Archive

Entries are ordered by the event or original note date. Reconstructed entries are labeled “event” here and show their actual publication date on the post.

2026

• 09 Oct — A Distillation Weight Needs Its Temperature · research note

• 07 Oct — Perplexity Has a Tokenizer in Its Denominator · research note

• 05 Oct — NFC Cannot Fix the Wrong Diacritic · engineering note

• 02 Oct — Three Judge Calls Are Still One Item · research note

• 30 Sep — Count the Raster After Decoding · engineering note

• 28 Sep — A Ranking Metric Cannot Choose a Threshold · research note

• 25 Sep — The Perceptron Convergence Bound Has Units · research note

• 24 Sep — One Field Is Not a List Operation · engineering note

• 21 Sep — Test the State Space, Not the Noise · engineering note

• 20 Sep — A Holdout Is a Set of Equivalence Classes · research note

• 27 Aug — Defining a Reproducible Arithmetic Experiment in CalcGPT · engineering note

• 27 Aug — The 2.4M-Parameter Model Won—Until the Text Got Noisy · research note

• 27 Aug — Change One Slot, Watch What Else Moves · research note

• 27 Aug — Ten Thousand Adaptation Steps Could Not Remove One Scraper Artifact · research note

• event 27 Aug — The Panel Was Weak on Items and Useful for Ranking Systems · research note

• 27 Aug — Three Emulator Bugs, Three Different Tests · engineering note

• event 18 May — A Defensive Ollama Client for Local Judge Inference · engineering note

• event 02 May — The Interval Belongs to the Comparison · research note

• 14 Apr — The Artifact Chain Behind TinyFabulist · research note

• 17 Mar — What Must Be Fixed Before the Diacritics Comparison · method note

• 24 Feb — Decision Gates for an Open-Weight Judge Panel · method note

• event 15 Jan — TF3-RO: Training and Compressing a Romanian Model from Scratch · release note

2025

• 08 Dec — Designing a 51M-Parameter Romanian Model from Scratch · engineering note

• 27 Oct — One Year In: The Decisions Behind the Output · retrospective

• event 27 Oct — What the Diacritics Study Measured at InnoComp 2025 · research note

• 29 Sep — Romanian Diacritics: The Errors Hidden by Overall Accuracy · research note

• event 09 Sep — TF2: Open Models for English–Romanian Literary Translation · release note

• 04 Aug — What LoRA Changes in the TF2 Translation System · engineering note

• event 26 Jul — A Logistic-Regression Baseline Worth Keeping · retrospective

• 07 Jul — When BLEU and a Literary Rubric Answer Different Questions · research note

• 09 Jun — Why TF2 Chose Translation Before Romanian-Native Generation · research note

• 12 May — Treating LLM Judges as Measurement Instruments · research note

• event 29 Apr — Three Million Stories Are Still a Sparse Sample · research note

• event 29 Apr — TF1-EN-3M: From Six Slots to Three Million Fables · release note

• event 18 Mar — A Map of LLM-Generated Synthetic Data · release note

• 10 Mar — What Breaks When Fable Generation Reaches Millions · engineering note

• 03 Feb — Why TinyFabulist Starts with a Specification · research note

• 07 Jan — The 546× Speedup That Wasn’t One Number · retrospective

• 04 Jan — A Perceptron You Can Audit · retrospective

2024

• 09 Dec — Papers That Changed the TinyFabulist Design · retrospective

• 28 Oct — Three Decisions at the Start of the PhD · retrospective