Gradient Noise Scale
Estimates the critical batch size using stable exponential-moving-average estimators. During gradient accumulation it requires no extra forward or backward passes.
Is my batch size wasting compute?
Founder-led AI research lab · 2026
Intelligence
in motion.
Phason Labs studies how models change internally, tests which signals survive controls, and turns reliable methods into tools and independent audits.
Phase space / φ + Δφ
Move cursor to perturb field
Move cursor to perturb field
Scroll to enter
The most important advances in AI may come from understanding how intelligence reorganizes itself. Not simply from making today’s systems larger.
Active work
Evidence,
not theatre.
Two active projects. Each tied to a concrete question, public code, and an honest statement of maturity.
Does a training metric still predict held-out performance after ordinary explanations are controlled?
A metric can look impressive under pooled correlation while mostly tracking learning rate, architecture, task, seed, validation loss, or another design variable. Marginal Baseline Evaluation compares raw association against controlled partial rank-correlation to expose that failure mode.
$ pip install mbe-eval
$ mbe-eval-demo --bootstrap 200
$ mbe-eval-demo --bootstrap 200
680Trained models
480CIFAR-10 image models
200Character transformers
40+Candidate metrics
Current evidenceOne metric.
One metric.
Three different stories.
FIM_norm is the motivating case study. Its apparent relationship changes after controls and across task domains—exactly the behavior MBE is designed to reveal.
AuditRaw ρPartial ρClassification
Image / default controls−0.662−0.218Survives
Text / default controls−0.291+0.014Washout
Full pool / default controls+0.225−0.203Reverse-inversion
Raw pooled correlation is insufficient for validating ML training metrics. The useful question is whether signal survives ordinary baselines and experimental design variables.
Evidence / metric evaluationTwo architectures.
Two architectures.
Every row auditable.
ResNet-18 and WRN-28-2 have separate raw-data bundles, frozen analyses, and standard-library audits.
810kExample-level records recomputed
24 / 24Complete executions
6 / 6Exact clean model pairs
5.5e-15Maximum metric discrepancy
Across both architectures, AUM reached 0.9872 aggregate AUROC and beat mean loss in all twelve noisy runs. In the controlled synthetic benchmark it reached 0.99995 mean AUROC and 99.5% mean precision at the top 20%, against a 20% random baseline.
AUM led EL2N in all six symmetric-noise runs; EL2N led AUM in all six class-conditional runs. The ordering replicated with the corruption regime rather than collapsing into a single headline.
Publication boundarySynthetic corruption, one dataset, two related CNNs, and four regime clusters. The separate 48-row MBE audit is withheld from promotional evidence because its random negative control missed the frozen point-estimate threshold, although its confidence interval crossed zero. The defective first attempt and corrected result remain public.
Three PyTorch diagnostics. Each answers a decision you face during training.
TrainTools turns training behavior into actionable signals: whether batch size is wasting compute, whether a network is losing its ability to learn, and whether continued training is worth the cost.
Plasticity Probe
Measures dormant-unit fraction and effective rank of activation covariance, combining them into a plasticity score from zero to one.
Is my network losing the ability to learn?
TrainGuard
Fits power-law or exponential curves to validation loss, bootstraps uncertainty, and estimates whether additional training is worthwhile.
Should I stop training yet?
Integration
Two lines into an existing loop.
Core support targets PyTorch 2.0+ and Python 3.9+. Optional extras add curve fitting, plotting, and HuggingFace Trainer integration.
$ pip install traintools[full]
from traintools.callbacks.pytorch import TraintoolsTracker
tracker = TraintoolsTracker(model, loss_fn)
decision = tracker.step(step=step, inputs=x, targets=y, val_loss=val_loss)
from traintools.callbacks.pytorch import TraintoolsTracker
tracker = TraintoolsTracker(model, loss_fn)
decision = tracker.step(step=step, inputs=x, targets=y, val_loss=val_loss)
Evidence / instrumentationDiagnostics without
Diagnostics without
changing training.
Public releases were tested on a frozen synthetic classification protocol, with every configuration trained normally and again with instrumentation.
48 / 48Exact final-model matches
100 / 100Injected faults detected
0 / 50Clean-control warnings
1.19×Median probe-heavy CPU runtime
Instrumentation left every final parameter hash, validation loss, and test accuracy unchanged. The suite detected NaNs, invalid labels, extreme class imbalance, and amplified gradients in all 25 trials per condition.
The runtime figure comes from a tiny CPU-bound, probe-heavy benchmark. It is descriptive evidence of this test configuration, not a production-overhead estimate.
Public telemetry / 26 July 2026
Installs are
not adoption.
PyPI install/download counts are public traction signals. They show that packages are being fetched, but they do not by themselves prove sustained human usage.
MBE
2,802 PyPI install/download events recorded so far. This is real package activity, not yet proof of sustained users.
2,802Total PyPI install/download events
2,443Last 30 days
199Last 7 days
23Latest recorded day
- Package observed: mbe-eval, not the unrelated PyPI package named mbe.
- Source: Pepy public PyPI statistics through 25 Jul 2026; CI traffic is included.
- Interpretation: early install traction is visible; active human usage is not established.
TrainTools
3,193 PyPI install/download events recorded so far. Continued weekly activity suggests curiosity beyond a zero-noise launch.
3,193Total PyPI install/download events
2,363Last 30 days
262Last 7 days
23Latest recorded day
- Latest public release observed: v0.6.2.
- Source: Pepy public PyPI statistics through 25 Jul 2026; CI traffic is included.
- Interpretation: real package activity is present; this should not be described as 3,193 confirmed users.
Lab method
How the work
gets done.
The lab’s philosophy becomes credible only when it constrains the research process.
Frame a falsifiable question
Start with a claim that evidence could genuinely weaken.
Instrument before benchmarking
Build visibility into the process before deciding what score matters.
Control simple explanations
Compare new metrics and methods against ordinary baselines.
Record negative results
Washouts, inversions, and failed intuitions remain part of the evidence.
Release reproducible artifacts
Code, protocols, and evidence ledgers accompany public claims.
Research atlas
A field of
questions.
Active work anchors the map. Speculation stays at its edge.
Field notes
Thinking
in public.
Not a content feed. A quiet record of hypotheses, methods, failed intuitions, and changes of mind.
N–001Open question
Beyond loss: what should training diagnostics measure?
Separating descriptive metrics from metrics that change decisions.
N–002Research thesis
Why intelligence may be a reorganization problem
Internal structural change as a source of capability.
N–003Open question
Tools before benchmarks
Instrumentation can reveal the question a benchmark should answer.
The lab
Independent
by design.
Phason Labs is a founder-led AI research lab built by Aparajeet Shadangi in Bhubaneswar, India.
The lab is currently focused on model-internals research, open diagnostic tools, and independent evidence audits for machine-learning teams and maintainers.
- Founder
- Aparajeet Shadangi
- Location
- Bhubaneswar, India
- GitHub
- github.com/AparajeetS ↗