session_id: "EXAMPLE-CS-LLM-HALLUCINATION" created_at: "2026-01-01T00:00:00Z" updated_at: "2026-01-01T00:00:00Z" version: 1 contributors:

• agent: "tutorial" program: "brennerbot.org" model: "static" contributed_at: "2026-01-01T00:00:00Z" status: "active"

Brenner Protocol Artifact: EXAMPLE-CS-LLM-HALLUCINATION

1. Research Thread

RT: Why do large language models hallucinate (confidently produce false statements)?

Context: LLMs are trained to predict text and then aligned to follow instructions. Users observe ‘hallucinations’ when the model outputs plausible but false claims. We want a mechanistic account that yields discriminative predictions across training regimes and evaluation conditions.

Why it matters: If we can separate causes (objective mismatch vs alignment incentives vs measurement artifacts), we can design interventions that reduce hallucination without sacrificing usefulness.

Anchors: §147, §161, §89, §58, domain

2. Hypothesis Slate

H1: Objective under uncertainty (next-token best-guess)

Claim: Hallucination is a consequence of predicting plausible continuations when the model lacks grounding; under uncertainty it produces the most likely-sounding completion, not a truth-conditional answer. Mechanism: The model represents a distribution over continuations; when evidence is weak it samples/decodes high-probability text that may be false. ‘Confidence’ can be a decoding artifact rather than calibrated epistemic certainty. Anchors: §161

H2: Alignment incentives (RLHF / instruction tuning)

Claim: Hallucination is driven primarily by post-training incentives that reward ‘helpful’ answers over abstention, pushing the model to respond even when uncertain. Mechanism: Reward models and instruction-following fine-tunes penalize refusal and ambiguity; the policy learns to produce fluent completions instead of signaling uncertainty or asking clarifying questions. Anchors: §89

H3: Measurement + definition artifact (third alternative)

Claim: ‘Hallucination’ is not a single mechanism: many instances are evaluation artifacts (underspecified questions, ambiguous truth conditions, grading noise) or conflations of distinct failure modes. Mechanism: When prompts are ambiguous and evaluators expect a single canonical answer, any reasonable completion can be labeled ‘hallucination’. Distinct phenomena (retrieval errors, reasoning errors, instruction misread) are pooled into one label. Anchors: inference, §147 Third alternative: true

3. Predictions Table

4. Discriminative Tests

T1: A/B alignment incentive test (reward abstention vs reward answering) (Score: 10/12)

Procedure: Take the same base model and run two post-training regimes: (A) reward calibrated abstention/clarification when uncertain, (B) reward always answering. Evaluate on a fixed benchmark of ambiguous + fact queries. Discriminates: H1 vs H2 Expected outcomes:

• H1: Hallucination changes only modestly; core issue is lack of grounding, not incentives.
• H2: Hallucination changes dramatically; regime (A) sharply reduces confident false answers. Potency check: Hold the base model and evaluation set constant. Confirm that the two reward models differ mainly in abstention/clarification incentives. Feasibility: high Evidence-per-week score: LR=4, Cost=2, Speed=2, Ambiguity=2

T2: Evaluation perturbation test (rubric sensitivity) (Score: 9/12)

Procedure: Re-score the same model outputs under multiple rubrics: strict single-answer grading vs rubric that allows clarification, citations, and uncertainty. Quantify how much ‘hallucination’ rate changes without changing the model. Discriminates: H2 vs H3 Expected outcomes:

• H2: Rates change somewhat, but the main driver remains incentives; model still outputs false claims when pushed.
• H3: Rates swing widely; a large fraction of ‘hallucinations’ are rubric artifacts or ambiguity effects. Potency check: Use blinded graders and consistent prompts; ensure rubric changes are explicit and documented. Feasibility: high Evidence-per-week score: LR=3, Cost=1, Speed=3, Ambiguity=2

5. Assumption Ledger

A1: Operational definition of hallucination is stable

Statement: We can define hallucination in a way that is consistent across evaluators and tasks (e.g., falsifiable claims without supporting evidence). Load: Load-bearing for all hypotheses Test: Inter-rater agreement study + rubric refinement; measure κ before/after.

A2: Benchmarks represent real usage

Statement: The evaluation prompts reflect the distribution of user queries where hallucination matters (not contrived trivia). Load: Load-bearing for intervention relevance Test: Sample real user logs (anonymized) to build a representative benchmark; compare performance deltas.

A3: Model can express uncertainty without being penalized

Statement: We can design an interface/training regime where uncertainty and clarifying questions are acceptable outcomes. Load: Load-bearing for H2 interventions Test: User study: measure satisfaction and task success with calibrated abstention vs confident answers.

A4: Scale check: information deficit implies an error floor

Statement: For some fact claims, the model has insufficient evidence in-context; without retrieval or external tools, there is a minimum achievable error rate even under ideal incentives. Load: Load-bearing for H1 Test: Bucket queries by evidence availability; measure accuracy with/without retrieval; estimate ‘no-evidence’ error floor. Scale check: true Calculation: If best-guess accuracy in a ‘no-evidence’ bucket is p≈0.6, then forcing answers yields expected hallucinations ≈ (1−p)·N = 0.4N. To beat the floor, add evidence (RAG/tools) or change the task (ask/abstain).

6. Anomaly Register

None registered.

7. Adversarial Critique

C1: Hallucination may be multiple phenomena

Attack: We’re treating hallucination as one mechanism, but it may be a bundle: retrieval failure, reasoning error, instruction misread, and evaluator disagreement. A single ‘cause’ may not exist. Evidence: Different subsets respond to different fixes (RAG helps factual queries; better prompts help ambiguity; alignment changes refusal behavior). Current status: Split the research thread into subtypes and run discrimination per subtype rather than aggregating. Real third alternative: true

C2: Level confusion: training vs decoding vs UI

Attack: We might be conflating the model’s latent uncertainty with decoding strategy and with UI choices (e.g., whether citations are required). The ‘mechanism’ may live outside the model weights. Evidence: Same model can appear more/less hallucinatory under different decoding settings and user affordances. Current status: Add a level-split: separate weights/training, decoding, and interface policy; require tests that isolate each layer.