### 30-second answer
"Hallucination is when the model states confident falsehoods. I handle it in layers: prevent (RAG + low temperature + strict prompt), detect (faithfulness evaluator + fabrication regex), correct (one retry with missing names), and contain (never let LLM choose roster members — ML + DB only)."
Defense layers (in order)
Layer 1 PREVENT → RAG facts block, "ONLY use facts below", temperature 0.1
Layer 2 CONSTRAIN → LLM never outputs player ids; ML+DB supply the list
Layer 3 DETECT → FaithfulnessEvaluator (coverage + suspectedFabrications)
Layer 4 CORRECT → Retry prompt listing missed names (score < 0.8)
Layer 5 CONTAIN → Return team[] from DB even if summary is bad; summary is garnish
Layer 6 HUMAN → Thumbs-down on recommendations; implicit quality signal
What your evaluator catches vs misses
| Catches | Misses |
|---|
| Invented player names ("Babe Ruth" not in FACTS) | Wrong attributes ("retired", "aggressive") |
| Missing names from roster | Paraphrases that omit last names |
| Correct-looking prose with subtle factual errors |
Say this: "Faithfulness here is name coverage — a strict subset of full factual accuracy. I'd extend checks to debut/bats/throws if those were in FACTS."
Real example from your demo
First summary (loose prompt): invented "aggressive approach", "retired/active".
After structured FACTS + 0.1 temp + retry: less embellishment, measurable via faithfulness.score and suspectedFabrications.
Golden set: measure and tune your hallucination defenses
Automated layers (RAG, evaluator, retry) are hypotheses until you regression-test them. A golden set is a fixed list of prompts + expected behavior you re-run after every prompt, temperature, or model change — same idea as unit tests for LLM output.
#### Step 1 — Define what "pass" means (quality bar)
For grounded team summaries, a case passes when all of these hold:
| Check | Pass criteria |
|---|
| Coverage | faithfulness.score >= 0.8 (or your chosen threshold) |
| Fabrication | suspectedFabrications is empty |
| Constraint | Summary mentions every name in team[] |
| Containment | team[] itself came from ML+DB, not the LLM (always true in your architecture) |
Optional human spot-check: summary is readable and doesn't invent attributes ("retired", "aggressive") even when name coverage is 1.0.
#### Step 2 — Build the golden set (10–20 cases)
Mix happy path, edge cases, and known failure modes:
| # | Input prompt | What you're testing |
|---|
| 1 | Build a team like Abbott | Baseline RAG summary |
| 2 | Build a team like Aaron | Common name, many matches |
| 3 | Build a team like McGriff | Single clear seed |
| 4 | Same as #1, run 3× | Consistency at temp 0.1 |
| 5 | Build a team like Abbott with loose prompt (A/B) | Before/after FACTS block |
| 6 | Team of 5 with one hard-to-spell name | Coverage stress test |
| 7 | Generic: Why is the sky blue? | No faithfulness gate (control) |
Store as JSON or a test class — inputs are stable; outputs are scored, not exact-string matched (LLM wording varies).
Example record:
{
"prompt": "Build a team like Abbott",
"expectAction": "TEAM_GENERATE",
"minFaithfulness": 0.8,
"maxFabrications": 0
}
#### Step 3 — Run the suite and record metrics
After each change (prompt edit, temperature, retry threshold, model swap):
- Call
POST /v1/chat for each golden prompt (fixed seed player if possible). - Record per case:
faithfulness.score, suspectedFabrications, latency, retry fired (yes/no). - Compute pass rate = cases meeting all criteria / total cases.
Log to a spreadsheet or ai_metric lines — you want trends over time, not one-off eyeballing.
#### Step 4 — Inspect failures and tune the right knob
| What you see | Likely cause | Fix |
|---|
| Low coverage, no fabrications | Prompt doesn't demand all names; or "exactly 2 sentences" too tight | Add "mention every name"; relax to 2–3 sentences |
suspectedFabrications non-empty | Model inventing players | Tighten FACTS constraint; lower temp; add fabrication gate (fail/retry if non-empty) |
| Coverage 1.0 but prose invents attributes | Evaluator too narrow | Extend FACTS with debut/bats; add attribute-level checks or human review |
| Pass rate OK but retries fire 80% of the time | Threshold too strict or prompt weak | Fix prompt first; only then lower RETRY_SCORE_THRESHOLD from 0.8 |
| Pass rate low, summaries look fine | Evaluator too strict (e.g. last-name matching) | Fix metric before loosening bar |
| Pass rate drops after model change | New model ignores instructions | Re-tune temp + prompt; consider rules-first for facts |
Key principle: tune the quality bar from failure analysis, not vibes. The bar is the lowest score where outputs are good enough on the golden set at acceptable latency/cost.
#### Step 5 — Lock it in (regression gate)
- Treat prompt text like code: version it, review it, run the golden set in CI or before demo.
- Any change that drops pass rate below an agreed floor (e.g. 90%) blocks merge until fixed or the bar is consciously revised.
- Complement automation with periodic human review — catches Goodhart cases (coverage 1.0, useless prose).
#### How this connects to your code today
| Golden set check | Code hook |
|---|
| Coverage | FaithfulnessEvaluator.evaluate() → score |
| Fabrications | suspectedFabrications list |
| Auto-correct | RETRY_SCORE_THRESHOLD = 0.8 + retry block in TeamChatService |
| Prevent | FACTS/TASK prompt + chat(prompt, 0.1f) |
Interview line: "Hallucination handling isn't a one-time prompt tweak — I regression-test it. Golden set pass rate is how I know whether Layer 3–4 actually work after a change."
If they push
| Question | Answer |
|---|
| Can low temperature eliminate hallucination? | No — makes wrong answers consistent. Need RAG + verification |
| LLM-as-judge to detect hallucination? | Possible with bigger model; risk: judge hallucinates too. Prefer deterministic checks when facts are structured |
| Goodhart's law? | Model could list all names with nonsense between — coverage 1.0, useless text; need fluency + human eval too |
| Production gate? | Block response if suspectedFabrications non-empty; or return team without LLM summary |
| How do you tune the 0.8 threshold? | Sweep 0.6–1.0 on golden set; pick lowest value with ≥90% pass rate and acceptable retry rate |