Epistemic Blinding: An Inference-Time Protocol for Auditing Prior Contamination in LLM-Assisted Analysis

Apr 7, 2026•Michael Cuccarese•View PDF

TL;DR Highlight

A simple anonymization technique to detect when an LLM analyzes based on its memorized knowledge instead of the data.

Who Should Read

ML engineers or data scientists utilizing LLMs in data analysis pipelines. Specifically, AI agent developers concerned about the potential for LLMs to distort results by relying on pre-trained knowledge rather than the provided data.

Core Mechanics

LLMs subtly fill answers with memorized knowledge (parametric knowledge) the moment they see 'famous entities' such as gene names or stock codes, rather than relying on the provided data.
Epistemic blinding involves substituting entity names with anonymous codes (e.g., Gene_001, Company_042) and then A/B comparing the results with the original names exposed.
The key point is not that 'the blinded result is better,' but that 'without blinding, there is no way to know if the LLM followed the data or its memorization' – restoring auditability.
Through ShinkaEvolve, an evolutionary optimization using Claude Sonnet as a mutation operator, a scoring function was evolved over 13 generations using only numerical features without gene names.
Contamination is more severe when feature signals are weak. Consistency is 90% in AML with strong signals and 75% in GBM with weak signals, showing a predictable pattern of contamination.
Implemented with Claude Code skill, allowing users to automatically execute blinding → A/B comparison → de-anonymization with a single command without separate scripts.

Evidence

"In predicting the top 20 drug targets for 4 oncology cancer types, there was an 84% average agreement between blinded and unblinded results → 16% of predictions changed, but the recall of validated targets was the same in both conditions (average 2.75/cancer type).\nIn GBM (glioblastoma), DPP8 was #3 in the blinded condition but dropped to #9 when the name was exposed, while PTEN jumped 12 places from #15 blinded to #3 unblinded.\nIn S&P 500 value investing screening, 7 out of 20 (35%) of the top companies changed when the ticker was exposed, and ELV and CI showed a systematic bias of moving up in the rankings 4 out of 5 times when unblinded.\nEvolutionary optimization results: The final function among 53 candidate functions in generation 13 achieved a strict hit rate of including FDA-approved targets in the top 20 for 10 out of 18 diseases, with fitness improving from 68% to 82% compared to the initial naive function (Generation 0)."

How to Apply

"When asking an LLM to rank/score based on data, specify the entity columns (gene names, company names, candidate drug names, etc.) in a YAML config and anonymize with blind.py, then send each A/B prompt to a fresh session. If the results differ significantly, the LLM answered with its prior knowledge, not your data.\nIf feature values can identify entities (e.g., $3 trillion market cap + smartphone sector = Apple), normalize or bin that feature by sector to reduce structural identifiability before blinding.\nIf using an agentic workflow in the Claude Code environment, adding epistemic-blinding as a Claude Code skill will automatically run blinding → comparison → de-anonymize in one turn when requesting ranking/prioritization."

Code Example

snippet

# Installation and basic usage
# git clone https://github.com/mcuccarese/epistemic-blinding

# 1. Write YAML configuration file (config.yaml)
# datasets:
#   - path: genes.csv
#     entity_columns: [gene_symbol]
# task: "Rank the top 20 drug targets based solely on the provided features."
# seed: 42

# 2. Generate blinded prompt
python blind.py --config config.yaml
# → blinded_prompt.txt, unblinded_prompt.txt, mapping.json are created

# 3. Send each to LLM and compare results
python compare.py \
  --blinded blinded_output.txt \
  --unblinded unblinded_output.txt \
  --mapping mapping.json
# → Set overlap, Jaccard index, Mean rank delta, Kendall τ are output

# 4. Restore anonymous code → actual entity
python deblind.py \
  --response blinded_output.txt \
  --mapping mapping.json

# Use Claude Code skill (one command)
# "Rank these companies by value. Use epistemic blinding."
# → Automatically runs blind → A/B execution → deblind

Terminology

parametric knowledgeKnowledge memorized by the LLM from training data. It's like a student who has memorized the textbook answering a test question by mixing memorized content into the answer.

epistemic blindingA technique similar to double-blinding in clinical trials, where the LLM is given analysis targets with their names hidden to make judgments purely based on the data without learned biases.

contaminationHere, it refers to the phenomenon of the LLM's pre-trained memory secretly mixing into the data analysis results. Similar to a taste being infused into food.

auditabilityThe property of being able to trace where the analysis results came from. Code can be debugged to trace the execution path, but this solves the problem that LLMs cannot do that.

evolutionary optimizationAn optimization technique that automatically improves code/formulas in a way similar to biological evolution, by selecting good functions and discarding bad ones.

fame biasThe bias where LLMs rank famous entities (KRAS gene, Apple stock) higher without data basis. It's the same as the halo effect where famous people seem better.

Jaccard indexA metric for measuring how much two sets overlap. Used to compare how much the top 20 lists in blinded/unblinded conditions are the same. 1 means completely identical, 0 means completely different.

Kendall τA statistic for measuring how similar two ranking lists are in order. Used to see the agreement in rank order between blinded and unblinded results.

Related Resources

epistemic-blinding GitHub (blind.py, deblind.py, compare.py)

Original Abstract (Expand)

This paper presents epistemic blinding in the context of an agentic system that uses large language models to reason across multiple biological datasets for drug target prioritization. During development, it became apparent that LLM outputs silently blend data-driven inference with memorized priors about named entities - and the blend is invisible: there is no way to determine, from a single output, how much came from the data on the page and how much came from the model's training memory. Epistemic blinding is a simple inference-time protocol that replaces entity identifiers with anonymous codes before prompting, then compares outputs against an unblinded control. The protocol does not make LLM reasoning deterministic, but it restores one critical axis of auditability: measuring how much of an output came from the supplied data versus the model's parametric knowledge. The complete target identification system is described - including LLM-guided evolutionary optimization of scoring functions and blinded agentic reasoning for target rationalization - with demonstration that both stages operate without access to entity identity. In oncology drug target prioritization across four cancer types, blinding changes 16% of top-20 predictions while preserving identical recovery of validated targets. The contamination problem is shown to generalize beyond biology: in S&P 500 equity screening, brand-recognition bias reshapes 30-40% of top-20 rankings across five random seeds. To lower the barrier to adoption, the protocol is released as an open-source tool and as a Claude Code skill that enables one-command epistemic blinding within agentic workflows. The claim is not that blinded analysis produces better results, but that without blinding, there is no way to know to what degree the agent is adhering to the analytical process the researcher designed.