Skip to main content
Review an incident
League admins and stewards10 min read

Step 3 of 3 · Review an incident

AI Power-User Guide

When to use Racey's three AI features (Protest Summary, Race Report, Scoring Recommendation), what each one does, and where they fall short.

Before you start

  • Read the role and setup requirements in the guide before following its steps.
  • Sign in with the account and league role needed for the task.

What you’ll be able to do

Know which AI feature to reach for in each workflow · Understand the human-in-the-loop guardrails and rate limits · Spot the cases where AI output should be discarded, not edited

On this page

Racey ships three AI features for league admins and stewards. Each one is a focused tool that takes data already in your league (a protest, a set of race results, your league configuration) and produces a draft you can edit, copy, or discard. The AI does not make decisions for you. It writes the first draft so you can spend your time reviewing instead of typing.

This guide covers when to use each feature, what to expect from the output, where it shows up in the UI, and where it falls short.

Availability

All three AI features are free for every league — there is no plan gate on any of them (aiRaceReport, aiScoringRecommendation and aiProtestSummary are all on for every plan since 2026-08-27).

If a button is missing or an action fails, it is one of two operator-side reasons, not your plan: the platform-wide ai_features_enabled feature flag is off, or ANTHROPIC_API_KEY is unset on the server — in which case the route returns 503 AI_UNAVAILABLE.

Philosophy: assist, don't replace

These features are deliberately scoped to drafting and recommendation, not decision-making. They share three rules:

  1. Human in the loop. Every AI output appears in a dialog or side panel that you can copy, regenerate, edit, or dismiss. Nothing is published, applied, or sent to drivers automatically.
  2. Neutral framing. Prompts forbid the model from assigning blame in protests or speculating beyond the supplied data in race reports. Output reads like a draft from a junior assistant, not a verdict.
  3. Cited inputs only. The model only sees the structured league data we pass to it (protest description, race result rows, league configuration). It does not browse the web, read your other leagues, or have memory across runs.

If you are looking for full automation, these aren't it. If you want to cut a 20-minute task down to a 5-minute review, they will.


AI Protest Summary

A neutral, 2–3 sentence summary of an open protest, plus a short list of key points stewards should focus on. Lives in the steward ruling workspace.

When to use

  • A protest description runs to several paragraphs, references multiple drivers, or mixes timeline and opinion. You want a neutral restatement before you write your ruling.
  • You are catching up on a protest you did not file or witness, and want a fast read of what allegedly happened.
  • You are writing the "Reasoning" section of a ruling and want a starting paragraph you can edit instead of staring at an empty textarea.

If the protest is a one-line "T1 lap 1 contact, P5 hit P3" you do not need this. Use it on the messy ones.

What it does

When you click AI Summary in the "Issue Ruling" card on a protest page (/steward/protests/[protestId]), the steward UI calls POST /api/ai/protest-summary with just the protest id. The server fetches the protest, all parties (complainant, respondent, witnesses), the round and track context, and any attached evidence text, then asks Claude to summarize.

The model is instructed to:

  • Stay completely neutral — no blame, no fault assignment.
  • Use stewarding-panel language, not casual prose.
  • Identify involved parties by display name.
  • Limit the summary to 2–3 sentences.
  • Surface 2–4 key points stewards should focus on.

The response renders in a side drawer. You can:

  • Copy to reasoning — pastes the summary text straight into the ruling explanation textarea.
  • Regenerate — runs the prompt again with a fresh model call (often produces slightly different phrasing).
  • Close the drawer and ignore it.

A small "AI-generated. Always review for accuracy before using in a ruling." footer is included on the panel itself — that line is non-removable on purpose.

Limitations

  • The model only sees what's in the protest record. If the description omits a critical detail (e.g., "P3 was on cold tires after a yellow"), the summary will not include it. Garbage in, garbage out.
  • Evidence files are not parsed. If a steward attaches a video link or a screenshot, the model sees the URL or filename, not the contents. Telemetry, .ibt files, and image OCR are not part of the input.
  • Hallucination risk on edge cases. The model may occasionally invent a detail (e.g., a corner number that wasn't in the description). Always cross-check against the original protest before quoting.
  • No legal weight. This is a drafting aid. The actual ruling is yours. Copying the summary verbatim into a ruling is your call, not the platform's recommendation.
  • Not retroactive. This feature was added in April 2026; protests filed before then still work, but the feature did not exist when they were originally being adjudicated.

Example output

For a protest like:

Lap 12 at Spa, T5 (Les Combes). Complainant entered the corner inside, respondent had nose alongside on the outside. Respondent did not yield, contact made on apex, complainant spun and lost three positions. Multiple replays attached.

The model returns roughly:

{
  "summary": "On lap 12 at Spa-Francorchamps Turn 5 (Les Combes), an incident occurred between the complainant and respondent when both drivers were occupying the same corner. The complainant entered on the inside while the respondent was alongside on the outside, resulting in contact at the apex. The complainant subsequently spun and lost three positions.",
  "keyPoints": [
    "Both drivers had a reasonable claim to the corner at entry",
    "Contact occurred at the apex, not on entry or exit",
    "Multiple video replays are attached for review",
    "Complainant suffered measurable position loss (3 places)"
  ]
}

The steward UI renders the summary paragraph in the drawer and the Copy to reasoning button drops just the summary text into the ruling explanation field. Key points are shown alongside but are not auto-copied.


AI Race Report

The GRS pilot provides a saved, evidence-backed recap on the league results page: an event-specific headline, short opening, useful developments, session classifications, and championship impact. The complete-event view is the default, including ordinary single-race events. Other leagues retain their existing report experience.

Choose AI Report, select Whole event or a named session, and generate a draft. Heat, Feature, numbered races and twin duels retain their own classifications. Edit highlights lets you select supported wording and stories; Save revision retains the previous revision. Copy recap exports the saved revision for sharing. Staff with results-import permission can publish a saved revision to a configured Discord channel after review. A long recap keeps a short message and results link ahead of the full attachment.

Racey computes facts, event points and counted standings before the configured OpenAI model chooses supported wording. It ranks leadership or gap changes, different winners, reconciled all-laps wins or winner/most-laps-led contrasts, repeated podiums, event-point differences, recorded net gains, relevant class results and carefully scoped season milestones. The editor accepts supplied fact references, names and numbers only. Invalid model output uses a readable deterministic fallback. Connected story paragraphs appear before the classifications and also carry through to copy and Discord. The strongest computed developments are required, so a model cannot reduce the recap to a list of results. Repeated results are folded into a single story paragraph. Event-point advantages explain position, stage, bonus and penalty differences only where those components reconcile with the official total; multipliers remain explicit when they do not. No model is permitted to invent arithmetic or unsupported narrative.

External/imported drivers are included. Recaps work with manual CSV and non-iRacing results; the available evidence determines which stories can be told. Missing grid, incident, gap or lap records remain unknown. A net gain is not a count of passes. Laps-led totals do not establish lap order, overtakes or strategy. Incident points are not a count of incidents and do not establish blame or retirement causes.

Expected scoring sessions must have results or explicit supported cancellation. Practice and qualifying do not establish event scoring. An empty Heat is non-scoring only where the current scoring configuration makes it so. Incomplete or provisional events suppress event points and championship claims, including championship context attached to an individual-session report. Incomplete earlier scoring history also prevents championship comparisons and milestones.

Standings are computed through the selected round using current corrected results and rules, including drops, best-of rules, multipliers, manual adjustment values and engine tiebreak order. Private adjustment reasons are excluded. Heat-only recaps never attribute full-event points to the Heat. Links retain their access requirements; current standings links may include later rounds. Next-round context appears only for a recent current round and a verified future published schedule; unpublished times remain unknown.

A season milestone requires complete comparable history, stable participant identity, the same session type/order and class. “First recorded podium of this season” is not “first-ever podium.” A report becomes stale after source, identity, scoring-rule, schedule or narrative-version changes; regenerate before saving further edits or publishing. Old saved revisions remain unchanged.

Illustrative supported example (fictional fixtures): Alice wins the Heat and Bob wins the Feature. Casey finishes P3 in both. Bob's counted championship gap to Alice grows from four to eight points through Round 3. The recap can describe those results and the gap change, label the session and full-event scopes, show the scoring calculation and link to results. Those facts do not establish who passed whom or why anyone retired.

AI Scoring Recommendation

A one-shot recommendation of which scoring preset best fits your league, with a 2–3 sentence rationale and up to 3 suggested customizations. Lives on the league scoring configuration page.

When to use

  • You are setting up a new league and the eight presets (f1, nascar, indycar, imsa, oval_weekly, road_weekly, gt3, heat_feature) all look plausible and you do not know which to pick.
  • You inherited a league or are running a new format (multiclass endurance, heat + feature) and want a sanity check on whether your current preset still fits.
  • You are pitching a scoring change to your league and want a third-party rationale to share with drivers.

If you already know exactly which preset you want and how to customize it, skip this — the configuration UI is faster.

What it does

On the league scoring page (/league/[slug]/scoring), click Analyze with AI next to the preset selector. The UI calls POST /api/ai/scoring-recommendation with just the league id. The server reads the league's discipline (road, oval, dirt, endurance, multiclass), counts approved + pending registrations across all seasons, infers the race format and average race length from the most recent season's settings, and detects whether multiple car classes are in use.

That league shape is fed to Claude with the rules of each preset and instructions to recommend exactly one. The model can also suggest up to three customizations (e.g., "add 1 drop week", "increase laps-led bonus to 2 points") if the preset doesn't fit perfectly.

The recommendation renders in a dialog showing:

  • The current preset (for comparison).
  • The recommended preset, highlighted.
  • The reasoning, as the dialog description.
  • Any suggested customizations as a labeled key-value list.

You can Apply the recommendation (which loads the recommended preset into your scoring form — a confirm step protects against accidental clicks) or Dismiss. Customizations are advisory; they are not auto-applied. You implement them by editing the scoring fields after the preset loads.

If the model returns the same preset you already have with no customizations, the dialog shows "Already applied" and disables the Apply button.

Limitations

  • Heuristic, not optimization. The recommendation is based on discipline + driver count + format heuristics. It is not running a simulation against your historical results to find the points system that maximizes championship excitement.
  • Snapshot in time. It uses your current registration count and most recent season's format. If you are mid-pivot from sprint to endurance, run it again after you have updated your settings.
  • Eight presets only. The model is constrained to recommend one of the eight built-in presets. It will not invent a custom point scale from scratch — for that, build manually after applying the closest preset.
  • Customizations are descriptions, not edits. If the model says "add a clean-race bonus of 1 point", you have to go open that field and add it yourself. The dialog shows the suggestion; the form does not pre-fill.
  • No notion of your league culture. It does not know whether your drivers want depth at the back, top-heavy drama at the front, or a forgiving drop-week scheme. Use the suggestion as a starting point.

Example output

For a 28-driver multiclass road racing league running 90-minute races:

{
  "preset": "imsa",
  "reasoning": "Multiclass + endurance-length races (90 minutes) point strongly to IMSA-style scoring, which natively handles per-class results and rewards finishing position with a deep enough field for 28 drivers. The F1-style points distribution scales well across two or more classes without flattening the title fight.",
  "customizations": {
    "Add 1 drop week": "Useful for an endurance season where mechanical DNFs are more common than in sprints.",
    "Enable team scoring": "Recommended for multiclass since drivers often share a car across stints; team standings give the partnership its own championship."
  }
}

(Illustrative — generated to show shape.) The dialog renders imsa as the recommended preset, the reasoning as the description, and the two customizations as a definition list.


Rate limits

All three AI endpoints share the same rate-limit budget shape but use separate keys, so generating a race report does not eat into your protest-summary budget for the day.

EndpointLimitWindowKey prefixKeyed on
AI Protest Summary10 requests15 minutesai:protest-summary:userId (composed with IP)
AI Race Report10 requests15 minutesai:race-report:userId
AI Scoring Recommendation10 requests15 minutesai:scoring-rec:userId

When you exceed a limit you get HTTP 429 RATE_LIMITED with a "Rate limit exceeded. Try again later." toast. The window is sliding, so a request you made 15 minutes ago drops out of your budget continuously, not in fixed buckets.

Source: docs/security/rate-limits.md (rate-limit audit details — every AI surface enforces both the middleware tier limit and a per-feature handler limit).

If you are hitting these in normal use, something is off — the limits are sized to absorb "regenerate four times in a row" but not "loop the endpoint from a script." Reach out if you have a legitimate workflow that needs more.


  • Racey School — table of contents
  • Steward Guide — full protest, voting, ruling, and appeals workflow (AI Protest Summary slots into Section 7 "Issuing a Ruling")
  • League Admin Guide — covers scoring setup (Section 4) and publishing results (Section 9), the two surfaces where AI Race Report and AI Scoring Recommendation live

Reading recap sources

Story sources use links such as “Feature results” and “Points calculation”; internal reference IDs stay out of the reader panel. An explanation beside the event-points table states why counted championship gains are unavailable when earlier scoring data is incomplete. A dash is unavailable data, never a zero.

The GRS pilot links to /results/events/[roundId], a read-only page with the full event classifications, recorded laps led, class labels/positions and every driver’s event points. Drivers need an active sign-in and registration in the league; staff retain their existing results permission. Non-members and other leagues cannot read it. Drivers cannot read pending/provisional result sets or draft seasons. The page reads current corrected data without invoking AI, creating a report revision or exposing staff editing controls. A published result set with missing scoring sessions still shows provisional classifications and suppresses unsupported scoring. Results can differ from an earlier saved recap after corrections; the page and source links explain this.

Before you move on

  • Know which AI feature to reach for in each workflow
  • Understand the human-in-the-loop guardrails and rate limits
  • Spot the cases where AI output should be discarded, not edited