Repository Public Research preview
1% ============================================================================2% Limitations. Sources: docs/gates/GATE_E.md, GATE_D.md, docs/STATUS.md, artifacts/metrics/.3% ============================================================================4\section{Limitations}5\label{sec:limitations}67\begin{itemize}8 \item \textbf{Scale of the models.} The confirmatory model has 33.5\,M parameters and was trained for about 2.79 accelerator-hours. HTR-Small (130.8\,M) was under-trained on the same step budget, and the Base and Large design10 points were still in training at the time of writing (\pending{HTR-Base and HTR-Large results}). The conclusion11 that verbatim reconstruction is close to the identity ceiling is established for small models; whether larger12 models can exceed it is open.13 \item \textbf{One training rewriter.} All training pairs are Haiku~5.5 rewrites. Transfer to Opus rewrites was14 positive but below the SESOI, and transfer to held-out prompts and to the Sonnet/Haiku pilot set was negative.15 \item \textbf{Light rewrites are over-corrected.} The model degrades rewrites that are already close to the16 original (variants B and C, held-out prompts). It has no mechanism to estimate rewrite strength and abstain.17 \item \textbf{Factual cost.} Number recall (0.991) and placeholder exactness (0.992) are below identity's 1.000,18 and non-inferiority on numbers was not established. Every output carries preservation warnings, but a user must19 check flagged spans.20 \item \textbf{Corpus composition.} The corpus is mostly PMC biomedical and life-science prose from 2010--2017;21 pre-1950 material comes almost entirely from books, and the pre-1900 held-out period showed a lexical loss.22 \item \textbf{No human evaluation.} No independent rater has judged the reconstructions23 (\pending{independent blind rater study}).24 \item \textbf{Dual-latent objective untested at a scale where it could help.} The full M3 objective failed at25 pilot scale and only one M3-light seed was trained at Medium scale; H6 was not tested as specified.26 \item \textbf{Detector results are narrow.} The Pangram measurement covers 139 short passages, one detector and27 one detector version.28\end{itemize}29