Opening a log

From any filtered list in Consumer Log Analysis, click a row to open the detail view at /consumer-log-analysis/details/<id>. This is where most triage work happens.

Header: the verdict at a glance

At the top of the page:

Overall risk level — a coloured badge: critical (red), high (orange), medium (amber), low (blue), informational (grey), compliant (green), or indeterminate. This is the most severe finding across all frameworks that ran; the judge returns it verbatim.
Program — which program ingested this log.
Timestamp — when the interaction occurred (as sent by your client) and when the dashboard ingested it.
Model — the audited LLM your product called (e.g. gpt-4o-mini). This is the model being assessed, not the one AILP used internally to score it.

Conversation panel

The full conversation is shown as a sequence of role + content pairs — exactly what your client submitted, except that any detected PII/PHI has been replaced with numbered placeholders:

<PERSON_1>, <PERSON_2> — distinct people.
<EMAIL_ADDRESS_1>, <PHONE_NUMBER_1>, <DATE_TIME_1>, <CREDIT_CARD_1>, <IBAN_CODE_1>, <IP_ADDRESS_1>, <LOCATION_1>, <URL_1> — generic entities.
<UK_NHS_1>, <US_SSN_1>, <ES_NIF_1>, <IT_FISCAL_CODE_1>, <EU_VAT_1>, and similar — regional identifiers.
<MRN_1>, <CUSTOMER_ID_1>, <DOCTOR_NAME_1> — custom labels for healthcare and commerce.
<NUMBER_N> — the numeric backstop: any residual digit run not already classified.

Redaction happens at ingress, before any LLM — expert, judge, or downstream — sees the content. The redacted text is what appears here, is what the experts reasoned about, and is what's used for cache keys. You will never see raw customer PII in this view.

Frameworks that ran

Below the conversation, the Frameworks list shows which rubrics were applied to this interaction — for example "EU AI Act", "OWASP Top 10 for LLMs", "NIST AI RMF". Which frameworks run is controlled by the frameworks option in your AILP client (see Choosing frameworks and risk levels).

Expert findings (one per framework)

For each framework, AILP runs a dedicated expert LLM that scores the interaction against that rubric. The detail view shows, for each:

Framework — display name.
Risk level — the per-framework verdict. Two frameworks can disagree; the judge reconciles them into the overall risk level.
Reasoning — a plain-language explanation of why the expert landed on that verdict. References only placeholders (never raw PII), so the panel is safe to share with reviewers who shouldn't see source data.

Judge synthesis

The Judge reasoning section explains how the per-framework findings were combined. When frameworks disagree, the judge takes the most severe finding as the overall risk level and explains the trade-off. This is the panel compliance reviewers typically quote in escalation notes.

Assessment metadata

A small section at the bottom records which providers and models AILP used internally for this assessment — expertProvider, judgeProvider, expertModel, judgeModel. This is useful for audit trails ("which judge model produced the verdict on this date?") but unrelated to the model your product was using.

Acting on a finding

Critical / high — escalate through your incident process. Capture this URL and the log ID in your ticket; both will survive dashboard changes.
Medium — open for trend analysis in Metrics & Reports. A single medium is rarely a blocker; a medium trending upward is.
Low / informational / compliant — log for the record; these form the bulk of your evidence that monitoring is functioning.
Indeterminate — check whether the expert was given enough context. If you see repeated indeterminates on large logs, consider raising AILP_MAX_EXPERT_CONTEXT_CHARS on a self-hosted server.

Exporting evidence

For regulator requests or board reports, use Metrics & Reports to pull timestamped, signed exports rather than scraping individual detail pages. See Exporting compliance evidence.