assign_triggers() gains a conditions
argument that fixes the order of the condition codes, and
export_experiments() gains one that defaults to the order
of the design’s conditions. The codes used to follow the
order in which the conditions first appeared in the counterbalanced
trials, so a different seed could swap them, and
low_frequency was 101 in a design that lists
high_frequency first. The pipeline now passes the design’s
order, followed by any other condition in the order the item source
lists it, so the first condition the design lists is 101. The condition
codes of 12 of the 21 demonstration designs change with it.
assign_triggers() called without conditions
behaves as before.cohens_d_ci() uses the large-sample standard error of
d, sqrt(1/nx + 1/ny + d^2 / (2 * (nx + ny))) (Hedges &
Olkin, 1985; Borenstein et al., 2009). It left out the d^2
term, which treated the pooled SD as known and made the interval on a
manipulated dimension far too narrow. In the English frequency contrast,
d = 5.27 had the interval [5.01, 5.53] and now has [4.72, 5.83]. A
matched control, whose d is near zero, moves by at most 0.002 across the
21 demonstration designs, and no d, TOST p-value or verdict changes. The
comparisons, datasheets and run logs of the designs with a manipulated
dimension change with it. No stimulus moves.match_report() documents the direction of its signed
statistics. cohens_d and its interval are the reference
(first) condition’s mean minus the mean of the condition in that row,
and var_ratio is that condition’s variance over the
reference’s.lexsync_cache_clear() removes a fetched corpus from
the cache, together with any partial download of it, or with no argument
removes the whole cache. It returns the paths it removed,
invisibly.lexsync_cache_dir() only reports the cache’s path. It
created the directory on every call, so asking where the cache was left
an empty directory behind. fetch_corpus() now creates the
cache when a download into it starts, once the registry entry and its
URL have been accepted, so a refused call writes nothing. That download
begins by deleting any .part sidecar in the cache that has
gone a day without a write. An interrupt or a process that died
mid-transfer leaves one behind, and no handler removes it. A younger
sidecar may belong to a download still running in another session, so it
is left alone. A dest of the caller’s choosing behaves as
before.citation("lexsync") gives the DOI that CRAN assigned to
the package, 10.32614/CRAN.package.lexsync, with the CRAN page as its
URL. The 0.1.0 build on CRAN predates the DOI and gives the repository’s
address. The entry names CRAN as its publisher, and its text version is
the APA reference for software: author, year, title, version, “[Computer
software]”, CRAN and the DOI.templates/ directory found the unrelated
‘templates’ package there and failed. A test that compares the package
with repository files now accepts a root only if it holds lexsync’s own
DESCRIPTION under R_workflow/ and the Python
package beside it, and skips on CRAN.local_mocked_bindings() can replace an imported function,
as the corpus download tests do with download.file().CHANGELOG.md for the full,
cross-language history and the planned methodological roadmap.These notes follow the package through its development up to the release, so they describe each change against the state it replaced.
Depends: R (>= 4.0.0).
tools::R_user_dir does not exist before 4.0.0, and
round()’s post-4.0 algorithm shapes artefact bytes, so an
older R would fail obscurely or write different bytes without
complaint.describe_stimuli() was the one reduction still on
stats::median, which averages the two middle values through
mean()’s long-double accumulator while the Python engine
reduced through numpy. Both engines now sort and take the exact middle,
with (a + b) / 2 in plain double arithmetic for even n, and
regenerating every design moved no byte. Reproducibility and
parity claimed that every reduction shared one
compensated-summation algorithm, which the median never did, and now
describes the sort-and-middle rule beside it.define_by selects identically in
both engines. This engine always fell back to the whole pool. The Python
engine crashed with a bare KeyError on the same design and now mirrors
the fallback, pinned by twin tests.INTER_TRIGGER_S is substituted through
%.17g like its neighbours in the generated PsychoPy script.
It used as.character(), whose rendering of a non-integer
quotient differs from Python’s str(). The shipped default
still renders “0.01”, so no committed experiment byte moved.hash_unit() always has, so a latin1-marked condition read
from a user’s CSV can no longer rank by different digests in the two
engines.R_workflow/run_pipeline.R prefers an installed
lexsync over the edited sources, so edits silently did nothing until a
reinstall. It now states the loaded copy and the remedy on stderr at
startup.design_sha256, which must not depend on the operating
system), and its parity caption now matches the keyed-hash guarantee:
the engines produce byte-identical stimuli and trial order, where the
caption used to promise only the stimuli. The Python engine’s twin fixes
land in the same release: its writer refuses 16-digit integer columns as
this engine always has, its scalar rounder passes an overflowing scale
through, and run_pipeline(verbose = FALSE) keeps its
console silent.n_per_condition is now
an error. Every selector used to clip the request to the available pool
and said so at most through a verbose message, while the datasheet and
the generated Methods text kept stating the requested n. The new
matching: shortfall policy defaults to error,
and allow accepts the shrink. Its sibling
matching: on_insufficient_tolerance (default
relax) can likewise turn the silent tolerance-window
widening into a refusal, and a relaxation that does happen is now
recorded in the run log and the datasheet
(selection.window_relaxations), where before it was only
narrated to the console. No committed design experiences either
condition.joint and optimal matching can no longer
select the same word twice. With overlapping condition windows a word
could be paired with itself at zero cost, or mirrored across two sets,
appearing in both conditions. Both matchers now track used words as the
anchored matcher always has, and every matcher asserts its output holds
each word at most once. Every committed design has disjoint conditions,
and those select identically, byte for byte.condition reached Python’s hash-key guard as the string
“nan” and passed the very check built to refuse it, a blank
item cell turned numeric identifiers into “1.0”, and R
failed on the same table with a bare “missing value where TRUE/FALSE
needed”. Missingness is now checked before any coercion,
item is read as text, empties after trimming and duplicated
item-and-condition rows are refused, and both engines say the same
thing.equivalent: false. The
sentence is now conditional on the stored verdicts, prints the signed
worst difference and the configurable bound, and the analysis note names
the suggested formula as lme4 syntax, where it used to
imply that statsmodels accepts it. The datasheet also gains
the SHA-256 of the design and schema, the pairwise matchers’ candidate
cap and whether it fired, the equivalence bound tested against, and the
yaml, stringi and operating-system entries the
environment record omitted.continuous block now works over a supplied pool, and
two silent table-mode gaps are refusals. source: pool
reaches the continuous selector (it previously fell through to the
conditions matcher and crashed). A continuous table without
members is an error now, as is pool_filters on
a plain table design, where each used to do nothing silently while the
provenance recorded otherwise. Misspelt pool_filters or
define_by columns, duplicate condition names, reversed or
non-finite ranges, negative tolerances and non-integer n are likewise
refusals now..round_dp and a new vectorised numpy twin), replacing a
pairing of R’s round() with numpy’s, two rounders the
project measured to disagree at boundaries. Regenerating all 21 designs
moved no stimulus, report or experiment byte. A custom norm named in
match_on now also reaches the descriptives, comparisons and
realised-control record, which previously listed only the built-in
dimensions.language_tag is
shape-checked before it is used. Escaping them correctly would have
taken three rules per engine. .pyq() escapes newlines now
as well: an .osexp is line-oriented, so a raw newline
closed the inline-script block and let the rest of the value start a new
top-level item. Pinned by test-injection.R. No generated
artefact changed. Adversarially attacking that fix then found four ways
through it, all now closed and pinned. The largest: a response event’s
keys were joined into OpenSesame’s
set allowed_responses "a;b" with no validation at all, and
that is one line of a line-oriented format, so a key holding a double
quote closed the string, a newline ended the line, and the rest of the
value became new top-level items in the experiment, including an
inline_script whose body OpenSesame runs. Keys are
validated now. The shape guards were also anchored with $,
which in both Python’s re and R’s PCRE matches just before
a final newline, so a port address ending in one passed the check. They
are anchored at end-of-string now. A scalar keys: space or
blocks: practice was iterated character by character in
Python and kept whole in R, so one design gave two different
allowed-response lists and a block-restricted event ran everywhere in
one engine and nowhere in the other. And R’s HTML escape did not cover
U+2028/U+2029, which end a line in JavaScript but are not ASCII
controls: Python escaped them, R did not, so the same design produced
different bytes and R’s <script> was a syntax error
before ES2019.write_csv_utf8() reproduced readr’s format for
small magnitudes and left the top end alone, since nothing lexsync
computes reaches it. Nothing lexsync computes does. A joined norm table,
a supplied pool or an item table carries whatever columns the user has,
and those go straight into the stimuli CSV, so the guarantee held for
the shipped designs and failed silently for the user’s own data. readr’s
layout above 1e15 could not be reproduced in Python, which writes 1.5e16
as 15e15, the largest double as
17976931348623157e292, and the double nearest 5e22 as
4.9999999999999996e+22. Both engines now refuse such a
value and name the column. One engine accepting it would leave the two
writing different bytes. A value with two equally short decimal forms is
refused for the same reason. Everything verified to render identically,
across 465 values compared against readr’s own output, writes as
before.feedback event
or a blocks: restriction would not run. The emitter wrote
unicode(...), a Python 2 builtin that OpenSesame’s Python 3
inline workspace does not provide, so the experiment died with
NameError on the first trial. It also passed
None as a var.get() default, which OpenSesame
cannot distinguish from no default, so it raises where it should return.
Both are fixed, both spellings are now pinned by a test, and a
feedback event with no preceding response or
question is refused at generation time. Left to run time,
that one design error would surface as three different failures, one per
target.round() disagrees with Python’s builtin,
Python’s builtin disagrees with numpy’s, and even R’s
sprintf("%.3f") disagrees with Python’s "%.3f"
on 274 of them. Some values move by one in the last published digit. No
selection changes.condition cell, a routine data error neither reader
rejects, rendered "NA" in R and "nan" in
Python, so the two engines produced different trial orders from the same
design, reproducibly and with nothing to signal it. Booleans get a
pinned spelling and are still accepted.joint and
optimal matchers goes through the compensated reduction.
The cap does fire on shipped designs, so it decides which candidates
reach matching. It was verified to change no design’s selection.R CMD check --as-cran is clean: 0 errors, 0 warnings, 1
note (the standard new-submission note). Fixed on the way: three
undocumented arguments on select_continuous_stimuli(), an
unqualified utils::head(), and the package’s only example,
which called an unexported function from inside \dontrun{}
and so could never have worked if a user copied it. The package now has
executable, offline examples, and cran-comments.md no
longer misstates the check result.practice: and
fillers: item tables. Those trials are presented but not
analysed, so the stimuli file and the reports are written from the main
rows while the generated experiments run every trial. Practice comes
first. Fillers are interleaved with the main trials, because a block of
fillers at the end is not a filler but a second block a participant can
tell apart. A design declaring neither is unaffected, down to not
gaining a block column.feedback trial event, and a blocks: key
that restricts any event to named blocks. Together these confine
feedback to practice, which is the usual arrangement: feedback teaches
the mapping, and would contaminate the reaction times it is measuring.
Implemented for PsychoPy, OpenSesame and jsPsych alike. The PsychoPy
runner now returns the pressed key so it can be scored, and each runner
pauses at a block boundary.NaN in the generated
jsPsych experiment where a trial had no value for a field another block
supplies (a main-block trial in a design whose practice items carry an
answer). That is not valid JSON, and the R engine dropped
the key instead, so the two engines’ experiments differed. Both now drop
it.1 and
1.0 are the same number. Thirteen of the 18 shipped designs
differed byte for byte while the gate stayed green. Three differences
were serialisation (a whole number written 1 and
1.0, a boolean written FALSE and
False, a small value written 9e-4 and
0.0009) and one was not: two reported means differed in the
last decimal published, because numpy sums pairwise and R’s
mean() does not. Every reduction in the package now uses
one compensated-summation algorithm written out in both engines, whose
agreement follows from IEEE-754 and no longer from measurement. The
writers agree on every value, and the artefacts are now compared as
bytes. No R golden moved, because R’s two-pass mean was already the
correctly-rounded one.f was silently turned into
FALSE. readr reads a column whose values are
all f, t, T or F as
logical, while pandas keeps the string, so an item table using the
commonest two-choice key pair had its correct answer corrupted in one
engine. Item tables now read the condition label and the paradigm’s
presented fields as text in both engines.jsonlite’s default of four digits was truncating the
datasheet: a design declaring
tolerance_k: 0.1111111111111111 had it recorded as
0.1111, which does not reproduce the run the record exists
to describe. The JSON is now written at full display precision, and the
Python engine writes the same precision.categorisation: a category cue, then the
word to judge against it, with answer holding the correct
response key so the data are scoreable. Counterbalanced by Latin-square
rotation, so a participant never sees the same target twice.counterbalance.optimise (off by default) assigns item
sets to lists so the lists are equated on the declared dimensions, where
the old deal went by set rank. The search is a deterministic integer
descent with a keyed-hash tie-break, so it uses no random number
generator and both engines produce the same assignment.
balance_lists() is exported.pool: a supplied word list goes through
the matcher without having to masquerade as a corpus lexicon.
load_pool() is exported. The neighbourhood dimensions are
computed against the lexicon’s words, because a word’s neighbours are
its neighbours in the language.norms: block. They
are joined onto the lexicon before the candidate pool is built, so a
semantic dimension the corpus does not carry (concreteness, age of
acquisition, valence) can be filtered on, matched on or spanned like any
other. No norm data is bundled. Every joined table is recorded in the
materials datasheet with its checksum and its per-column coverage,
because a norm table can supply the very variable a design
manipulates.materials_source$norms
and, for a pair-keyed design, a relational block naming the
members, the pair count, the member lexicon and its checksum, and the
member-level dimensions separately from the relational ones.
selection$cross_engine no longer reports “n/a
(user-supplied items)” for a pair-keyed continuous design: that design
does perform a selection, and it is byte-identical across engines.merge_norms() returns the lexicon with the norm columns
appended, in the lexicon’s own row and column order. It
previously used merge(), which hoists the join column to
position 1 where pandas keeps the left frame’s order, so the two engines
disagreed on column order whenever on was not already
first. The join key is now case- and whitespace-folded on the lexicon’s
side too, where only the norm table’s was folded before: a lexicon
holding Dog used to match nothing and leave an
all-NA dimension. A norm column whose name already exists
on the lexicon is now an error. Renaming it to frequency.x
/ frequency.y, as merge() did, left the
matched dimension under a name nothing looks for.write_datasheet() and write_run_log() now
write LF on every platform. They used a text-mode connection, so on
Windows the datasheet, its Markdown rendering and the Markdown run log
came out CRLF while the Python engine’s twins were LF. The datasheet is
the provenance record. Its bytes no longer depend on the machine that
built it.add_pair_overlap() and
resolve_trial_timing() are now exported in fact as well as
in the documentation. Both were marked for export and documented, but
the NAMESPACE had not been regenerated, so
library(lexsync) did not make them available.items.generation.method: subsyllabic): whole
onset/nucleus/coda constituents are swapped for attested same-role,
same-length alternatives, preserving syllabic structure and length.
Byte-identical across engines, and the default letter-substitution
generator is unchanged.continuous block and
select_continuous_stimuli() spans a predictor’s range
evenly while holding the controls near-constant, with predictor-control
correlations and a regression suggested-model in the datasheet.
Byte-identical across the R and Python engines.mahalanobis (a
covariance-aware distance that down-weights correlated dimensions) and
optimal (a globally optimal assignment for two-condition
designs, using the suggested clue package). Unlike the
default methods, these use a covariance inverse and an assignment
solver, so the R and Python engines agree closely but not byte-for-byte
on them.matching.method, and a candidate pool too
small for the requested n_per_condition, now raise an
actionable error. Each used to fall back silently, to a default method
or to a short set. The Python engine raises the same message, and code
that relied on the old fallback will now stop.stringi is a new hard dependency (Imports). The
canonical word key is case-folded with ICU at the root locale, so the
key no longer depends on the machine’s locale and matches the Python
engine byte for byte. shiny, bslib,
DT and zip are new Suggests, for the Shiny
app.PARADIGMS(),
build_datasheet(), build_lexdec_stimuli(),
count_syllables(), generate_pseudowords(),
make_pseudoword(), methods_paragraph(),
required_fields(), resolve_events() and
write_datasheet().tost_equiv() now defaults to
bound_d = 0.5, the value the Python engine already used.
The R default was 0.4, so the equivalence bound is wider and a
comparison is easier to declare equivalent. Reported verdicts can change
for the same data.describe_stimuli() now orders rows by each group’s
first appearance, matching pandas. This changes the row order of
generated descriptives.seed|replicate|list|set|condition, a tuple that identifies
the trial uniquely under either counterbalancing recipe, so the
permutation is a pure function of the design: byte-identical from both
engines on any platform, and different for every seed. Previously the
order was drawn from sample(), which could never match
numpy’s generator for the same seed, so the trial lists were the one
engine-specific artefact. All 75 generated experiment files across the
15 designs bundled at the time are now byte-identical across the
engines, and the parity gate now compares the trial column
of the stimuli CSVs. Stimulus selection, pairing and lists are
unchanged, but every design’s trial order changes relative to the
previous artefacts. The package no longer uses R’s random-number
generator at all, so a seeded run cannot perturb the calling script’s
random stream and there is no RNG state to save or restore.merge_norms() preserves the lexicon’s row order, and
participant_table() crosses factors in
expand.grid() order in both engines.match_stimuli() raises where it used to re-pick an
already-used row, when a relaxed window re-admits candidates missing a
matched dimension. Such a candidate has no defined distance and is never
assigned, yet it still counted towards the pool-size guard. An
NA-depleted pool could therefore pass the guard and go on to emit the
same word in several sets.match_stimuli() and
select_continuous_stimuli() no longer select an all-NA row
when the tolerance window is NA (an anchor of a single item gives
sd = NA). An undecided comparison now resolves to
FALSE, as it does in Python, so the window is relaxed and a
real word is selected.