Flagship · Autonomous Data Scientist

Forty-eight wired modules. Forty-three typed tools. One sealed evidence chain.

The agent runtime that runs the platform. It plans the work, asks before spending, executes typed Python tools, validates generation against real physics, self-heals when a quality gate trips, and seals the entire conversation into the same signed evidence bundle as Mock and Synthesize.

Wired modules
48
Planner tools
43
Dispatch engines
17
Evidence
9-artefact BLAKE3
customer-cohort-q3 · /agentplan-mode
Build me a clean training set from /datasets/customers.csv, drop nulls > 5 %, encode categoricals, train a baseline.
proposed plan · 4 steps
data_cleandrop_threshold=0.05~5c
encode_categoricalmethod=onehot~3c
train_predictivemodel=randomforest~12c
shap_explainsample=200~3c
est. total23 credits · ~38 s
▶ data_clean ✓ 4.2 c 2.1 s
▶ encode_categorical ✓ 2.8 c 1.4 s
▶ train_predictive … 7.1 / 12 c
▷ shap_explain queued
estimated 14.2 s remaining · evidence chain rolling
Module map

Forty-eight specialised modules. One coherent runtime.

Six families: planning & approval, cost & scheduling, physics validation, self-heal & resilience, code generation & review, evidence & narrative. Each module has one job. The 17 highlighted below are the core control surface; 31 more specialised modules are wired end-to-end in code (approval_timeout, contract_context, dataset_resolver, evidence_parser, executor_edsha, executor_eval, executor_steps_ml, grounded_output, hypothesis_tree, memory, moderation, observer, patch_planner, policy_adapter, prompt_registry, replan_evaluator, replay_bundle, retry_escalate, rseg_calibration, rseg_evaluator, scenario_selector, turn_context, validation_runner, budget, hagp_pause, mmd, retry_policy, surrogate, diff_builder, domain_prompt, dataset_context).

ads / module map · 17 highlighted · 48 wired · 4 lanes
all green
PLANCOST / PHYSICSHEAL / CRITICCODE / NARRATIVEAPPlannerPEExecutorHEALSelf-HealGATEApproval GateShShadowCOSTCost ManagerAUDITAudit TrailPHYPhysics ValidatorELSEngine SchedulerPPProof PacketsCAABAssertion BuilderRPReplanRSEGResult GraderCACRChange ReviewerCGCCode CriticCCMChange ManagerNGNarrative Generator
active runs
7
self-heal cycles
3
approvals queued
1

Planning & approval

4 modules
AP

Agent Planner

Decomposes intent into typed step graph; estimates cost per node before any tool fires.

PE

Plan Executor

Runs the graph with structured concurrency, retries, and per-step checkpointing.

HAG

Human Approval Gate

Pauses on high-cost or high-risk steps; surfaces a typed approval card to the operator.

Sh

Shadow

Dry-runs proposed plans against last week's data to surface regressions before commit.

Cost & scheduling

3 modules
CM

Cost Manager

Hard caps per project, soft caps per turn; halts when burn rate exceeds policy.

CAT

Cryptographic Audit Trail

Per-tenant compute throttle with full audit traceability; preempts low-priority work when high-priority arrives.

ES

Engine-Level Scheduler

Routes work to the cheapest engine that satisfies the contract (Mock vs Synthesize vs ADS-driven).

Physics validation

1 modules
PV

Physics Validator

Constrains generation to physically-realisable signal shapes using a physics-honest response model.

Self-heal & resilience

5 modules
SHX

Self-Healing Execution

Watches the evidence chain; if a quality gate fails, repairs the failing artefact and replays downstream.

CAB

Consistency-Aware Builder

Re-derives downstream artefacts when upstream inputs change so the evidence chain stays consistent.

RP

Replan

Mid-run plan revision when an unexpected error or insight makes the original plan stale.

REG

Risk Evidence Graph

Risk-stratified evidence graph; flags brittle or low-confidence sub-paths for review.

CCR

Cross-Correlation Rebuilder

Preserves PK/FK and joint distributions when artefacts are regenerated.

Code generation & review

2 modules
CGC

Code-Gen Critic

Reviews tool-generated Python before execution; rejects unsafe or undefined behaviour.

CCM

Code Change Manager

Versions every tool-generated Python; lets the operator diff and roll back at any step.

Evidence & narrative

2 modules
PP

Proof Packets

Bundles per-step inputs, outputs, hashes, and parameters into the signed evidence chain.

NG

Narrative Generator

Produces a plain-English run-log alongside the JSON artefacts — for the audit reviewer.

Tool palette

Forty-three typed tools the agent can wield.

All five families are wired through both the chat interface and POST /v1/agent/tools/{name}/execute for direct invocation from the SDK or your own UI.

Data preparation

6 tools
data_clean
feature_engineer
transform_data
encode_categorical
scale_numeric
split_dataset

Machine learning

7 tools
train_predictive
tune_hyperparameters
cross_validate
evaluate_model
cluster
detect_anomalies
dimensionality_reduce

Explain & report

8 tools
shap_explain
feature_importance
profile_dataset
scatter_matrix
time_series_plot
decompose_timeseries
forecast
generate_report

Synthetic data

5 tools
mock_data
synthesize
fabricate_dataset
verify_synthesis
quality_report

Industrial / ICS

6 tools
scada_sim
ics_attack
inject_mock
analyze
verify
evidence_seal
Self-healing execution loop

Quality gate fails? It heals itself.

When a K-S divergence widens, a constraint slips, or a hash mismatch appears, the self-healing layer isolates the failing sub-graph, repairs the affected artefact, and rebuilds the chain — without losing what already passed.

The narrative file gets a paragraph explaining what happened and why the rebuild is sound. Your auditor reads English, not commit hashes.

step · 01

Sense

Self-healing execution watches every quality gate, every artefact hash, every chain segment in real time.

step · 02

Diagnose

The risk-evidence graph isolates the failing sub-graph; the cross-correlation rebuilder identifies which downstream artefacts are now stale.

step · 03

Repair

Replan rewrites the affected sub-plan; the code-gen critic reviews any generated code before re-execution.

step · 04

Re-seal

The consistency-aware builder rebuilds the downstream chain; proof packets commit the new cryptographic root; the narrative generator appends to the run-log.

Every engine under one orchestrator.

The engine scheduler routes the work to the cheapest engine that satisfies the contract; the physics validator keeps generation physically realisable; proof packets seal the result.

Mock Data

Sub-minute deterministic synth from a prompt.

Synthesize

Our trained synthesis engine on uploaded data.

Virtual SCADA

Physics-honest sensor telemetry over real protocols.

ICS Security

Ground-truth ATT&CK ICS attack mix injection.

Tool executor

43 typed Python tools in-process or queued.

Connectors

14 encrypted source adapters, browse + import.

evidence pipeline · job_9b3df1
live
K
Sealed contract
E
Engine run
Q
Quality gates
B
BLAKE3 chain
S
Sealed bundle
rows generated
10 000
qa score
0.957
chain root
a4f2…d801

Built so an autonomous agent can survive procurement.

Tenant-isolated

Every plan, prompt, dataset preview, and tool I/O lives in your tenant prefix only.

Spend-bounded

Soft + hard cost caps; the agent stops, asks, or downgrades the engine before exceeding policy.

Audit-ready

Sealed transcript artefact per project — prompt, plan, approvals, tool I/O, evidence root.

Reproducible

Re-run the same sealed contract + seed on a different cluster: the dataset hash matches byte-for-byte.

Bring a real prompt. Watch every module work.

90-minute working session: bring a representative dataset and one open question. We drive the agent end-to-end, you watch the plan card, the cost gate, the live exec stream, and the self-heal trigger if a gate trips. You keep the sealed transcript bundle.

ads / execution loop · job_9b3df1
in flight
01
Plan
4–8 steps · typed tools
02
Execute
engines + cost gate
03
Self-Heal
repair · replan · grade
04
Narrate
English proof · sealed
evidence.stream● rolling
▶ plan.proposed 4 steps · 23 credits
⌀ approval.gate 1 queued · awaiting hagp
▶ exec.data_clean ✓ 2.1 s · 4.2 c
▶ exec.train … 7.1 / 12 c
⟲ heal.replan K-S drift · patch bus_3
▣ narrate.chapter "Why we re-ran step 04"