AEGIS Trial SDK

AEGIS Trial SDK is the public local integration layer for a separately supplied compatible AEGIS Trial Core wheel. It is detection-only: it observes local model state, returns local events, and does not change a training run.

Initial Trial platform

The initial live Trial release requires Linux x86_64 and CPython 3.11. The supplied Core wheel currently has the form:

aegis_trial_core-0.1.2-cp311-cp311-linux_x86_64.whl

Version numbers and filenames do not establish scientific qualification. Use only a Core wheel whose supplied hash and release status explicitly identify it as authorized for the trial; engineering-candidate wheels must not be deployed.

The SDK exposes PyTorch, Hugging Face Trainer, and distributed integration helpers. Those APIs do not by themselves establish that a particular model architecture or multi-worker training configuration is validated for the Trial Core. Use only the configuration supplied with your trial materials.

Install

Install both components into the same CPython 3.11 environment:

python3.11 -m pip install /path/to/aegis_trial_core-0.1.2-cp311-cp311-linux_x86_64.whl
python3.11 -m pip install aegis-trial-sdk

When installing from a source checkout, replace the second command with python3.11 -m pip install ..

Before installation, check the Core wheel against the SHA-256 value supplied with your trial delivery:

echo "<supplied-sha256>  aegis_trial_core-0.1.2-cp311-cp311-linux_x86_64.whl" | sha256sum --check -

The public SDK does not install or bundle the Core. Constructing AegisMonitor without the separately supplied compatible wheel produces a local error.

Training-loop lifecycle

Call begin_step before the monitored forward pass and end_step after the normal user-owned training step. See examples/pytorch_training_loop.py.

from aegis import AegisMonitor

# This must equal the number of training forwards accumulated into each
# optimizer step. The monitor aborts on missing or extra captures.
monitor = AegisMonitor(model, expected_microbatches=1)
monitor.begin_step(step)
loss = model(batch).sum()
loss.backward()        # your ordinary training loop
optimizer.step()       # your ordinary training loop
event = monitor.end_step(loss=loss)
if event.detected:
    print(event.anomaly_state, event.detection_confidence)

AEGIS does not modify gradients, optimizer state, learning rate, checkpoints, workers, or training control. It only returns the event.

If the training loop uses gradient accumulation, construct the monitor with the same explicit count. For example, use expected_microbatches=4 when four forward and backward passes precede each optimizer step. A mismatch raises an error instead of silently evaluating an incomplete optimizer step. The compatible Core combines every configured observation before producing the local event.

Event schema

AegisMonitor.end_step() returns an immutable AegisEvent with:

• anomaly_state: a generic local state, such as clear or detected.
• detection_confidence: a numeric confidence when available, otherwise None.
• recommended_next_step: informational text for the caller. In the Trial, it does not select, authorize, or execute an intervention.
• detected: a convenience property derived from anomaly_state.

The Core's signal_detected decision field is not emitted as a separate public event field; callers should use event.detected.

Use aegis.integrations.AegisCallback(monitor) with a Hugging Face Trainer, or aegis.integrations.gather_events(event, torch.distributed) to collect one event per worker on rank zero. The distributed example is an API integration reference, not a broader validation claim. See examples/ddp_integration.py.

Data and network boundary

AEGIS runs locally. The public SDK does not upload training data, model state, or event logs, and opens no network connections. See DATA_FLOW.md and PRIVACY.md.

aegis-demo --preview writes a clearly simulated evaluation-report preview.

Public SDK release integrity

Every published SDK release should include a SHA256SUMS file alongside its wheel and source distribution. Verify the downloaded artifact against that file before installation:

sha256sum --check SHA256SUMS

SHA256SUMS covers the public SDK artifacts only. The separately supplied Trial Core has its own delivery manifest and SHA-256 value; never substitute one checksum for the other.