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?
AI research lab · 2026
Intelligence
in motion.
We study how intelligent systems reorganize internal representations, adapt their reasoning, and develop new capabilities.
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.
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)
Controlled benchmark / 18 July 2026
Measured,
then claimed.
We tested the public PyPI releases of TrainTools and MBE on a frozen synthetic classification protocol: 48 configurations, each trained once normally and once with instrumentation.
48 / 48Exact final-model matches
100 / 100Injected faults detected
0 / 50Clean-control warnings
0.99995Mean AUM AUROC
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. AUM reached 99.5% mean precision at the top 20%, against a 20% random baseline. Median runtime was 1.19x in this tiny CPU-bound, probe-heavy test; it is not a production-overhead estimate.
PTDB-1 architecture replication / 22 July 2026
Two architectures.
Every row auditable.
ResNet-18 and WRN-28-2 now 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. AUM led EL2N in all six symmetric-noise runs; EL2N led AUM in all six class-conditional runs. The metric ordering replicated with the corruption regime.
Boundary: synthetic corruption, one dataset, two related CNNs, and four regime clusters. Gradient-norm and stopping claims are withheld; the protected holdout remains sealed.
Publication boundaryThe accompanying 48-row MBE audit is not used as 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.
Correction logWithheld MBE report
Correction logWithheld MBE report
Public telemetry / 18 July 2026
Attention is
not adoption.
Package downloads and repository traffic are weak public signals. They help track curiosity, but they do not prove active users.
MBE
Stronger curiosity signal: substantial GitHub views and clones, but no public engagement markers yet.
~725PyPI downloads since 23 Jun, mirrors excluded
122Unique GitHub viewers over 14 days
100Unique GitHub cloners over 14 days
0Stars, forks, watchers, and open issues
- Largest observed PyPI days: 175 downloads on 13 Jul and 102 on 17 Jul.
- GitHub traffic peaked on 13 Jul with 99 unique viewers and 37 unique cloners.
- Interpretation: organic curiosity is plausible; public adoption is not established.
TrainTools
Early package activity with release-shaped spikes; adoption signal is still weak.
545PyPI downloads last month, mirrors excluded
253PyPI downloads last week, mirrors excluded
37PyPI downloads last day, mirrors excluded
37Unique GitHub cloners over 14 days
- Latest public release observed: v0.6.2 on 16 Jul.
- GitHub showed 6 unique viewers over 14 days, with 0 stars and 0 forks.
- Interpretation: some touches and installs; not enough evidence for organic users yet.
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 an independent AI research lab founded by Aparajeet Shadangi in Bhubaneswar, India.
At the beginning by choice: forming directions, building open tools, and finding questions worth reorganizing around.
- Founder
- Aparajeet Shadangi
- Location
- Bhubaneswar, India
- GitHub
- github.com/AparajeetS ↗