N'Ko as Computational Infrastructure: Script-Native Speech Recognition, a Phonemically Interpretable Error Metric, and Admissible Tone Correction

Abstract

This manuscript consolidates a multi-paper research program on , Manding
automatic speech recognition, script visibility in large language models, and
trajectory-conditioned decoding. The central argument is that should not be
treated as a decorative or interchangeable rendering of Manding language. For
machine-learning systems it functions as computational infrastructure: it determines
what tokenizers can represent, what hidden circuits are available, how acoustic
evidence is aligned to symbols, whether reported error rates measure speech
recognition or merely agreement with an inherited orthographic convention, and how
tone itself can be encoded and reconstructed from acoustic evidence.

The paper integrates five written project papers and subsequent audit notes into a
single canonical account. The representation studies show that current LLM families
accept Unicode strings while internally underrepresenting the script through
inflated translation cost, weak activation geometry, entropy gaps, sparsity
inflation, kurtosis deficits, and poor circuit-duplication yield. The speech papers
show a progression from early CTC systems to frozen-Whisper-feature decoders and
then to a trajectory-conditioned Transformer CTC decoder. In the canonical
architecture, 1280-dimensional Whisper large-v3 features are projected into a
768-dimensional decoder space, downsampled temporally, and decoded by a six-layer
Transformer CTC head. The anticipatory component computes a seven-dimensional
trajectory state $z_t$ for each timestep--commitment, uncertainty, transition
pressure, recovery margin, phase stiffness, novelty, and stability--and injects it
as an attention-logit bias $B_{ij}^{(m)}$ before CTC emission.

The mathematical claim is a measurement claim, not an automatic leaderboard
guarantee. I formalize a transparent-script proposition: if a normalized script map
$f_N:\Phi\rightarrow\Sigma_N$ is bijective over the target phoneme inventory, then
character edit distance over $f_N(\phi_{1:U})$ preserves the phoneme-edit structure
up to explicitly modeled normalization choices. A Latin transcription relation with
variable-length digraphs, optional tone marking, and spelling variation does not
have the same property. This makes CER a more phonemically interpretable
metric for Manding ASR than Latin WER, even though it is not a perfect phoneme error
rate and still depends on normalization, reference quality, and tone/diacritic
policy.

The strongest retained ASR artifact is an archived checkpoint trained on a
-pair Bambara corpus snapshot, with a 232,476/29,060/29,060
train/validation/test split, learning rate 0.0003, batch size 32, dropout 0.1, seed
42, and reported test of . This is the canonical public anchor
for discussing ``the 20\
that direct script-native ASR reached a meaningful error regime on a large
Bambara corpus; it is not a newly completed May 2026 strict reproduction, not a
result for later internal decoder variants, and not a closed proof that beats
Latin under every matched hyperparameter setting.

The Anticipation Geometry Partition (AGP) is likewise bounded carefully. AGP is not
the acoustic model that produced ; it is the post-ASR geometry and
governance layer that partitions transcript spans into stable, boundary, uncertain,
and novelty states before allowing conservative correction. The resulting
manuscript is best read as a systems-and-measurement paper: it connects script
invisibility in LLMs, script-native ASR architecture, trajectory geometry,
phonemically grounded evaluation, and row-level correction governance around a
single archived CER anchor.

Tone, historically the program's largest open hedge, is treated here as a structured
reconstruction problem rather than a deferred limitation. The same trajectory
geometry $z_t$ is reused at three levels---decode, govern, and correct---and tone
resolution is framed as the product of a linguistic prior over tone marks and
acoustic evidence about pitch. This positions two companion papers as the
reconstruction pillar of the program: a contextual tone model that resolves tone from
text and Featural Acoustic Coding (FAC), which resolves tone from *acoustic
$F_0$* and treats the syllable codebook as a featural sound code. A further
companion develops admissible expert routing (MAOE-) and an applied protocol
direction turns this stack into linguistic computation that workers, not GPUs, can
perform. The present paper is the umbrella and the measurement spine; the companions
are cited where they carry the empirical load, and every tone-related number is
reported with its provisional, OCR-noisy provenance.

Introduction

Modern multilingual NLP and ASR systems often inherit a hidden assumption: script is
a surface channel. A sentence may be written in Latin, Arabic, Cyrillic, Devanagari,
or another script, but the model is expected to learn the same linguistic content
once enough data is supplied. That assumption is convenient for model builders, but it
is scientifically unsafe for low-resource and indigenous scripts whose structure is
not visible in pretraining corpora, tokenizers, or standard evaluation metrics.

makes the failure concrete. It is a West African script designed for Manding
languages and encoded in Unicode at U+07C0--U+07FF [citation: unicode2006nko]. The
script also belongs to a social and intellectual movement around Manding literacy,
standardization, and linguistic self-representation [citation: donaldson2017clear].
Unlike Latin Bambara orthography, was deliberately engineered around Manding
phonology, including segmental distinctions and tonal marking. For a CTC ASR system
[citation: graves2006connectionist], that difference is not cosmetic. The model emits a
symbol sequence through time, and the scientific interpretation of its errors depends
on whether those symbols approximate acoustic-phonemic units or a historically
adapted spelling convention.

The research program behind this manuscript began as several papers rather than one.
Dead Circuits diagnosed invisibility inside LLM activations.
Living Speech documented the construction of a script-native ASR
stack. Script Invisibility Is Structural tested whether the activation
pattern extended across model families. Does Script Design Matter? developed
the trajectory ASR argument and produced the archived checkpoint.
Deployment Properties explored generalization, domain transfer, row-level
artifacts, and the correction layer that later became AGP. Later project notes
sharpened the story further: the compact trajectory scalars appeared more important
than heavier decoder additions, word error rate looked scientifically weak for tonal
Manding ASR, and the strongest immediate public claim became the metric-validity
argument around CER rather than an overbroad leaderboard claim.

This canonical version therefore makes a narrower but stronger thesis. is
computational infrastructure for Manding language technology. It changes model
visibility, acoustic decoding, metric interpretation, and the design of downstream
correction. The result is important because it anchors that thesis in a
large ASR artifact, but the number must be presented with its provenance and caveats.
The value of the work is not only that a checkpoint reported approximately 20\
it is that the surrounding evidence explains why that number is more meaningful than
Latin WER for this setting and why trajectory geometry belongs naturally in a
script-native ASR pipeline.

Caption: Core contributions of the canonical manuscript.

Initial Hypothesis Stack

Before formalizing the research questions, it is useful to reconstruct the
hypothesis stack that motivated the project. This prevents the manuscript from
appearing to have been organized around one surviving number after the fact. The
original program was broader: it asked whether changes model internals,
decoder alignment, evaluation metrics, deployment behavior, personalization, and
knowledge provenance. The canonical paper keeps the supported and relevant parts of
that program, while explicitly labeling the hypotheses that remain incomplete or
outside the present evidence base.

\begingroup

longtable{@{}p{0.07\linewidth}p{0.34\linewidth}p{0.29\linewidth}p{0.20\linewidth}@{}}

Caption: Initial hypotheses reconstructed from the project roadmap, handoff notes, and paper outlines. The later research-question section formalizes the subset that this canonical manuscript can support rigorously.

ID & Initial hypothesis & Intended test or observable & Current disposition

\endfirsthead

ID & Initial hypothesis & Intended test or observable & Current disposition

\endhead
IH1 & is not merely low-resource; it is structurally invisible inside LLMs
trained on corpora where the script is absent or nearly absent. & Layerwise
activation norms, entropy, sparsity, kurtosis, tokenizer allocation, and translation
tax on parallel English/ inputs. & Retained and supported by the LLM
diagnostic papers.

IH2 & Script invisibility is not model-specific. The same failure signature should
appear across architecturally distinct model families. & Repeat the activation scan
on Qwen-family and Mistral-family models and compare the geometry of failure. &
Retained and supported within the tested model families.

IH3 & Direct script-native ASR is technically feasible using frozen Whisper
features and a trainable CTC decoder. & Build audio-to- CTC systems and track
CER across BiLSTM, Transformer, LoRA/Whisper, and trajectory variants. & Retained;
the archived checkpoint is the strongest ASR anchor.

IH4 & A bijective or near-bijective script gives CTC a cleaner target than Latin
Bambara because target symbols correspond more directly to acoustic-phonemic events.
& Matched and Latin CTC runs with identical architecture, data, optimizer,
split, and scoring protocol. & Retained as a metric/architecture hypothesis; the
fully matched superiority proof remains unresolved.

IH5 & Trajectory bias acts as a bijection amplifier: dynamic speech-state information
should help most when the output symbols preserve phonemic structure. & Compare
baseline, graph, trajectory, and combined decoders in and Latin under matched
conditions. & Historically supported; full artifact bundle incomplete.

IH6 & should generalize better to unseen vocabulary because novel words
compose from familiar phoneme-character units. & Train on SEEN-only utterances and
evaluate on utterances containing UNSEEN words. & Retained as historical
generalization evidence.

IH7 & Vocabulary expansion should be more operationally useful in transparent-script
systems because adding words or graph entries preserves phonemic structure without
retraining the acoustic model. & Compare SEEN-only and full-data or graph-expanded
conditions on unseen-word utterances. & Partially retained; evidence is useful but
not the main anchor.

IH8 & Speaker adaptation or test-time training should be easier when the decoder's
labels are phonemically aligned and trajectory state captures speaker-specific
articulation patterns. & Diarize Djoko speakers, update decoder layers sequentially,
and estimate per-speaker CER or loss improvement slopes. & Planned/pending; not
claimed as a result.

IH9 & Out-of-domain deployment requires row-level provenance, uncertainty, and
correction governance rather than scalar CER alone. & Build prediction/reference
rows with trajectory summaries, partitions, provenance fields, and correction
decisions. & Retained as AGP architecture and artifact protocol.

IH10 & Training a personalized model on indigenous-script representations may change
behavior relative to colonial-language or English representations of the same
speaker data. & Compare LoRA adapters trained on English versus -encoded
speaker data using hidden-state and behavior metrics. & Out of scope for the
canonical ASR paper; preserved as future personalization work.

IH11 & Compressed -based knowledge representations can preserve provenance
through auditable transformation chains. & Track conversation-to-curation-to-
translation-to-sigil-to-inscription transformations and measure compression,
retention, and verification cost. & Out of scope for the canonical ASR paper;
preserved as future provenance work.

longtable
\endgroup

This table also explains why some claims are deliberately downgraded later in the
paper. The initial program contained aggressive empirical predictions: should
beat Latin under matched CTC conditions, trajectory should help more than
Latin, and adaptation should improve with speaker exposure. Some of those ideas have
supporting evidence, but not all of them have current artifact-complete support. The
canonical manuscript therefore treats the initial hypotheses as the research
program's generative map, then separates them from the narrower set of claims that
can be defended today.

Related Work and Technical Frame

Unsegmented sequence learning..
The ASR systems in this project use CTC, which was introduced to train recurrent
networks directly on unsegmented sequence-labeling problems such as speech and
handwriting [citation: graves2006connectionist]. CTC is relevant here because Manding
speech annotation is not frame-aligned; the model must learn a monotonic relation
between acoustic frames and character sequences. This makes the output alphabet more
than a formatting choice. The alphabet defines the units that CTC learns to align.

Whisper features and parameter-efficient adaptation..
The speech papers build on frozen Whisper large-v3 encoder features. Whisper was
trained at web scale for multilingual speech recognition and translation
[citation: radford2023robust], but the project does not assume that Whisper already
solves Bambara-to- transcription. Instead, Whisper provides acoustic features
while the script-native decoder learns the target writing system. The broader project
also used LoRA-style parameter-efficient adaptation [citation: hu2022lora] in the
language and speech stacks, but the canonical ASR anchor is not a
LoRA-Whisper result. It is a trajectory CTC decoder result recorded under its own
artifact path.

Model-family diagnostics..
The LLM papers use Qwen and Mistral-family models as diagnostic instruments for
script visibility. Qwen2/2.5 technical reports and Mistral 7B provide the external
model context [citation: yang2024qwen2,qwen25technical,jiang2023mistral]. The internal
research question is not whether those models are strong in general. It is whether
they carry functional internal structure for once Unicode acceptance,
tokenization, activation geometry, and downstream behavior are measured separately.

The metric problem..
is not a direct measure of speech understanding. It is a word-level
Levenshtein metric computed after tokenization, normalization, and reference
selection. It becomes scientifically interpretable when four assumptions roughly
hold: word boundaries are stable, spelling conventions are stable, the written word
preserves the acoustic distinctions being tested, and an edit at the word level is a
reasonable proxy for a recognition error. Those assumptions are often acceptable for
high-resource ASR benchmarks with standardized orthography. They are much weaker for
Latin-script Bambara and related Manding ASR.

Latin Bambara weakens the metric in several ways. Tone is often absent even though
tone can distinguish lexical meaning. Digraphs can encode a single sound with
multiple characters, so a one-phoneme error can become a multi-character spelling
error. Conversely, two acoustically different forms can collapse to the same Latin
string when tone or vowel quality is not represented. Word segmentation and spelling
variation then add a second layer of uncertainty: a model can be penalized for
choosing a different written convention rather than for failing to recognize the
speech signal. In this setting, Latin WER mixes acoustic error, orthographic
convention error, tokenization error, and reference-author preference.

CER addresses the same utterance at a different measurement level. Because
explicitly represents vowels, many consonantal contrasts, tone classes, and
nasalization marks, a normalized character edit is closer to a phonemic edit
than a Latin word edit. This is the metric advantage. It should not be overstated:
CER is not automatically phoneme error rate. A scorer must declare whether it
counts Unicode code points, grapheme clusters, base letters, combining tone marks,
spacing, punctuation, digits, and normalization variants. For example, a base letter
plus tone mark can be counted as multiple Unicode scalars unless the scorer defines
a phonemic or grapheme-cluster unit. The paper's claim is therefore conditional:
under a declared normalization policy, CER is a more interpretable proxy for
Manding phonemic accuracy than Latin WER.

This distinction changes how results should be reported. A public ASR result should
not only report ``CER'' or ``WER''; it should report the unit of scoring, the
normalizer, the treatment of tone and combining marks, and the reference inventory.
Without those details, two systems may appear comparable while measuring different
objects. The anchor is valuable because it is a native-script result,
but future work should still strengthen the scorer by publishing a normalized
phonemic-unit evaluation alongside raw character edits.

Orthographic transparency and ASR labels..
The argument extends the orthographic-depth tradition, which treats writing systems
as differing in how directly spelling maps to sound [citation: katz1992orthographic].
In ASR, orthographic depth becomes a label-design problem. A decoder can emit
graphemes, phonemes, subword units, bytes, word pieces, or hybrids; systematic
comparisons show that the choice of label unit remains an empirical and
architectural variable rather than a mere notation change
[citation: zeineldeen2021graphemephoneme]. For high-resource languages, a phoneme
decoder can rely on pronunciation lexica or trained grapheme-to-phoneme systems. For
low-resource Manding ASR, those tools are not guaranteed to exist at the quality
needed for benchmark evaluation. A transparent script partly solves the label
problem by making the ordinary written target closer to the phonemic target.

For CTC, this is not only a linguistic preference. CTC learns a monotonic alignment
between acoustic frames and target labels. If /ny/ is represented as a single
script-level target, the alignment has one label event. If it is represented as
n+y, the alignment has two character events for one phonemic event.
If tone is unmarked, the target omits an acoustic distinction entirely. If a subword
token crosses a phoneme or syllable boundary, the label may become even less
localized in time. These effects do not make Latin unusable, but they make Latin
labels a less direct measurement instrument for the phonemic ASR question.

Standard speech scoring toolkits such as NIST SCTK make WER/CER-style scoring
operational [citation: nist2021sctk], but a scoring toolkit cannot make the reference
units linguistically valid by itself. Low-resource African ASR work on radio archives
demonstrates the practical need for speech technology in this setting
[citation: doumbouya2021]; the present project asks a narrower measurement question:
which script makes the error signal scientifically interpretable for Manding speech?

The Script Advantage: , Latin, and Phonemic Mapping

The script advantage claim requires linguistic precision. is not simply a
font for Bambara, nor is it identical to one spoken language. It is a right-to-left
alphabetic script and written standard used across Manding varieties, including
Bambara/Bamanankan, Maninka, Jula/Dyula, Mandinka, and related written practices
[citation: unicode2006nko,donaldson2017clear]. Spoken Manding varieties remain diverse;
the computational claim is that supplies a more explicit written interface to
the phonology that those varieties share than Latin Bambara orthography usually
does. The word ``advantage'' therefore means an advantage in representation and
measurement, not a claim that a script alone solves acoustic modeling.

At the Unicode layer, occupies U+07C0--U+07FF. That range contains digits,
alphabetic letters, tone marks, nasalization marks, punctuation, and related symbols.
The core vowel inventory is represented by named vowel letters, including U+07CA
NKO LETTER A, U+07CB NKO LETTER EE, U+07CC NKO LETTER I, U+07CD NKO LETTER E,
U+07CE NKO LETTER U, U+07CF NKO LETTER OO, and U+07D0 NKO LETTER O. The consonant
inventory contains letters such as U+07D3 NKO LETTER BA, U+07D5 NKO LETTER TA,
U+07DE NKO LETTER KA, U+07E1 NKO LETTER MA, U+07E2 NKO LETTER NYA, U+07DC
NKO LETTER GBA, and U+07DA NKO LETTER RRA. For ASR, the important property is that
many speech units that are multi-character or inconsistently represented in Latin
orthography have direct script-level targets in .

\caption{Representative script mappings relevant to Manding ASR. The table is not a
complete phonology; it illustrates why CER is closer to a phonemic edit
metric than Latin WER.}

This is the practical source of the bijective argument. In the idealized normalized
setting, the decoder's output alphabet is close to the inventory of acoustic
contrasts: a vowel is a vowel letter, a palatal nasal can be one script unit, a
labial-velar stop can be one script unit, and tone/nasalization can be explicit
combining information. Latin Bambara can still be useful for reading, teaching,
search, and interoperability, but it is not the same measurement object. It often
encodes one sound with multiple characters, leaves tone underrepresented, and
inherits spelling variation from colonial and corpus-specific conventions.

The difference matters for CTC. CTC learns monotonic alignments between time frames
and output symbols. If one phonemic event maps to one normalized target, then
the alignment problem and the error metric are easier to interpret. If the same
event maps to a Latin digraph or to an unmarked tonal word form, then the model can
be penalized for a spelling boundary rather than for an acoustic boundary. This is
why the paper compares CER to Latin WER as measurement systems, not merely as
two display choices for the same transcript.

The script advantage is therefore conditional. It depends on a clean target
column, a declared Unicode normalization policy, explicit treatment of tone and
combining marks, and an evaluation script that counts the intended units. If those
conditions are violated, CER can lose its interpretability. This is why the
data-quality section treats script contamination and normalization drift as central
scientific risks rather than as minor preprocessing details.

Research Questions, Hypotheses, and Claim Boundaries

The initial hypothesis stack above reconstructs the broad research program. This
section narrows that stack into the claims this canonical manuscript can evaluate
with defensible evidence. It is a systems-and-measurement study rather than a single
leaderboard paper. Its research questions therefore distinguish four objects that
are often collapsed in low-resource ASR papers: model-internal script representation,
metric validity, acoustic-decoder architecture, and post-ASR correction governance.
The questions are stated at the level of measurable constructs so that a future audit
can replicate, falsify, or narrow each claim independently.

Caption: Research questions, constructs, and primary observables.

longtable{p{0.08\linewidth}p{0.25\linewidth}p{0.25\linewidth}p{0.25\linewidth}}

Caption: Hypotheses with null or falsifying cases and current evidentiary status.

ID & Alternative hypothesis & Null / falsifying case & Current status

\endfirsthead

ID & Alternative hypothesis & Null / falsifying case & Current status

\endhead
H1 & is internally underrepresented in tested LLM families beyond ordinary
low-resource performance variation. & Tokenizer and activation diagnostics show no
consistent deficit relative to supported scripts, or deficits appear in only
one isolated metric without convergence. & Supported by converging diagnostics in
the written activation papers; scope limited to Qwen/Mistral-family tests and their
protocols.

H2 & Normalized CER is a more phonemically interpretable Manding ASR metric
than Latin WER. & The target normalization map is not bijective over the
claimed inventory, or Latin word scoring preserves the same phonemic edit structure
under the same assumptions. & Supported as a formal measurement proposition, not as
an empirical superiority claim.

H3 & A trajectory-only Transformer CTC decoder can reach the approximate
20\
archived artifact is missing required metadata, has inconsistent hashes or row
counts, or a strict reproduction under identical settings repeatedly fails to enter
the same error regime. & Supported by the preserved archived
checkpoint; the strict May 2026 reproduction did not complete and remains unresolved.

H4 & Trajectory geometry is useful as a dynamic ASR state and as a correction-risk
signal. & Trajectory channels do not improve or explain any archived ASR behavior,
and AGP gates accept unsupported regressions at unacceptable rates in row-level
tests. & Partially supported: historical trajectory evidence and architecture are
defined; AGP has smoke tests but lacks a full benchmark over the anchor test set.

longtable

The primary endpoint for the ASR anchor is character error rate on the 29,060-row
test split, reported with both numerator and denominator. Secondary endpoints are
validation loss, row-level prediction/reference completeness, split integrity,
vocabulary integrity, and artifact hashes. For the LLM studies, the endpoint is not
a single behavioral score but convergence across representation diagnostics. For
AGP, the endpoint is not merely a lower final CER; a valid correction layer must
also report accepted regressions, rejected improvements, abstentions, and
per-partition behavior.

The evidence standard differs by question. The LLM visibility claim is supported by
converging diagnostics across tokenization, activation norms, entropy, sparsity,
kurtosis, and translation behavior. The metric claim is supported by the formal
transparent-script proposition and is conditional on normalization. The ASR anchor
claim is supported by preserved artifact metadata and checkpoint provenance, but the
strict May 2026 audit did not complete and therefore cannot be cited as a new
reproduction. The AGP claim is architectural and experimental-preparatory: AGP has
smoke-test evidence and a row-level contract, but it is not yet a full benchmark
result over the test set.

These boundaries are part of the result. The project contains artifact-backed
evidence, historical evidence, smoke tests, and incomplete audits. Treating them as
one undifferentiated evidence pool would make the paper weaker. Separating them makes
the contribution auditable.

Operational Definitions

This can be stated as a measurement proposition.

proposition: Transparent-script edit preservation.
Let $\Phi$ be a normalized phoneme inventory and let
$f_N:\Phi\rightarrow\Sigma_N$ be a bijective transcription map over that
inventory after the paper's explicit normalization choices. For any two phoneme
strings $\phi_{1:U}$ and $\psi_{1:V}$, edit distance over the strings
$f_N(\phi_{1:U})$ and $f_N(\psi_{1:V})$ preserves the phoneme-level edit structure:
each insertion, deletion, or substitution in $\Phi$ corresponds to exactly one
insertion, deletion, or substitution in $\Sigma_N$, and vice versa. Therefore CER
over normalized strings is an interpretable proxy for phoneme-level edit
rate, modulo explicitly declared treatment of tone, combining marks, punctuation,
and spacing.

Proof sketch..
Because $f_N$ is bijective, it has an inverse $f_N^{-1}$ on $\Sigma_N$. Any edit
script between $\phi_{1:U}$ and $\psi_{1:V}$ maps through $f_N$ to an edit script of
the same length between their renderings. Conversely, any edit script between
the renderings maps through $f_N^{-1}$ to an edit script of the same length
between the phoneme strings. The minimum edit length is therefore preserved. The
equivalence holds only over the normalized inventory named in the proposition; if a
normalizer deletes tone, collapses combining marks, or changes spacing, the metric
must report that policy.

A Latin transcription relation $f_L:\Phi^*\rightarrow\Sigma_L^*$ with
variable-length digraphs, optional tone marking, and spelling variation does not
satisfy the bijective condition. One phoneme may require multiple Latin characters,
one Latin character sequence may correspond to different phonemic analyses, and word
boundaries introduce segmentation choices that are not phoneme boundaries. Latin WER
therefore measures a mixture of recognition, spelling, word segmentation, and
reference-convention agreement. This proposition is a metric-validity claim, not a
guarantee that a particular model must always obtain lower CER than a Latin
model.

Methods and Artifact Protocol

The canonical manuscript uses three classes of evidence. The first class is
published-in-repository paper evidence: the five LaTeX manuscripts in
paper/current/, their figures, and their local result dependencies. The
second class is artifact evidence: checkpoint files, results.json, split
metadata, vocabularies, row-level prediction/reference exports, and SHA-256 hashes.
The third class is project-history evidence: handoff notes, publication-readiness
notes, and session-derived caveats about incomplete audits or missing artifact
chains. Only the first two classes should support numerical claims. The third class
is useful for explaining why claims are bounded.

The ASR training stack for the anchor uses frozen Whisper large-v3 encoder features
of dimension 1280, a learned projection to a 768-dimensional decoder space, stride-4
temporal downsampling, sinusoidal position information, and six Transformer CTC
decoder layers. The trajectory branch computes seven dynamic scalars per timestep and
maps them into attention bias. The N'Ko vocabulary contains 66 output classes in the
project implementation; the Latin comparison vocabulary contains 41 classes. The
anchor model is recorded as a 46.8M-parameter trajectory CTC decoder.

The retained anchor is evaluated on 29,060 test examples from the 290,596-pair
snapshot. The handoff and result metadata record 216,225 character edits over
1,050,967 reference characters, yielding 20.57\
future work is stricter than the old papers: any public benchmark row should include
the run name, script, mode, data snapshot, split hash, vocabulary hash, model
checkpoint hash, learning rate, batch size, seed, prediction row count, reference row
count, and CER numerator/denominator. This is the standard needed to prevent another
training-parameter mismatch from being mistaken for a scientific contradiction.

Caption: Evidence levels used throughout the canonical paper. This table is the guardrail that keeps the paper from mixing archived benchmarks, incomplete audits, and hypothesis-generating experiments into one unsupported headline.

The same discipline applies to the phrase ``canonical.'' A canonical result is not
the most exciting number in the project history; it is the number that can be stated
without forcing the reader to trust conversational memory. For this manuscript, that
means the checkpoint is canonical as an archived anchor, while the
low-learning-rate matrix is canonical as a non-comparable artifact-complete
engineering bundle. The historical eight-way comparison is important but secondary,
because its role is to explain why the trajectory hypothesis became worth testing.

The scoring protocol is intentionally simple. References and hypotheses are
normalized by the project text normalizer, scored with Levenshtein edits, and
reported as edits divided by reference characters. For a paper-ready ASR run, the
same normalization function must be recorded alongside the result. Without that,
even a correct CER number is not a complete scientific object: a different
normalizer can change the denominator, remove or retain combining marks, collapse
spaces, or hide script contamination. This is especially important for because
the value of the metric depends on preserving the script's phonemic information, not
merely rendering a string in Unicode.

The May 2026 audit changed the publication protocol even though it failed to
produce a replacement benchmark. It showed that paid training should never be
launched before three checks pass locally and remotely: first, the pair file and
feature cache must have expected hashes and counts; second, the trainer must be
compiled and pinned by hash; third, a sanity run must emit the same artifact
contract expected from the full run. The audit also showed why a result table should
include the learning rate in the table itself. The later 31\
expected to reproduce the anchor because they used a different
optimization regime.

Data Quality and Normalization Caveats

The project should treat data cleaning as a research result, not as a hidden
preprocessing footnote. The April project notes record a label-contamination
discovery in the N'Ko column: Latin and IPA-like characters appeared inside rows that
were supposed to be script-native labels. Those notes attribute a large CER
change to the cleaning function, but the exact contamination count and improvement
should not be promoted as a public benchmark until the script, manifest, and before
and after hashes are packaged. The publishable claim is narrower: low-resource ASR
datasets can contain script contamination severe enough to change model conclusions,
so every run in this project must report the label-cleaning function used.

The strict audit also exposed an implementation-level data issue before training
could complete. Some Whisper feature tensors were two-dimensional time-by-feature
matrices, while others carried an extra leading singleton dimension. The loader had
to normalize those tensors before the sanity gate would pass. That failure is not a
scientific result, but it is a reproducibility lesson: for a corpus this large, the
artifact contract must validate tensor shape, feature count, pair-file hash, split
size, vocabulary hash, and prediction/reference row counts before paid training is
trusted.

There are three distinct data risks in the project. The first is script
contamination: a row claims to be but contains Latin, IPA-like, or otherwise
non-target characters. The second is normalization drift: two runs use different
rules for spaces, punctuation, Unicode combining marks, or target filtering. The
third is feature-label mismatch: the acoustic tensor exists but no longer matches
the row identity, duration, or expected tensor shape. These risks are operationally
different. Script contamination corrupts the target alphabet; normalization drift
changes the metric; feature-label mismatch corrupts the input-output pairing.

Caption: Data-quality risks and the corresponding publishable mitigation.

This is why the paper should not hide the failed audit. The failed audit is part of
the scientific record because it explains why the project moved from ``run more
benchmarks'' to ``freeze a canonical artifact and document the evidence ladder.''
That move is not a retreat from rigor. It is the condition for rigor: the paper
becomes more publishable when it tells readers exactly which results are artifacts,
which are historical, and which are still planned.

Script Invisibility in LLMs

The first layer of the project is diagnostic. A model can accept characters
without possessing circuits that make those characters computationally useful. The
written LLM papers therefore avoid evaluating only final translation quality. They
measure what happens inside the network.

Translation tax..
The activation papers define a translation or representation tax as the excess
internal effort required to process relative to a better-supported script. In
one protocol, Qwen3-8B exhibited an average tax of 2.94x. In the cross-model protocol,
the corresponding tax estimates were 3.30x for Qwen3-8B, 3.59x for Qwen2.5-7B, and
2.67x for Mistral-7B. These values should not be merged as if they came from one
identical measurement pipeline. Their importance is directional: each protocol finds
that input induces a substantially less efficient internal trajectory than
well-supported text.

Hidden-state geometry..
The brain-scan work measures layerwise activation norms, entropy, sparsity, and
kurtosis. The Paper 1 diagnostics reported a 1.2--1.7 bit entropy gap, approximately
2.2x higher embedding sparsity for , and a 78.1\
The cross-model paper reported the same qualitative pattern across the tested
families, with activations roughly 66--72\
spanning 64.6--93.5\
pooled statistic. It is the repeated geometry: enters the model as Unicode but
does not propagate as a well-supported internal representation.

[figure: ../../figures/brain_scan_l2_comparison.png]

Caption: Layerwise activation-norm comparison from the brain-scan experiments. The figure is best read with the entropy, sparsity, kurtosis, and tokenizer analyses in the underlying papers. Its role here is to illustrate the structural nature of script invisibility rather than to serve as the only evidence.

Tokenizer evidence..
The tokenizer evidence explains one mechanism. Arabic, another right-to-left script,
has thousands of vocabulary entries in major multilingual tokenizers, while is
often represented by fallback behavior or tiny coverage. This isolates the issue
from script direction alone. The problem is data starvation and representation
absence, not merely right-to-left rendering.

Why this matters for ASR..
The LLM findings motivate the speech work. If generic language models have weak
internal structure for , then a low-resource ASR system cannot rely on a
generic multilingual language model to supply script competence after decoding. It
needs script-native targets, script-native normalization, and metrics that preserve
the acoustic distinctions the script was designed to encode.

The diagnosis can be summarized as a three-zone failure model. In the first zone,
the model fails before semantics because the tokenizer fragments or backs off on the
script. In the second zone, the model passes Unicode strings forward but lacks
high-energy internal directions for them, producing weak activation norms,
high-entropy representations, and sparse or brittle hidden states. In the third
zone, the model can produce fluent output in a dominant bridge language, but the
bridge destroys the script-specific distinctions that matter for the target
language. The speech work is designed to bypass all three zones: it does not ask a
general LLM to transliterate or repair Bambara after the fact; it trains a decoder
to emit directly from acoustic evidence.

Caption: Failure zones exposed by the LLM papers and their ASR design response.

The LoRA remediation experiments in the representation papers should be read in this
frame. They do not prove that a small adapter solves for every downstream
task. They show that the deficit is partly trainable: when the model receives
script-aware adaptation, some activation and output behavior moves in the right
direction. That makes the invisibility result more actionable. It is not an
argument that is intrinsically impossible for large models; it is an argument
that current general-purpose pretraining leaves measurable structural holes.

Script-Native Speech Recognition

The speech papers record an engineering progression. The early BiLSTM CTC system was
small and functional but weak, with approximately 56\
Transformer CTC system used frozen Whisper features, a 46.9M-parameter decoder, six
Transformer layers, hidden size 768, and stride-4 temporal downsampling; it reached
approximately 33\
loss behavior, with validation loss moving from 0.884 to 0.290 and a 29.4\
50-sample held-out evaluation, but the WER improvement was not statistically strong
in the recorded paired comparison. This sequence matters because it prevents a false
interpretation: ASR did not appear fully formed. It emerged through
architecture, normalization, data-cleaning, and metric work.

Caption: Condensed progression of the script-native ASR stack. The values summarize the written paper artifacts and are not all from the same evaluation regime.

The corpus associated with the canonical ASR anchor contains paired
examples split into 232,476 training examples, 29,060 validation examples, and 29,060
test examples. The target is character transcription. The model uses frozen
Whisper encoder features and a trainable Transformer CTC decoder whose trajectory
variant modulates attention using learned dynamic state.

The bridge systems in the earlier speech papers are useful because they show what
the canonical system deliberately avoids. A Latin bridge can appear attractive:
there are more existing tools, more pretrained assumptions, and more familiar
evaluation conventions. But bridge decoding introduces its own error classes. Latin
Bambara can represent a single sound with a digraph; tone may be absent or
inconsistently marked; conversion rules can miss phonemes such as /g/ or confuse
symbols borrowed from IPA-like transcriptions; and right-to-left rendering problems
can hide target-side defects in visual inspection. A bridge result therefore mixes
speech recognition error, transliteration error, orthographic-convention error, and
display error. Direct CTC reduces the number of moving parts.

This does not mean every number is automatically better than every Latin
number. Latin can sometimes win under a smaller output vocabulary or a cleaner
training condition, as the historical baseline table shows. The claim is that
makes the error easier to interpret. If a direct decoder substitutes
one character for another after normalization, the error is usually closer to a
phonemic confusion than a Latin word-level error is. That interpretability is what
makes the anchor worth discussing even while the strict reproduction
remains incomplete.

The implementation path also explains why the paper should not collapse V1, V3, V4,
and Paper 4 into one performance curve. V1 tested feasibility with a small CTC
system. V3 tested whether frozen Whisper features could support a larger
script-native decoder. V4 tested adaptation and confidence behavior. The Paper 4
trajectory model tested whether dynamic state could be injected into the decoder.
Those systems answer different engineering questions. The canonical paper uses them
as a development history, not as a single leaderboard.

The Anchor

The result that can be discussed publicly is precise. The local Paper 4 reproduction
cache preserves an archived trajectory ASR checkpoint reporting
test CER. Its metadata records script , mode trajectory,
trajectory-only decoding, learning rate 0.0003, batch size 32, dropout 0.1, seed 42,
best validation loss 0.6358872798606507, and 47 trained epochs.

longtable{p{0.34\linewidth}p{0.54\linewidth}}

Caption: Canonical archived ASR anchor.

Field & Value

\endfirsthead

Field & Value

\endhead
Artifact root &

Script and mode & trajectory CTC

Training branch & Trajectory-only; no residual trajectory-attention branch and no test-time adaptation branch

Learning rate, batch size, dropout & 0.0003, 32, 0.1

Seed & 42

Train / validation / test & 232,476 / 29,060 / 29,060

Best validation loss & 0.6358872798606507

Epochs trained & 47

Reported test &

Results SHA-256 & 252aecd6e323f7d50cefd3c1e507ddaf035d9f0ac4f78d67766c4cf6ed5d24a7

Vocabulary SHA-256 & e3ab620c9d2f971603d76f953f2be40bf9283dfd99d6428c7d51a9a73246ea67

Best checkpoint SHA-256 & ab1fe47f96c2c434d8f301ae065b3292d592b9a4f5accf1d09acc97ca2c03b59

longtable

The project therefore has a canonical artifact for the ``20 CER'' discussion: an
archived checkpoint reporting CER. The later training that
consumed May 2026 compute did not finish a strict reproduction that can replace that
artifact. The anchor remains usable, but it must be described as an archived
checkpoint result, not as a newly reproduced result.

The reason the later runs around 31\
replications is also clear. They were not run under the same parameter regime. The
preserved low-learning-rate bundle used
$1\times10^{-4}$ while the anchor used $3\times10^{-4}$, and the strict audit that was
supposed to resolve the mismatch did not complete. That is a training-parameter
caveat, not a disproof of the archived checkpoint.

The archived result also has a narrow architectural identity. It is not the later
TAR branch, not the residual trajectory-attention variant, not the trajectory plus
test-time-training branch, and not AGP. The result belongs to a trajectory-only CTC
decoder using the recorded learning rate, batch size, dropout, seed, split, and
vocabulary. In public writing, that specificity prevents two opposite mistakes: it
prevents overstating the newer experiments as if they inherited the anchor, and it
prevents dismissing the anchor because later experiments used different settings.

The unfinished strict audit should be described as a failed replacement attempt, not
as a failed anchor. The audit was launched to answer the correct question: can the
20.57\
with row-level artifacts? It did not produce the required final
results.json, prediction file, reference file, and partition metrics. That
means the old artifact remains the best retained evidence. It does not mean the old
artifact should be erased; it means the paper must label it accurately.

Historical Evidence That Motivated the Trajectory Hypothesis

The project history contains a stronger comparative story than the canonical anchor
alone, but it must be labeled correctly. Earlier internal 297K-scale runs compared
N'Ko and Latin across baseline, graph, trajectory, and combined decoder conditions.
Those results motivated the phrase ``bijection amplifier'': dynamic or structural
decoder mechanisms seemed to help more than Latin because characters
are closer to acoustic units. However, the complete local artifact chain for all
eight historical runs is not currently restored, so these numbers should appear as
hypothesis-generating evidence rather than as the primary benchmark.

Caption: Historical eight-way comparison from internal logs. These values explain the origin of the trajectory hypothesis, but they are not the canonical benchmark because the full artifact bundle is not locally restored.

The compositional-generalization results are also important because they test a
different prediction: a phonemically transparent script should degrade more gently on
unseen words, since novel words recombine known sound-symbol units. In the preserved
project notes, an experiment trained on seen vocabulary and evaluated on utterances
containing unseen words. N'Ko degraded from 32.75\
degraded from 31.43\
gap was 37.81pp and the Latin gap was 41.46pp, a 3.65pp smaller gap for N'Ko.

\caption{Historical compositional-generalization evidence. These values support the
hypothesis that composes unseen words from known phonemic units, but they
remain separate from the anchor.}

These historical experiments should not be discarded. They explain why the paper
models trajectory geometry, compositionality, and metric validity rather than only
reporting one checkpoint. Their limitation is provenance, not conceptual value. The
manuscript therefore presents them as prior internal evidence that motivated the
strict audit, while keeping the 20.57\
artifact-backed ASR result.

The vocabulary-expansion experiment adds a second robustness signal. In the
SEEN-only control, Latin was slightly better on seen words: 15.05\
16.09\
sharply, but degraded less: 53.90\
both scripts recovered almost the same amount of the unseen-word gap, 13.75pp for
and 13.78pp for Latin. The remaining unseen-word CER was still lower for
: 40.15\
claim. Full-data training helps both scripts, but retains a 2.58pp advantage
on rare-word utterances and a 3.62pp smaller residual gap in the recorded
experiment.

\caption{Historical vocabulary-expansion evidence. The recovery from SEEN-only to
full-data training is nearly identical, but the residual unseen-word gap remains
smaller for .}

[figure: ../../figures/fig5_expF_compositional.pdf]

\caption{Historical Exp F compositional-generalization figure. The figure should
be read as context for the trajectory and metric hypotheses, not as a replacement
for the archived anchor.}

[figure: ../../figures/fig6_expH_vocab_expansion.pdf]

\caption{Historical Exp H vocabulary-expansion figure. Full-data training helps
both scripts, but the recorded unseen-word residual remains smaller for .}

The reason these tables belong in the canonical manuscript is not that they close
the proof. They do not. They show that the research program has a coherent empirical
shape. The anchor establishes that a script-native trajectory decoder can reach the
20\
key architectural idea. The compositional and vocabulary tables explain why script
choice matters beyond in-distribution test CER. Together they justify a future
matched audit without pretending that the audit has already succeeded.

Trajectory Geometry

Trajectory conditioning is the architectural bridge between acoustic decoding and the
larger AGP program. Speech unfolds through time; phonetic evidence does not arrive as
isolated frames. A decoder benefits when it can distinguish stable regions, imminent
boundaries, recovery zones, uncertainty, and novelty.

This mechanism is especially plausible for because the output script preserves
more acoustic structure than Latin Bambara spelling. Trajectory geometry is a model
of where the speech signal is moving; it is most useful when the target symbols
retain the movement-relevant distinctions.

Caption: Interpretation of the seven trajectory channels in the canonical speech-dynamic framing. Earlier notes sometimes used motion-inspired names; this paper formalizes the channels as ASR state variables.

This table also resolves an ambiguity in the project history. Some earlier notes
described the scalars through a handwriting or pen-stroke analogy: velocity,
curvature, acceleration, and related motion terms. That analogy was useful for
intuition, but it is not the publishable explanation of the ASR model. In this paper,
trajectory means dynamic acoustic state. The relevant movement is the movement of
evidence through time: toward a symbol, across a boundary, through uncertainty, or
into a domain region the model has not seen before.

The script dependence follows from the target representation. If an acoustic
boundary corresponds cleanly to a character boundary, then transition pressure
is directly useful. If the same acoustic transition is encoded through a Latin
digraph, optional tone mark, or spelling convention, the decoder receives a less
direct target. This does not imply that Latin cannot be modeled. It implies that the
trajectory signal has a cleaner supervised target in .

What the Conservative Ablations Say

The later same-snapshot low-learning-rate matrix is valuable, but it is not a
strict comparison to the anchor. It preserves five artifact-complete
runs with 29,060-row prediction/reference contracts, all under learning rate
0.0001. The results were baseline 31.38\
residual variant at 31.69\
at 31.12\

\caption{Preserved low-learning-rate matrix. These runs are useful for engineering
diagnosis and AGP packet construction, but they are not directly comparable to the
anchor because the anchor used learning rate 0.0003.}

The correct conclusion is modest. The low-learning-rate results show that the
training and artifact pipeline can produce complete row-level outputs across
conditions, and they reinforce that optimization settings dominate if they are not
held fixed. They do not validate or invalidate the result. They also do
not prove that the residual-attention branch or the test-time adaptation branch
improves ASR. In fact, the later proposal notes reinterpret the residual-attention
branch as likely a negative result: trajectory scalars themselves appear more
important than adding deeper residual trajectory attention.

The abbreviation problem also needs to be corrected for outside readers. TAR in the
internal run names referred to a trajectory-attention residual branch: a heavier
decoder variant that tries to inject trajectory state deeper into attention. TTT
referred to test-time training or adaptation: an inference-time procedure intended
to update a small part of the model on new speaker or domain evidence. Neither term
should appear in a public abstract without expansion, and neither should be
identified with the archived checkpoint. The anchor was the simpler
trajectory-only CTC path.

The low-learning-rate matrix is still scientifically useful because it is a negative
control on the project's own enthusiasm. If a heavier trajectory branch under a
different learning rate does not beat a simpler baseline, then the paper should not
claim that more geometry automatically helps. The better interpretation is
mechanistic: compact trajectory state may be valuable when it biases the decoder in
the right place, while additional residual branches can overconstrain or
miscalibrate the attention stack. That is a real finding, even though it is not the
headline result.

The matrix also shows why future comparisons must be matched at the launch-script
level. A fair comparison requires identical corpus snapshot, pair hash, feature
cache, split, vocabulary construction, optimizer, learning rate, batch size,
dropout, patience, seed, epoch budget, early-stopping rule, and scoring normalizer.
Changing any one of these can be defensible engineering; changing it silently makes
the run unusable for the claim ``model A beats model B.'' The canonical paper
therefore treats the low-learning-rate runs as artifact-complete ablations under
their own regime, not as failed reproductions of the anchor.

[figure: ../../figures/fig1_cer_comparison.pdf]

Caption: CER comparison figure from the Paper 4 line of work. Public versions should separate archived anchor evidence, historical trajectory evidence, and non-comparable low-learning-rate evidence.

AGP: Anticipation Geometry Partition

AGP is the post-ASR Anticipation Geometry Partition. It should not be described as
the model that produced the result. The acoustic pipeline is audio to
Whisper features to trajectory CTC to raw transcript. AGP begins after that:
it consumes the raw transcript, row-level metadata, and trajectory-derived signals,
then partitions spans into states that govern whether correction is allowed.

Caption: Operational AGP states. AGP turns trajectory geometry into a correction policy rather than an unconstrained text rewrite.

This distinction matters because a language-model correction layer can easily improve
surface fluency while corrupting speech evidence. AGP exists to prevent that failure.
Its goal is conservative admissibility: accept local repairs that reduce error,
reject unsupported rewrites, and expose uncertainty rather than hiding it in a fluent
sentence.

The current AGP smoke tests are preliminary but informative. A hand smoke test moved
CER from 14.29\
edits. A synthetic stress test moved 13.33\
neutral accepted cases. An archived low-CER real slice moved 76.04\
one improved accepted case and no worse accepted cases. These numbers are evidence
that the correction gate can be conservative; they are not a benchmark claim over the
test set.

The formal AGP unit is a row, not a sentence-level anecdote. A row contains the raw
ASR hypothesis, the reference when available, edit counts, character denominator,
trajectory summaries, a partition label, candidate correction spans, admissibility
decision, and final text. When retrieval or provenance is available, the row also
stores sources used, transliteration variants, normalized forms, and a provenance
score. For deployment, the same row carries TTS eligibility fields: overlap risk,
music risk, speaker cleanliness, and exclusion reason. This schema matters because
it makes AGP auditable. A reviewer can inspect not only whether CER changed, but
which edits were accepted, rejected, and why.

Caption: AGP row contract. The row-level design prevents a correction layer from silently rewriting transcripts without evidence.

The admissibility policy is deliberately asymmetric. A correction is cheap to
propose but expensive to trust. AGP should accept a change only when the local
evidence supports it and the partition permits it. Stable spans usually need no
change. Boundary spans may admit small repairs because boundaries are precisely
where CTC errors often occur. Uncertain spans should often abstain unless multiple
signals converge. Novelty spans should be treated as data-discovery events rather
than as language-model editing opportunities. In other words, AGP is not a
post-hoc beautifier; it is a risk model for transcript modification.

A full AGP benchmark should therefore report more than final CER. It should report
the number of proposed edits, accepted edits, rejected edits, abstentions, accepted
improvements, accepted neutral edits, accepted regressions, rejected improvements,
and per-partition CER deltas. The most important safety metric is accepted
regressions. A correction system that sometimes improves CER but frequently accepts
worse edits is not acceptable for building a corpus or a TTS training set.

Deployment and Row-Level Artifacts

Scalar CER is insufficient for the next stage of the research. AGP needs rows:
feature identifier, split, script, mode, raw ASR hypothesis, reference text, edit
counts, reference character counts, trajectory scalars, partition labels, correction
proposal, admissibility decision, and final text. The later packet-corpus work
created that structure for the preserved low-learning-rate runs, with 29,060
prediction rows and 29,060 reference rows per available run.

The deployment papers also preserve Djoko and consensus-pipeline artifacts:
8,985 segments, 8,985 transcriptions, 269 consensus rows, 6,625 speaker-labeled rows,
258 episodes, seven speakers, and five eligible speakers. Those data matter because
the eventual system is not just an ASR benchmark. It must operate in real domains
with speaker variation, episode structure, visual context, and disagreement among
transcribers. They should be presented as deployment substrate, not as final
CER-bearing proof.

The broader Djoko extraction effort was larger than the first consensus slice. The
handoff records 1,124 downloaded videos out of a 2,001-video YouTube channel and
32,826 thirty-second audio segments. It also records a severe domain gap: the first
batch produced only about 8.8\
was applied to soap-opera audio. That number should be taken seriously. It means the
archived checkpoint is a within-distribution ASR anchor, not a
production model for conversational broadcast media.

The reason the Djoko material still belongs in the paper is that it defines the next
research interface. The source has no reliable burned-in subtitles, so OCR is not a
solution. The project instead uses audio segmentation, ASR hypotheses, Latin bridge
hypotheses where useful, visual scene analysis, speaker clustering, consensus
filtering, and row-level provenance. In that environment, AGP is not an optional
decoration. It is the only safe way to separate stable transcript rows from rows
that are uncertain, novel, noisy, overlapping, musical, or speaker-ambiguous.

[figure: ../../figures/fig7_djoko_quality.pdf]

\caption{Djoko quality-control figure from the deployment paper. The figure
illustrates the deployment filtering problem; it is not presented as a final
ground-truth CER benchmark for the checkpoint.}

The deployment protocol also clarifies the difference between corpus building and
model scoring. For model scoring, every row needs a reference. For corpus building,
many rows do not yet have a trusted reference, so the system must preserve
uncertainty, provenance, and exclusion decisions. A stable AGP row can be used for
search or candidate review. A boundary row can become a correction-training example.
An uncertain row should remain visible but not silently enter the training corpus. A
novelty row can point to new vocabulary, new speakers, or domain drift. This is how
the research can continue without pretending that out-of-domain ASR is already
solved.

The Public Narrative Around

The project can talk about CER, but only in a disciplined way. The
publishable sentence is:

quote
An archived trajectory ASR checkpoint in the repository reports
test CER after training on 290,596 Bambara speech pairs under recorded settings.
quote

That sentence is strong because it is exact. It avoids the mistakes that would weaken
the work: claiming the result was reproduced in May 2026, claiming matched
Latin-vs- superiority from non-matched runs, attributing the anchor to a later
decoder variant, saying AGP improved the benchmark, or implying production readiness.

The surrounding narrative should be that is an anchor for a larger
scientific case. The LLM papers show why is invisible to general models. The
speech papers show that direct audio-to- decoding is technically real. The
metric paper argues that CER is a more meaningful target than Latin WER for
Manding ASR. The trajectory work gives a dynamic decoding mechanism. AGP gives the
correction and governance layer needed before deployment. This is a coherent research
program even if the final strict reproduction must wait.

The public abstract should therefore avoid the structure ``we achieved 20.57\
therefore N'Ko beats Latin.'' That sentence invites the wrong scrutiny because the
matched comparison is not closed. A stronger abstract begins with the infrastructure
problem: widely used AI systems can encode characters while still lacking
usable internal structure for them; Latin-script metrics then hide the problem by
scoring orthographic convention rather than script-native phonemic evidence. The
result enters as evidence that a direct ASR path is technically
viable enough to make the measurement problem concrete.

For a model card, the wording should be even narrower. It should name the artifact
as an archived checkpoint, report the exact split and learning rate, state that the
strict May 2026 reproduction did not complete, and warn that the checkpoint is not a
production Djoko or conversational-broadcast model. For a project page, the broader
narrative can be stronger: reveals a three-part failure in modern AI systems,
covering representation, decoding, and metrics. For a research paper, the value is
the synthesis: the project connects mechanistic LLM diagnostics with script-native
ASR and a conservative correction framework.

The phrase ``20 CER'' should be treated as a handle, not as a final conclusion. It
is useful because it gives readers a memorable technical anchor. It is dangerous if
it becomes the entire story. The paper's real contribution is that it explains why a
20.57\
why a language technology stack for Manding should be evaluated in the script that
preserves the relevant linguistic structure.

The Reconstruction and Tone Pillar

The four levers introduced above---representation, alignment, measurement, and
governance---describe how a script shapes recognition, the path from signal to
symbol. A fifth lever closes the loop in the other direction. The same featural
codebook that makes recognition interpretable also makes reconstruction
expressible: tone, and ultimately sound, can be written in a form that is editable,
auditable, and phonemically explicit. This section states the reconstruction pillar,
gives the corrected tone-mark inventory it depends on, and fixes the unifying
architecture that ties decoding, governance, and correction into one object. It is
deliberately a framing-and-protocol section. The empirical load is carried by two
companion papers, and every number below is reported with its provenance.

Tone as prior times evidence

Tone is the largest open hedge in the recognition stack, and the limitations section
below still lists it as unfinished. The contribution here is structural: tone
resolution is not one model but the product of two independent sources. A linguistic
prior asks which tone is plausible given the surrounding text; an acoustic
evidence channel asks which pitch the speaker actually produced. Writing tone
resolution as prior times evidence makes the division of labor explicit and assigns
each half to a companion paper. The prior is a contextual tone model trained on
OCR-extracted toned , which learns tone from text. The evidence is Featural
Acoustic Coding (FAC), which reads tone from acoustic $F_0$ and, more broadly, treats
the syllable codebook as a featural sound code in which the script's tone,
length, vowel, and onset structure already carry most of the descriptors a sound
needs. Neither half is the acoustic model that produced the anchor;
both act after decoding, inside the governance gate described next.

The corrected tone-mark inventory

The reconstruction argument depends on stating the script's tone system correctly.
encodes tone with seven combining marks in the range U+07EB through U+07F1:
short and long forms of high, low, and rising tone, together with a long descending
(falling) mark at U+07EE. All four tone shapes---high, low, rising, and
falling---are therefore native to the script; length is an orthogonal axis carried by
the short/long distinction. This matters because an earlier internal annotation of
the syllable codebook had mislabeled U+07EE and U+07EF as generic length marks and
listed only five tone marks, which would have implied that falling tone was absent
and had to be added by a designed extension. That is incorrect. The script already
expresses all four tone shapes; any designed featural extension contemplated for FAC
is therefore purely timbral---harmonicity, spectral spread, dynamics---and
never tonal. Correcting this inventory removes a factual error that had propagated
from the codebook into draft figures and keeps the reconstruction claim honest.

An empirical tone prior, reported as provisional

On a corpus of toned harvested by vision-language OCR from lesson videos,
tone is overwhelmingly a register phenomenon: roughly 75\
level high or low and only about 1\
remainder unmarked. The practical consequence is that acoustic tone resolution for
is, to first order, a high/low register classification problem rather than a
contour-tracking problem, which is exactly the regime a register-relative classifier
handles well. A text-only baseline that predicts tone from a bigram language model
leaves roughly half of tone marks wrong, which is the bar that adding acoustic
evidence must beat. These numbers are reported as provisional. Independent
OCR passes by two strong vision-language models agree on only about 30 to 50\
tone classes, so the corpus is adequate for estimating gross distributional facts but
is not yet trustworthy tone ground truth. The decisive measurement requires read
speech of known tone-marked text, where the gold tone is authored rather than
inferred; that experiment is specified but not yet run, and the limitations section
records it as an open gate alongside the unfinished anchor audit.

One geometry, three levels: decode, govern, correct

The reconstruction pillar does not introduce a separate machine. It reuses the
trajectory geometry $z_t$ that already drives decoding. The same state that biases
CTC emission (decode) also defines the AGP partitions that decide whether a span may
be edited (govern), and tone correction is simply the admissible edit that the
correct stage applies under that gate (correct). Figure~[ref: fig:three-levels] shows
the unifying view, including the self-improving loop in which better tone resolution
produces cleaner labels and a lower-error retrain. The figure also records the two
directions of travel on the codebook. The analysis arrow runs from signal to
and is the recognition stack of this paper. The synthesis arrow runs
from back to a signal: FAC's decoder demonstrates it for sound, and the same
-as-score idea drives a separate generator for body motion in a computational
choreography line. The notation is shared; the generators are distinct builds, and
this paper does not claim that the recognition front-end is swapped to produce motion.

N'KO AS COMPUTATIONAL INFRASTRUCTURE
  one featural codebook . one trajectory geometry . many signals

 signal (mic / mocap) --> front-end features --> N'Ko featural codebook
                                                 (3,024 tonal syllables)
 +-------------------------------------------------------------------+
 | 1. DECODE   Anticipatory Transformer + CTC                        |
 |    features --> z_t (7-dim) --> TONELESS N'Ko  (~20% CER anchor)   |
 +---------------------------------+---------------------------------+
                                   | toneless inscription
 +---------------------------------v---------------------------------+
 | 2. GOVERN (AGP)  same z_t reused as a correction policy           |
 |    partition spans: stable / boundary / uncertain / novelty       |
 |    trust gate: accept / abstain / review                          |
 +---------------------------------+---------------------------------+
                                   | eligible spans
 +---------------------------------v---------------------------------+
 | 3. CORRECT  conservative, AGP-gated                               |
 |    tone = text prior (contextual)  x  acoustic register (FAC)     |
 |    --> TONED N'Ko                                                  |
 +-------------------------------------------------------------------+

 SELF-IMPROVING LOOP
   OCR lessons --> VLM --> toned corpus --> AGP partition --> stable pairs
        ^                                                          |
        +---------------- retrain <-- lower CER <------------------+

\caption{One geometry, three levels. The trajectory state $z_t$ is reused to decode,
to govern correction, and to resolve tone. The codebook carries two arrows:
analysis (signal to , the stack of this paper) and synthesis
( to signal), the latter proved for sound by the FAC decoder and reused as a
score for body motion in a separate generator.}

Program structure and companion papers

This manuscript is the umbrella and the measurement spine of a larger program. The
companions carry the empirical and engineering load and are written for their own
venues, so they are cited rather than absorbed. The representation companions
(script invisibility in LLM activations and its structural persistence across model
families) establish that generic models underrepresent . The recognition
companions develop the script-advantage CTC result, speaker adaptation and
generalization, and trajectory attention residuals. The reconstruction
companions are the two halves of tone described above: the contextual tone model and
FAC. A further governance companion, a Mixture of Anticipatory Orthogonal
Experts for (MAOE-), generalizes the AGP gate into a routing
architecture whose experts are orthogonal in authority rather than in capacity, with
a deterministic admissibility layer deciding what each partition is permitted to do.
Finally, an applied protocol direction reframes the entire stack as a market
for linguistic computation, in which the work of validating tone, transcription, and
cultural use of is performed by people whose competence no GPU can replace.
The science is the pillars; the protocol is one way the resulting capability can fund
the data that the pillars need. A systems track (neural-engine offload, distributed
on-device training, and quantized retrieval) is named separately because it explains
how the program was trained cheaply, not what it claims.

Limitations

The principal limitation is that the strict May 2026 anchor audit did not complete.
The project rented compute, hydrated the full feature cache, fixed a mixed tensor
shape bug, passed the sanity gate, and launched strict audit processes, but the audit
did not produce the verified final result artifacts needed to replace the archived
checkpoint.

The second limitation is comparability. The low-learning-rate matrix and several
historical trajectory comparisons are useful, but they cannot be collapsed into one
headline table. Parameter mismatch, artifact-chain incompleteness, and dataset
cleaning differences must be stated. In particular, the label-contamination discovery
from the project history is scientifically important: low-resource ASR datasets can
contain script contamination that changes CER by large margins if not cleaned
consistently.

The third limitation is metric precision. CER is more phonemically
interpretable than Latin WER, but it is not a perfect phoneme error rate. A stronger
future paper should formalize the mapping from grapheme sequences to phonemic
units, including tone, nasalization, punctuation, normalization, and boundary marks.
The reconstruction pillar above begins this formalization for tone, with a corrected
seven-mark inventory and a prior-times-evidence decomposition. What remains is a
decisive measurement on read speech with authored tone-marked gold, because the
current OCR-derived tone corpus is only provisional ground truth.

The fourth limitation is AGP evaluation. AGP has a coherent architecture and early
smoke-test behavior, but it still needs a full row-level benchmark with accepted,
rejected, improved, neutral, and worsened edits reported separately.

The fifth limitation is corpus representativeness. The -pair anchor is a
large project snapshot, but it is not a complete sample of all Manding speech. It is
closer to clean read-speech and curated paired data than to conversational,
multi-speaker, noisy, dramatic, or music-adjacent audio. The Djoko extraction work
exists precisely because the real deployment distribution is harder. Public claims
should therefore separate within-distribution ASR from out-of-domain transcription.

The sixth limitation is community and orthographic authority. is not only a
technical alphabet; it is a living script with scholarly, pedagogical, religious,
and community practices. A model can output Unicode and still violate
community expectations about spelling, tone, punctuation, or register. Publication
should invite review from readers who know Manding language practice, not only from
machine-learning reviewers.

The seventh limitation is the absence of a finished matched Latin-vs- anchor
under the same hyperparameters as the result. Historical comparisons
and low-learning-rate ablations are informative, but the cleanest future test would
train N'Ko and Latin decoders under a single frozen protocol with identical data
snapshot, optimizer, learning rate, seed schedule, patience, output artifacts, and
scoring code. Until then, the paper should defend CER as a better metric and
the archived checkpoint as a strong script-native artifact, not as the
final matched superiority proof.

The eighth limitation is compute closure. The project reached the point where more
paid GPU training was no longer justified without first tightening artifacts,
manifests, and publication language. That is a practical limitation, but it also
improves the research story. The paper can close the current phase by saying: this
is what we know, this is what remains unresolved, and this is the exact protocol a
future funded audit must run.

Conclusion

should be treated as computational infrastructure for Manding speech systems.
It affects what LLMs can see, what ASR decoders learn, what error metrics mean, and
how correction should be constrained. The project does have a legitimate way to talk
about the 20\
test CER under recorded settings on the -pair snapshot. That
claim is publishable when presented as an artifact-backed anchor rather than as a
freshly reproduced leaderboard result.

The broader paper should argue for a research program, not a single unqualified
number. Script invisibility explains why generic models fail . Script-native
ASR shows that direct transcription can work. Trajectory geometry explains why
dynamic speech state belongs in the decoder. AGP defines a conservative path from raw
ASR output to trustworthy correction. Tone reconstruction extends that same path to
the script's last unresolved axis, resolving tone from a text prior and acoustic
evidence under the same governance gate. The next phase is not to invent a new story;
it is to preserve this one with stricter reproduction, cleaner row-level artifacts,
matched comparisons, and the decisive read-speech tone measurement when funding
returns.

The consolidated conclusion is therefore methodological. The project began with the
intuition that might help ASR because it is phonemically transparent. The
stronger conclusion is that script affects the whole measurement stack. Before ASR,
script affects tokenization and internal model geometry. During ASR, script affects
the output labels that CTC learns to align with speech. After ASR, script affects
whether CER, WER, and correction decisions correspond to speech evidence or to
orthographic convention. A system that ignores script is not neutral; it chooses the
infrastructure of the dominant representation and then calls the result language
technology.

The anchor gives the paper a concrete center. It shows that direct
ASR is not merely an ethical preference or cultural argument; it reached a
technically meaningful error regime on a large Bambara corpus snapshot. The failed
May 2026 strict audit does not erase that contribution, but it changes how the
contribution must be stated. The result is an archived benchmark anchor awaiting
future strict reproduction, not a freshly closed leaderboard.

The immediate publication path should follow that truth. The flagship paper can
present script invisibility and computational infrastructure. The metric paper can
argue why CER is the right target for Manding ASR. The ASR technical report
can preserve the checkpoint, the trajectory hypothesis, the
low-learning-rate negative controls, the Djoko deployment lessons, and the unfinished
audit in one honest artifact. That is enough to conclude the current research phase
without spending more money, while still giving future collaborators a precise map
of what to reproduce next.

plainnat
../references