ALL SYSTEMS
Behavior Constraint Satisfaction
0.0%
▲ +0.4%vs 24h prior
Drift Detection · 7-Day WindowSTABLE
D-7
D-6
D-5
D-4
D-3
D-2
D-1
No distributional anomalies detected in window
Reward Signal Decomposition
Alignment monitors RLHF pipelines in real time — surfacing reward hacking, goal drift, and distributional shift before they reach production.
Reward Hacking Detection·Goal Drift Monitoring·Distributional Shift Alerts·RLHF Pipeline Integration·Audit-Ready Exports·Real-Time Eval Scoring·PyTorch · JAX · TF · vLLM·SOC 2 Type II Ready·Reward Hacking Detection·Goal Drift Monitoring·Distributional Shift Alerts·RLHF Pipeline Integration·Audit-Ready Exports·Real-Time Eval Scoring·PyTorch · JAX · TF · vLLM·SOC 2 Type II Ready·
01 · MonitorContinuous behavior surveillance
Continuous behavior surveillance
at every training step.
Alignment hooks into your RLHF loop via a single decorator. Every reward signal, policy update, and KL divergence measurement is captured, decomposed, and surfaced in under 200ms.
Reward Hacking Score
0.003σ
NOMINAL↓ -12%
Goal Drift Index
0.018
NOMINAL→ stable
Distributional Shift
0.041KL
WATCH↑ +8%
Constraint Violations
0/hr
NOMINAL→ stable
Live Event Streamstep 18,401
02:14:07INFOConstraint eval batch #4421 completed
02:14:09INFOReward decomposition within expected bounds
02:14:12WARNKL divergence spike: 0.043 → 0.061 on layer 28
02:14:12INFOAuto-checkpoint triggered at step 18400
02:14:15INFOGoal representation stable: cosine sim 0.994
02:14:18INFOBehavior constraint satisfaction: 99.2%
Python · Quick Integration
import alignment
@alignment.monitor(
metrics=["reward_hacking", "goal_drift"],
alert_threshold=0.05,
checkpoint_on_anomaly=True
)
def training_step(batch, model, optimizer):
loss = model(batch)
optimizer.step(loss)
return loss
# Hooks into your loop — zero overhead
# on nominal steps (<0.3ms p99)02 · EvaluateStructured evals that run
Structured evals that run
inside your training loop.
Define evaluation suites in YAML or Python. Alignment executes them at configurable intervals — every N steps, on checkpoint, or triggered by a metric threshold crossing.
SuiteCasesPass RateLast RunStatus
reward_hacking_v22,840
98.7%
14m agoPASSgoal_stability_bench1,200
99.1%
14m agoPASSdistributional_ood680
94.2%
1h agoWATCHconstraint_adherence3,100
99.2%
14m agoPASSYAML · Eval Suite Definition
# alignment.eval.yaml
suite: reward_hacking_v2
trigger:
every_n_steps: 500
on_metric_threshold:
metric: kl_divergence
value: 0.05
cases:
- source: ./evals/reward_hacking/
- source: huggingface://alignment-bench/rh-v2
scoring:
method: categorical
pass_threshold: 0.95
on_fail:
alert: slack
block_checkpoint: trueAPI Response · application/json
{ "eval_run_id": "eval_20260224_141237", "model": "rlhf-v4-checkpoint-18400", "timestamp": "2026-02-24T14:12:37Z", "suites_run": 4, "overall_pass_rate": 97.8, "flags": [ { "suite": "distributional_ood", "severity": "watch", "metric": "kl_divergence", "value": 0.061, "threshold": 0.05, "recommendation": "monitor_next_500_steps" } ], "constraint_satisfaction": 99.2, "reward_decomposition": { "helpfulness": 0.84, "harmlessness": 0.96, "honesty": 0.91 }}
03 · AuditEvery decision, logged and
Every decision, logged and
exportable on demand.
Immutable audit trails with cryptographic signing. Generate compliance reports for internal review boards, regulatory submissions, or incident post-mortems in under 30 seconds.
Immutable Audit Log
SHA-256 signed · tamper-evident2026-02-24 14:12:37 UTCPASS
Eval suite reward_hacking_v2 completed
AUD-2026-0224-001 · system
2026-02-24 14:14:12 UTCFLAG
KL divergence threshold crossed on layer 28
AUD-2026-0224-002 · system
2026-02-24 14:14:12 UTCINFO
Auto-checkpoint created at step 18400
AUD-2026-0224-003 · system
2026-02-24 14:19:04 UTCACK
Human reviewer acknowledged KL flag
AUD-2026-0224-004 · mwilson@lab.ai
2026-02-24 14:22:18 UTCEXPORT
Constraint satisfaction report generated
AUD-2026-0224-005 · priya.nair@lab.ai
Export Formats
📄
.PDF
Human-readable compliance report
{ }
.JSON
Machine-readable audit log
⊞
.CSV
Metric time-series export
⊕
.SPDX
Software provenance manifest
Compliance Frameworks
✓NIST AI RMF
✓ISO/IEC 42001
✓EU AI Act Art. 13
✓SOC 2 Type II
✓NIST CSF 2.0
04 · IntegrateDrop into your existing stack.
Drop into your existing stack.
Three commands to first signal.
Native support for PyTorch and JAX with full capability coverage. TensorFlow and vLLM at partial parity. Custom framework support via our instrumentation API.
Compatibility Matrix
PyTorch
JAX
TensorFlow
vLLM
Custom
Real-time hooks
●
●
●
◐
◐
Checkpoint intercept
●
●
◐
○
◐
Reward decomp
●
●
●
◐
◐
KL tracking
●
●
●
●
◐
Eval runner
●
●
◐
●
◐
●Full
◐Partial
○Roadmap
Python ≥3.10CUDA 11.8+PyTorch ≥2.0JAX ≥0.4Linux / macOSDocker ready
Setup · 3 Commands
1
Install
$ pip install alignment-sdk
2
Initialize
$ alignment init --project my-rlhf
3
Authenticate
$ alignment connect --api-key $ALIGN_KEY
JAX · Advanced Configuration
import jax
import alignment.jax as alm
# Configure monitoring scope
config = alm.Config(
project="rlhf-constitutional-v3",
checkpoints=True,
eval_suites=["reward_hacking_v2"],
alert_channels=["slack://safety-eng"],
)
# Wrap your train_step
@alm.instrument(config)
@jax.jit
def train_step(state, batch):
loss, grads = jax.value_and_grad(
loss_fn)(state.params, batch)
return state.apply_gradients(grads=grads)
# That's it. Alignment handles the rest.05 · Deploy
Ready to deploy with confidence?
Request sandbox access or download the technical spec sheet to evaluate Alignment against your current pipeline.
99.97%
Uptime
<200ms
Avg eval latency
1,240+
Models monitored
48M
Constraint checks/day