Research methods / Security identity
Why a Ticker Is Not a Security Identifier
Eight rejected dividend records, an issuer-filing contradiction, and the evidence required before attaching corporate actions to a historical security.
Methods, evidence & open questions
Research
Research perspectives and technical notes. Start with market data and portfolio accounting, or browse the AI and engineering work.
Markets & data
Identity, events, and accounting
Research methods / Portfolio accounting
A Total-Return Index Is Not an Executable Portfolio
A specification of close-reinvested return indices and raw-share portfolio accounting, including dividend entitlement, event sequencing, and funding constraints.
AI & engineering
The wider research archive
Perspective / Model evaluation
What would make a 400M-parameter model worth releasing?
A proposed framework for testing task usefulness, deployment cost, and release readiness.
Experiments / Evaluation methodology
Measuring training noise before comparing small language models
Same-seed retraining and seed changes moved validation scores more than most of the mechanism effects we tested on small models.
Engineering / Activation checkpointing
When attention settings change during checkpoint recomputation
A CPU reproducer demonstrates that attention-backend settings must stay consistent between forward and checkpoint recomputation, with three repairs.
Engineering / Activation checkpointing
Why checkpointed attention failed during backward
A CPU reproducer shows how an attention context can end before checkpoint recomputation, and three ways to keep execution settings consistent.
Experiments / Representation learning
Corrected statistic alone restores spread but fails rank
The corrected anti-collapse statistic restored feature spread but left effective rank near one, showing why both metrics matter.
Experiments / Representation learning
Fixing feature spread did not fix representation collapse
A four-variant synthetic experiment separates feature spread from dimensional collapse and tests a combined stability repair.
Engineering / Tokenization
Faster BPE Without Changing a Single Token
Semantics-preserving optimization of BPE fitting and encoding, with paired runtime measurements, differential validation, and explicit memory tradeoffs.
New to the topic? Read the price-adjustment guide, dividend-date guide, and look-ahead bias guide. Follow publications with RSS.