# HindsightBench: A Black-Box Behavioral Audit Protocol for Parametric Hindsight in Time-Indexed LLM Decision Tasks

arXiv: https://arxiv.org/abs/2607.18867 · HTML: https://mlsanity.com/paper/2607.18867/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 19 of its 30 references that are in the map.
- First posted on arXiv 21 Jul 2026.

## Builds on (most cited first, 19 of 19)

- [Auditing large language models: a three-layered approach](https://mlsanity.com/paper/2302.08500/) (2023)
- [Black-Box Access is Insufficient for Rigorous AI Audits](https://mlsanity.com/paper/2401.14446/) (2024)
- [Evaluating Quantized Large Language Models](https://mlsanity.com/paper/2402.18158/) (2024)
- [Dated Data: Tracing Knowledge Cutoffs in Large Language Models](https://mlsanity.com/paper/2403.12958/) (2024)
- [Assessing Look-Ahead Bias in Stock Return Predictions Generated By GPT Sentiment Analysis](https://mlsanity.com/paper/2309.17322/) (2023)
- [Accuracy is Not All You Need](https://mlsanity.com/paper/2407.09141/) (2024)
- [Profit Mirage: Revisiting Information Leakage in LLM-based Financial Agents](https://mlsanity.com/paper/2510.07920/) (2025)
- [Look-Ahead-Bench: a Standardized Benchmark of Look-ahead Bias in Point-in-Time LLMs for Finance](https://mlsanity.com/paper/2601.13770/) (2026)
- [From Knowing to Doing: A Memory-Controlled Benchmark for LLM Trading Agents on Stock Markets](https://mlsanity.com/paper/2605.28359/) (2026)
- [Quantized but Deceptive? A Multi-Dimensional Truthfulness Evaluation of Quantized LLMs](https://mlsanity.com/paper/2508.19432/) (2025)
- [A Fast and Effective Solution to the Problem of Look-ahead Bias in LLMs](https://mlsanity.com/paper/2512.06607/) (2025)
- [Detecting Lookahead Bias in LLM Forecasts](https://mlsanity.com/paper/2512.23847/) (2025)
- [DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining](https://mlsanity.com/paper/2603.11838/) (2026)
- [ExAnte: A Benchmark for Ex-Ante Inference in Large Language Models](https://mlsanity.com/paper/2505.19533/) (2025)
- [Simulated Ignorance Fails: A Systematic Study of LLM Behaviors on Forecasting Problems Before Model Knowledge Cutoff](https://mlsanity.com/paper/2601.13717/) (2026)
- [Instruction Tuning Chronologically Consistent Language Models](https://mlsanity.com/paper/2510.11677/) (2025)
- [LLMLagBench: Identifying Temporal Training Boundaries in Large Language Models](https://mlsanity.com/paper/2511.12116/) (2025)
- [Summoning the Oracle to Slay It: Mitigating Look-Ahead Bias in Financial Backtesting with Large Language Models](https://mlsanity.com/paper/2605.24564/) (2026)
- [NumLeak: Public Numeric Benchmarks as Latent Labels in Foundation Models](https://mlsanity.com/paper/2605.30393/) (2026)

## Abstract (arXiv)

Large language models leak parametric knowledge of what followed a historical date into decision tasks indexed by that date -- not necessarily a lookup of the realized outcome, but knowledge of the period all the same. Existence is settled; what users lack is a cheap way to audit a given model. We present HindsightBench, a black-box audit protocol that profiles parametric hindsight in any time-indexed LLM decision task at probe-level cost (no backtests, no logprobs, no corpus access). It chains a four-arm date-manipulation matrix (revealed/date-only/masked/transplanted), dual memory probes (date recovery; outcome recall), and six metrics -- trigger strength, transplant effect, post-cutoff placebo, recoverability, behaviorally effective cutoff, and recall-accuracy dissociation -- with explicit gates where identifiability is data-dependent. Applied to 15 models from seven vendors on a 258-node vintage-correct macro panel, it yields three patterns: (i) the date-trigger reflex is not a scale phenomenon -- it tracks training recency, though what installs it is not identified here: absent across every 2024 open-weight row where it is measurable, including a 70B tier with cutoff-aligned recall propensity, present in every tested 2026-generation model, and switching on within one vendor lineage (Qwen3 -> Qwen3.6) in the same MoE family at ~3B active; (ii) effective cutoffs span 22 months across vendors and precede vendor-reported dates by up to eight months, invalidating calendar-window placebos; (iii) results are not invariant to serving -- BF16 serving of an FP8-referenced model breaks the trigger estimate's stability while AWQ-INT4 preserves it, and a provider-locked reasoning regime makes one probe non-convergent -- so the protocol pins quantization and thinking regime as part of its contract. We release the panel, preregistrations, audit rows, transcripts, and one-command regeneration.