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

# AI for Evidence & Standards

> Audit-time AI assists - evidence pack narrative, pre-audit Q&A, completeness gate, standards coverage delta, plain-English clause summaries, posture weekly digest.

# AI for evidence & standards

When the work shifts from "the test failed" to "the auditor needs a story", Roboticks layers six AI assists across evidence packs, standards, and the posture dashboard. The goal is the same as everywhere else in the product: keep the engineer in control, drop the manual prose tax.

<h2 id="narrative">
  Evidence-pack narrative
</h2>

Surface: **Evidence pack detail → "Generate narrative"** action.

For any sealed evidence pack, Opus synthesises an executive-summary / audit narrative the certification auditor reads first. It pulls together:

* The pack's requirement coverage matrix.
* The standards subscriptions in effect at seal time.
* The hash-chain head + verification status.
* The release tag, branch, and commit lineage.
* The set of failing or skipped tests and their excuses (deferred, accepted risk, etc.).

The output is a 2-3 page Markdown narrative the engineer can paste into the audit cover letter or hand back into the [pack customisation](/evidence/customization) workflow. The narrative is stored on the pack so re-generating costs nothing.

Task type: `EVIDENCE_PACK_NARRATIVE` — Opus, 60k input budget, 24 token-cost.

<h2 id="qa">
  Pre-audit Q\&A
</h2>

Surface: **Evidence pack detail → Ask box.**

Free-form questions grounded in *this specific pack's* contents. Useful before a customer call ("what's our worst uncovered clause in IEC 61508 right now?") or to prep an audit ("how did we close the gap on REQ-014 between v2.3.0 and v2.4.0?").

Each turn is a Sonnet call with the pack manifest in the context. The conversation persists per pack.

Task type: `EVIDENCE_PACK_QA` — Sonnet, 40k input budget, 10 token-cost per turn.

<h2 id="completeness">
  Pack completeness gate
</h2>

Surface: **Evidence pack detail → Completeness card.**

Given a sealed pack + a target standard the org subscribes to, Sonnet calls out which clauses of the standard are *not* yet satisfied by the pack. Output is a table of `(clause, gap, suggested action)`. Used in three ways:

1. As a pre-seal advisory: run completeness against the standard, fix gaps, then seal.
2. As a regression check: re-run on an already-sealed pack to know whether the next release needs to close the gap.
3. As an audit-prep pass: hand the auditor the report alongside the pack and beat them to the obvious questions.

Task type: `EVIDENCE_PACK_COMPLETENESS` — Sonnet, 50k input budget, 12 token-cost.

<h2 id="coverage-delta">
  Standards coverage delta
</h2>

Surface: **Standard detail → Coverage AI** tab.

Across all of the project's requirements, which clauses of this standard currently have zero linked requirements? The prompt embeds a slice of the standard's clauses (typically 30-80 depending on size) and asks Sonnet to enumerate the gaps with one-line interpretation per clause.

Output renders as a coverage delta table with a "draft requirements for these clauses" action that hands off to the AI requirement generator.

Task type: `STANDARDS_COVERAGE_DELTA` — Sonnet, 20k input budget, 8 token-cost.

<h2 id="clause-summary">
  Plain-English clause summary
</h2>

Surface: **Standards detail → individual clause →** hovers, expanded cards.

Standards prose is dense by design. The clause summary is a Haiku-tier rephrase ("In plain English: …") that sits inline next to the official text. It is **not** legal advice — see the [standards disclaimer](/standards/disclaimer) — but it makes triage of "which clauses do I even need to think about" possible at the engineer level.

Task type: `STANDARDS_CLAUSE_SUMMARY` — Haiku, 2 token-cost. Cached per clause version so a project full of users opening the same clause pays once.

<h2 id="posture-digest">
  Posture weekly digest
</h2>

Surface: **Posture dashboard → top banner.**

A Sonnet call run weekly that summarises:

* This week's coverage movement (requirements newly covered or newly uncovered).
* Gap-trend slope across subscribed standards.
* Test flakiness movement.
* Significant evidence-pack events (new seals, archival).

Customers can ask a follow-up via the banner's chat link, which spawns a regular conversation against the digest.

Task type: `POSTURE_WEEKLY_DIGEST` — Sonnet, 15k input budget, 8 token-cost. Scheduled by the platform — there is no user-facing trigger button beyond "regenerate now".

## What the AI here does NOT do

* It does not sign anything. Every narrative, completeness report, or coverage delta is a draft; the engineer reviews and approves before the pack seals.
* It does not edit standards prose. The summaries are always shown next to the official text, never replacing it.
* It does not see content beyond the pack manifest + the standard's clause catalogue + the conversation thread. Source code, MCAPs, and per-test logs are not in the AI context.

## Plan gating

All six surfaces require non-zero `ai_tokens_balance`. Practical reality:

* **Free** plan's 100 000 grant covers a handful of clause summaries and a single pack completeness check.
* **Team** (5 000 000 grant) covers the typical weekly cadence: one narrative, several QA conversations, completeness on every seal, the weekly digest.
* **Enterprise** (∞) has no rate cap beyond Bedrock-side limits.

## Next

<CardGroup cols={2}>
  <Card title="Evidence packs (no-AI)" icon="file-shield" href="/evidence/overview">
    The shape of the artifact AI works on top of.
  </Card>

  <Card title="Standards coverage" icon="scale-balanced" href="/standards/overview">
    Subscribing to standards, pinning versions.
  </Card>

  <Card title="Requirements & traceability AI" icon="file-lines" href="/ai/requirements-and-traceability">
    The upstream surfaces that fill the evidence pack.
  </Card>
</CardGroup>