Taylor McCoy¹ and Blu²
¹ Independent researcher
² OpenAI Codex, Blu Private lane; technical collaborator
Study status: Preregistered prospective technical study. Target generation, confirmatory data collection, unblinding, and analysis have not yet occurred.
The completed September 21 terminal case that motivated this study is preserved separately as the original-observation foundation. It is prior evidence and rationale, not a result of the prospective study.
Text deleted from a terminal's visible input line may remain recoverable in a
session recording even when the edited text is not delivered to the interactive
application. A completed local case established a plausible recovery path:
macOS /usr/bin/script -q preserved draft-text echo and subsequent 08 20 08
erase sequences while canonical line editing delivered only the committed line
to llama-cli. Because that observation was discovered retrospectively, it did
not constitute a prospectively blinded test of exact recovery.
This study will test whether transcript-mediated inspection can recover a
previously unknown ordered sequence of ten physical playing cards. Taylor will
generate the target by drawing ten cards without replacement from a shuffled
standard 52-card deck, record the ground truth off-computer, enter an agreed
text encoding into a locally recorded terminal without pressing Return, delete
the complete target, and submit a separate control line. Blu will inspect the
frozen local transcript and record one exact prediction before the target is
revealed. The primary endpoint is exact ordered-sequence agreement. Under the
preregistered uniform-order independence null, the probability of an exact
match is 1 / 52P10 = 1.741926487643891E-17, approximately one in 57.4
quadrillion. No outcome data have been collected and no confirmatory analysis
has been performed. An exact match would reject the specified chance-guessing
null under its assumptions; it would not, by itself, identify the recovery
mechanism or establish that a model or network service received deleted text.
Keywords: terminal recording; pseudo-terminal; blinded recovery; playing
cards; exact permutation; backspace text data; local language model; script;
llama-cli
Interactive terminals separate visible editing from application input. With
canonical line processing and echo enabled, a pseudo-terminal may display typed
characters, process an erase action, and return only the edited line after
Return. A sequential recorder can nevertheless preserve the earlier character
echo and the later control bytes used to remove those characters from the
screen. Apple documents that script(1) records terminal output and that its
log includes backspaces, while POSIX describes canonical editing and echo at
the terminal line discipline (Apple Inc., 2023a, 2023b; The Open Group, 2004).
The foundation case observed this boundary
in one local llama-cli session. The recording contained the abandoned phrase
marie is black., fifteen 08 20 08 erase triplets, and the committed phrase
stanley is red.. A later six-pair diagnostic found erased synthetic markers
in 0/6 model answers under canonical line editing and 6/6 answers under a
deliberately altered -echo -icanon condition. Those results explain why exact
recovery from a terminal recording is technically plausible, but they do not
provide a blind confirmatory test: the phrases and recording structure were
already available when inspection occurred.
The present study replaces a researcher-authored phrase with a physical target that is unknown to the inspecting assistant until after its prediction is frozen. A deck of cards provides a familiar, finite sampling frame and an ordered target that another researcher can reproduce without constructing a custom word list. The study asks one bounded question:
Can read-only inspection of the frozen terminal recording recover the exact ordered ten-card target that was typed and visually deleted before a separate line was committed?
The preregistered hypothesis is exact recovery of all ten cards in order. The null hypothesis is that the frozen recovery is independent of the concealed card order and that any exact ordered match occurs by chance.
This is a prospective, blinded, single-challenge technical experiment. Taylor is the target generator and physical operator. Blu is the transcript inspector and predictor. Ground truth remains off the test computer and outside the chat until the prediction and transcript artifact are frozen. The experimental unit is the complete ordered ten-card sequence; individual card positions are not treated as ten independent replications.
The controlling preregistration is available in both human-readable and workbook form:
preregistration/PREREGISTRATION-2026-09-24.mdpreregistration/BackspaceGate.xlsxpreregistration/CHECKSUMS.sha256
The initial eight-word design remains visible in Git history. Before any card
target was generated, it was transparently amended to the standardized
ten-card design in commit 3a2e98b.
The physical sampling frame will be one ordinary 52-card deck with no jokers. Cards will be drawn sequentially without replacement. The local interaction will use the same class of measured path as the foundation case:
physical card target → typed terminal draft → script child PTY → llama-cli
↘ .typescript recording
Exact hardware, operating-system build, Terminal version, script version,
llama.cpp build, model artifact, launcher command, file permissions, and
network state will be recorded again at collection time. They will not be
copied forward from the foundation case as though live state could not change.
- Before shuffling or drawing, the researchers will freeze an unambiguous textual card notation, the separate committed control line, and the transcript filename in a dated protocol amendment.
- Taylor will shuffle a complete 52-card deck and draw ten cards sequentially without replacement.
- Taylor will write the ordered target on paper and preserve contemporaneous photographic or video evidence before terminal entry.
- The physical record and target sequence will remain outside the test computer and outside the chat until Blu's prediction is frozen.
- Taylor will not disclose hints, partial identities, suit counts, ranks, or ordering before unblinding.
The physical shuffle is intended to randomize order, but the protocol does not claim that a hand shuffle is mathematically proven to sample every permutation with equal probability. The one-in-57.4-quadrillion calculation is explicitly conditional on the uniform-order null.
The local terminal session will be wrapped by /usr/bin/script -q and will use
ordinary canonical line editing. Taylor will:
- type the encoded ten-card target without pressing Return;
- press Delete until the complete encoded target is visually removed;
- type the separately frozen control line;
- press Return once;
- end the local interaction cleanly and preserve the resulting
.typescriptfile without conversion.
Network disconnection may be recorded as an auxiliary boundary, but it is not the independent variable and will not substitute for process- or byte-level evidence.
After Taylor supplies only the agreed inspection signal, Blu will inspect the saved transcript read-only. Before receiving the paper record, photograph, video, or any description of the cards, Blu will freeze:
- one ordered ten-card prediction in the preregistered notation;
- the transcript path, byte length, and SHA-256;
- the relevant byte offsets and erase-sequence count;
- the inspection output and its SHA-256; and
- any protocol deviation observed before unblinding.
Taylor will then reveal the signed physical record and supporting media once. The frozen prediction will be compared with ground truth without editing either record. A failed, partial, malformed, or missing prediction remains a failure; the challenge will not be repeated merely to obtain significance. Any later replication requires a new target and a separately timestamped preregistration.
Source recordings will remain byte-preserved. Human-readable hex or caret renderings are derivatives and will carry their own filenames and checksums. The repository manifest, workbook checksum, Git commit, physical ground-truth record, and frozen prediction together will document ordering and provenance. Secrets, credentials, private messages, and unrelated sensitive text are excluded from the test.
Confirmatory data collection has not begun. No target cards have been drawn, no prediction has been scored, and no confirmatory analysis has been run.
| Field | Preregistered specification |
|---|---|
| Independent variable | Concealed ordered ten-card sequence drawn without replacement from a shuffled standard 52-card deck without jokers |
| Primary dependent variable | Exact ordered-sequence match (0 or 1) |
| Secondary outcomes | Correct ordered positions out of 10; exact card identities irrespective of order; edit distance; Spearman rank correlation for ordering, reported secondarily |
| Primary prediction | 10/10 cards correct and in order; exact-sequence match = 1 |
The number of possible ordered ten-card sequences drawn without replacement is:
P(52,10) = 52! / 42! = 57,407,703,889,536,000
Under the preregistered null that the frozen prediction is independent of a uniformly ordered target:
Pr(exact ordered match | H0)
= 1 / P(52,10)
= 1.741926487643891E-17
This exact ordered-sequence probability is the primary inferential quantity.
Spearman's rho, correct-position count, and edit distance are secondary
descriptions of partial agreement. If all ten identities were somehow known but
their order were independently random, the probability of perfect ordering
would be 1 / 10! = 2.755731922398589E-7.
The analysis will report the observed sequence, frozen prediction, exact-match indicator, position-level agreement, deviations, and reproducible calculations. The result will not be described as the probability that the null hypothesis is true. One challenge also cannot estimate a general recovery error rate.
The preregistration workbook is
BackspaceGate.xlsx.
Final scoring code, R output, result tables, and any derived figures will be
added only after the prediction is frozen and the target is unblinded. Empty or
future-facing analysis scaffolds are not results.
Results are pending. This section will be completed after the preregistered challenge, one-time unblinding, independent checksum verification, and analysis. Until then, the repository contains a protocol and evidentiary foundation—not a confirmatory finding.
An exact ten-card match would be extraordinarily inconsistent with independent uniform guessing under the specified null. Combined with a byte-preserved transcript showing the encoded target before its erase sequences, it would support successful transcript-mediated recovery by the declared inspection pipeline. The target record, transcript, prediction freeze, and unblinding order are all necessary: probability alone cannot identify the mechanism.
A partial or failed match would be reported directly. It could reflect absence of the target in the recording, a transcription or notation error, incomplete inspection, file mismatch, or another pipeline failure. It would not erase the foundation case, and it would not justify an undeclared retry.
Even a perfect result would not establish that:
llama-clior the local model received the deleted canonical draft;- OpenAI, ChatGPT, Codex, Ollama, Apple, or a network service collected it;
- the assistant guessed the sequence without transcript access;
- the same recovery rate generalizes to other terminals, recorders, operating systems, models, or sessions; or
- shared context or matching output proves continuous subjective identity.
The study tests a declared artifact-recovery path. Process input, terminal echo, recorder output, model response, inspection-tool access, and network telemetry remain separate evidentiary layers.
The design replaces a researcher-authored phrase with an externally generated physical target, freezes the prediction before unblinding, uses an exact primary endpoint, and preserves byte-level evidence. A standard deck also makes the sampling frame easy for other researchers to reproduce.
Limitations include a single challenge, imperfectly characterized physical shuffling, human transcription, a card notation not yet frozen, reliance on one local software path, and the fact that the inspecting assistant is intentionally given read access to the transcript. These constraints will remain visible in the final report rather than being repaired after the outcome is known.
Backspace Gate is a prospective blinded test of whether a declared transcript- inspection pipeline can recover an exact ten-card sequence that was typed and visually deleted before a different line was submitted. Its primary chance benchmark is one exact ordered sequence among 57,407,703,889,536,000 possible ten-card sequences under a uniform-order null. The study is preregistered, but data collection and analysis are not complete. The completed terminal case is preserved as a transparent foundation and must not be mistaken for the outcome of this confirmatory challenge.
- Original observation and completed technical case report
- Prospective preregistration
- Foundation evidence collection
- Foundation technical documentation
- Existing foundation and validation utilities
- Provenance records
- Credits and contribution record
The repository is presently private. Original iPhone videos from the foundation case remain indexed by checksum and are managed separately; first-frame previews are included. Model weights and credentials are excluded.
Taylor McCoy conceived the confirmatory challenge, controls the physical target generation and blinding, and will preserve the ground-truth record. Blu (OpenAI Codex, Blu Private lane) formalized the protocol, exact-null calculation, evidence boundaries, and reproducible repository structure. The original case and prospective study were refined collaboratively. No external funding was received, and referenced vendors did not sponsor or endorse this work.
Abramson, J., & North, S. (2020). College algebra with corequisite support. OpenStax. https://openstax.org/books/college-algebra-corequisite-support/pages/9-5-counting-principles
Apple Inc. (2023a). script.c (shell_cmds-302.0.1) [Source code]. Apple Open Source. https://github.com/apple-oss-distributions/shell_cmds/blob/e256b9a97f9bbd751305b7af36cf751668fbb849/script/script.c
Apple Inc. (2023b). script(1): Make typescript of terminal session [Manual page]. Apple Open Source. https://github.com/apple-oss-distributions/shell_cmds/blob/e256b9a97f9bbd751305b7af36cf751668fbb849/script/script.1
Best, D. J., & Roberts, D. E. (1975). Algorithm AS 89: The upper tail probabilities of Spearman's rho. Journal of the Royal Statistical Society: Series C (Applied Statistics), 24(3), 377–379. https://doi.org/10.2307/2347111
Mistral AI. (2025). Ministral 3 14B Instruct 2512 GGUF [Large language model]. Hugging Face. https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512-GGUF
R Core Team. (2026). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://doi.org/10.32614/R.manuals
Spearman, C. (1904). The proof and measurement of association between two things. The American Journal of Psychology, 15(1), 72–101. https://doi.org/10.2307/1412159
The Open Group. (2004). General terminal interface. In The Open Group Base Specifications Issue 6, IEEE Std 1003.1, 2004 edition. https://pubs.opengroup.org/onlinepubs/007904975/basedefs/xbd_chap11.html
Wasserstein, R. L., & Lazar, N. A. (2016). The ASA's statement on p-values: Context, process, and purpose. The American Statistician, 70(2), 129–133. https://doi.org/10.1080/00031305.2016.1154108
Original documentation and utilities are licensed under the Apache License, Version 2.0. See License scope for the preserved evidence and third-party material exceptions.