AI draft contains a consequential claim, help me see what supports it and what needs judgment, so I can respond safely without rebuilding the answer from scratch.

## Goal

Reduce critical policy errors by at least 30% relative while keeping median handle-time increase below 8%, escalation increase below 1.0 pp, and agent throughput decline below 5%.

## Non-goals

V1 will not:

- auto-send any response;
- expose a model confidence score to customers;
- use gate overrides as an agent performance score;
- guarantee factual correctness from confidence alone;
- cover unsupported languages, incident queues, abuse/threat, or legal-hold workflows;
- replace policy owners or approve exceptions;
- train on customer content outside the approved retention and consent policy.

## Product contract

### Inputs

- AI draft split into claims;
- retrieved evidence with source, owner, version, and timestamp;
- claim risk tier;
- calibrated confidence;
- account and queue eligibility state.

### Decision

`PASS`, `VERIFY`, or `ESCALATE` at draft and claim level.

Risk tier is evaluated before confidence:

```text
if ineligible_workflow: ESCALATE
else if high_risk AND evidence_missing_or_conflicting: ESCALATE
else if high_risk AND confidence_below_threshold: VERIFY
else if medium_risk AND evidence_missing: VERIFY
else: PASS
```

Threshold values are model-versioned and owned by Applied ML + Support Quality.

### Outputs

- supported, unsupported, and conflicting claim states;
- source excerpts beside the relevant claim;
- plain-language gate reason;
- one recommended action; and
- auditable override or escalation event.

## Requirements and acceptance criteria

### FR-01 — Claim evidence

**Requirement:** Each consequential claim shows supporting evidence or an explicit absence state.

- **Given** a retrieved source supports a claim, **when** the draft renders, **then** the claim links to source title, section, version, and updated date.
- **Given** no source clears the retrieval threshold, **when** the draft renders, **then** the UI says “No supporting source found”; it does not leave an empty citation area.
- **Given** two active sources conflict, **when** the draft renders, **then** both sources are shown and the claim cannot pass automatically.

### FR-02 — Risk-aware gate

**Requirement:** The system selects PASS, VERIFY, or ESCALATE using risk, evidence, and confidence.

- **Given** a high-risk financial claim with missing evidence, **when** the decision service runs, **then** the result is ESCALATE regardless of model confidence.
- **Given** a low-risk how-to claim with current supporting evidence and confidence above threshold, **when** the service runs, **then** the result is PASS.
- **Given** the decision service times out, **when** the composer loads, **then** sending the AI draft is disabled and a manual-response path remains available.

### FR-03 — Verification workflow

**Requirement:** VERIFY identifies the exact claim and action needed.

- The instruction names the unsupported or conflicting claim; generic “check this reply” copy does not pass acceptance.
- Opening evidence preserves draft cursor and scroll position.
- After a relevant edit, the agent can re-run the gate without regenerating the whole reply.

### FR-04 — Escalation

**Requirement:** ESCALATE routes context without requiring duplicate entry.

- **Given** an escalation, **when** the agent confirms, **then** conversation, draft, evidence state, and gate reason are attached to the destination queue.
- Customer-visible text is not sent until an authorized agent acts.
- If routing fails, the product preserves the unsent draft and gives a retryable error with reference ID.

### FR-05 — Override

**Requirement:** Authorized agents may override VERIFY, not mandatory ESCALATE cases.

- Override requires one structured reason and optional note.
- The interface states that reasons improve policy and model quality; they are not presented as performance scoring.
- Every override records model, evidence, policy, and UI versions.

### FR-06 — Auditability

**Requirement:** Quality reviewers can reconstruct the decision state.

- A completed conversation stores the draft shown, final text hash, gate result, claim states, source IDs/versions, agent action, and timestamps.
- Raw restricted customer content is not copied into the analytics warehouse.
- Audit data follows the approved retention policy and access controls.

### FR-07 — Accessibility and latency

- Claim states use text and iconography, not color alone.
- Gate actions are keyboard operable and screen-reader labeled.
- p75 gate decision latency is ≤ 500 ms after draft completion.
- Evidence loading cannot shift the send action beneath the pointer.

## Metrics

Primary, guardrail, and mechanism metrics are defined in [EVENT_SCHEMA.md](EVENT_SCHEMA.md) and the [experiment readout](EXPERIMENT_READOUT.md).

## Dependencies

| Dependency | Failure behavior |
|---|---|
| Claim extraction | Fall back to draft-level VERIFY |
| Evidence retrieval | Explicit missing-evidence state |
| Risk classifier | Use queue-level maximum risk |
| Confidence service | Do not treat missing confidence as high confidence |
| Escalation router | Preserve draft; retry with reference ID |
| Audit pipeline | Do not block send after gate decision; queue secure retry |

## Rollout

Employee sandbox → quality shadow mode → 5% → 10% → queue-specific ramps. Auto-send is not part of any stage. See [ROLLOUT_PLAN.md](ROLLOUT_PLAN.md).