Beat the Guesser / 别被猜中 — product and provenance notes
15 September 2026

This is an original implementation of a left/right prediction duel. The interface,
icon, scorecard and game code were made for this product. It includes no extracted
game assets, recognizable third-party melodies, character art or sampled audio.
No existing MindReader source code or media was copied into the app.

The general idea of predicting binary choices is established. Research references:
https://www.cs.williams.edu/~bailey/applets/MindReader/
https://marisae.github.io/MindReader/webVersion/
https://lots.pajitool.com/en/predict-my-next

Topic research recorded on 15 September 2026:
https://www.reddit.com/r/singularity/comments/1wgse1y/ask_your_llm/
https://www.reddit.com/r/ChatGPT/comments/1wgseg8/try_it_yourself/
These discussions contain anecdotes about random choices. They are not controlled
benchmarks, proof of global popularity, or validation of this game's predictor.

The opponent fixes its next prediction before accepting the player's next choice.
The displayed tally is based on the actual prediction/choice history. It runs
locally without calling a language model. It has no access to thoughts or future
input, and its accuracy is not guaranteed. A 1,000-choice run is entertainment,
not a scientific randomness, intelligence or personality assessment. Overall and
recent hit rates report only actual matches; the recent window contains up to
the latest 100 choices, not a population benchmark.

The version 1.1.0 predictor is an original causal ensemble. It combines contexts
of up to ten earlier choices at three recency rates, same/inverse lag predictors
up to forty moves back, and recent choice/result reaction contexts. Its expert
weights adapt to their actual earlier forecasts using two discounted log-loss
weightings. The locked prediction is scored before updating any model with the
current choice. Restoring a saved run validates its history by replaying it.
Synthetic pattern and random-control tests check implementation behavior; they
do not establish a guaranteed improvement against humans or a psychological claim.

Version 1.1.0 keeps the current run in local browser storage when available.
Reloading the same challenge resumes that run. A rematch starts fresh, and
learning does not carry across different games. Local progress does not sync
between browsers, devices or domains. Individual choices are not sent to the
shared analytics service.

Version 1.2.0 adds an optional local commitment/reveal inspector. Before a move
is accepted, it publishes a SHA-256 commitment to that prediction plus a fresh
256-bit random nonce and versioned run/seed/model/turn/previous-hash context.
The prediction, nonce and exact encoded payload are revealed after the move.
The full pending hash can be copied; downloadable audit JSON supports independent
recomputation and chain checks. The hash-chain mechanism and vector graphics are
original code for this app, not copied game assets.
The digest uses an original synchronous implementation of SHA-256 from NIST
FIPS 180-4; random nonces use the browser's crypto.getRandomValues API.

The pending nonce and commitment are kept with local progress to preserve the
same next-move commitment across a reload. Older saved runs start the audit at
the next unplayed turn; no commitment evidence is invented for earlier turns.
The audit remains local and adds no remote prediction service or server receipt.

Security boundary: a hash checks whether revealed content matches a commitment
already held by the verifier. A later exported file alone cannot establish that
its hash was shown before the player's input. Independently retain the full
commitment before a move to check it afterward. Self-reported timestamps, an
unanchored hash chain and an operator-controlled server are not independent
witnesses. This inspector does not certify that the website or its script is
incapable of alteration.

Primary background references checked on 15 September 2026:
NIST Secure Hash Standard: the SHA-256 algorithm.
https://www.nist.gov/publications/secure-hash-standard
MIT 6.857 course notes: commitments using a hash and random nonce.
https://people.csail.mit.edu/alinush/6.857-spring-2015/l14-public-key.html
W3C Web Cryptography API: random values, SHA-256 digest and script trust limits.
https://www.w3.org/TR/2017/REC-WebCryptoAPI-20170126/