Comparable experiments
Share model ingredients, data identities, budgets, and evaluation rules across strategies.
Hyphae Transformer Controlled transformer research
Keep the question, experiment, and result together.
Hyphae Transformer is a PyTorch research framework for controlled residual-strategy experiments. Later work studies small controllers over a frozen model and bounded navigation with explicit evidence contracts.
Make architecture and control experiments assessable through fixed protocols, retained artifacts, and visible failures.
01 / Why it matters
Why this work matters
Explicit research objects
Keep hypotheses, seeds, manifests, checkpoints, and verdicts together.
Frozen-model control
Train a small controller while retaining the backbone’s identity and host-owned policy boundaries.
Visible uncertainty
Keep inconclusive primary results and failed external evaluations beside successful runs.
02 / How it works
How it works
The main stages make the project’s boundaries visible.
- 01
Define
Fix the hypothesis, data, seeds, effect threshold, and budget
- 02
Validate
Check contracts and run a small correctness pilot
- 03
Evaluate
Execute the declared training and held-out comparisons
- 04
Retain
Publish results, artifact identities, limitations, and the next decision
03 / Capabilities
Current capabilities
Residual research
Controlled comparisons of decoder-only residual strategies, including shared-gate ReZero.
Experiment management
Typed manifests, resumable checkpoints, evidence records, budgets, and reports.
Bounded controllers
Frozen-backbone experiments with answer, request-evidence, and abstain decisions.
Navigation bundles
Deterministic published controller artifacts, including calibrated two-step navigation v2.
04 / Evidence
Evidence and its scope
The historical 30M result and later control/navigation work are separate, bounded lines of evidence.
paired final-NLL improvement
30M campaign, eight paired seeds
95% interval
Specific to that comparison
calibrated navigation bundle
Framework remains alpha 0.2.0
The 30M experiment cleared its preregistered practical threshold. Navigation results apply to the stated fixtures and certificates, with the backbone frozen; neither establishes general model quality.
What it does now
- Runs controlled architecture campaigns with fixed source, data, and budgets.
- Retains successful, inconclusive, and failed experimental outcomes.
- Publishes bounded control and navigation artifacts over frozen model features.
Next directions
- Test observed effects across further model sizes, corpora, and architectures.
- Expand independently annotated external evaluation before production claims.
- Advance navigation only through separately specified and measured protocols.
Limits to keep in view
- An alpha research framework, not a production model service.
- The navigation bundle’s v2.0.0 identifier does not change the framework version to 2.0.0.
- Fixture and held-out results do not establish an unrestricted autonomous agent.
- Earlier depth results, external-shadow failures, and budget defects remain part of the record.
Hyphae Transformer