Analyst
A permission-aware investigator over your alerts, incidents, assets and playbooks.
Your analysts stop losing the first forty minutes of every incident to assembling context.
A conversational assistant operates over alerts, incidents, assets, approved internal knowledge and the event store. It builds cited timelines, finds similar past incidents, retrieves approved playbooks, and drafts hunting queries, case notes, shift handovers and executive summaries.
- The M0 event store and connected alert sources
- Asset context and approved response playbooks, indexed locally
- Prior incidents — subject to evidence permissions
The mechanism, not the marketing. If a step below would not survive an engineer's scrutiny, it should not be on this page.
Operating path
From signal to auditable evidence
01Observed inputs
- The M0 event store and connected alert sources
- Asset context and approved response playbooks, indexed locally
- Prior incidents — subject to evidence permissions
02Deterministic workflow
- 1 Nothing is re-adjudicated Alerts keep the verdicts their originating tools gave them. The copilot assembles context around a decision that has already been made by something authoritative; it does not quietly re-decide it.
- 2 Retrieval is trimmed at query time, fail-closed Every candidate chunk is checked against the requesting analyst's own permissions and the source system's ACL snapshot. Stale, missing or unresolved ACLs deny rather than allow. A connector service account's broad read access never becomes the analyst's access.
- 3 Context is bounded before the model sees it The authorised evidence set for this case — not the mailbox, not the repository, not the log corpus. Evidence-capture=none also prohibits hidden prompt, trace or cache retention of that content.
- 4 Every factual claim carries lineage Source identifier and version accompany the statement. Anything the retrieved evidence does not support is marked inference or unknown. It is never asserted and then softened.
- 5 Fact, inference, hypothesis and recommendation stay typed They are distinct in the output structure, not merely in tone, so the model cannot collapse a guess into a finding by writing confidently.
- 6 Tools split into read, draft and execute The copilot holds read and draft. Execute requires deterministic policy plus explicit authorised approval, and runs through the target system's own controls — the EDR contains the host, not us.
- 7 The whole invocation is recorded Model and adapter version, prompt-template version, retrieval sources, requesting identity, tool calls, policy context, output disposition, latency and reviewer outcome.
03Evidence outputs
- An invocation record per AI call: model and adapter version, prompt-template version, retrieval-source identifiers, requesting identity, tool calls, policy context, output disposition, latency, reviewer outcome
- Every tool request, approval, denial, execution and result in the tamper-evident store
- Analyst disposition captured for evaluation
02:10 on a Saturday: “why is this alert risky?”
One analyst on shift, an EDR alert on a finance workstation, and six systems holding one piece of the answer each.
Under two minutes to a cited timeline that would have taken most of the shift to assemble by hand — and every claim in it traceable to an event, including the honest admission of what was missing.
The authority ladder. Green rungs are authoritative — a rule, a scanner, a sandbox, or a person. Blue is where the model contributes. Amber is where a human decides.
What the model adds
The assembly work — pulling six systems into one cited narrative, and saying what is missing from it.
What it cannot claim
It cannot establish that an intrusion occurred. Model output alone is not proof of compromise.
- Never contains a host, disables an account, or deploys a detection rule.
- Never presents inference as observed fact. Observed fact, inference, hypothesis and recommendation stay visibly distinct.
- Never retrieves data the requesting analyst could not open themselves. A connector service account's broad read access never becomes the user's access.
- Never answers from retrieval without naming the source and its version.
- An invocation record per AI call: model and adapter version, prompt-template version, retrieval-source identifiers, requesting identity, tool calls, policy context, output disposition, latency, reviewer outcome
- Every tool request, approval, denial, execution and result in the tamper-evident store
- Analyst disposition captured for evaluation
A specification, not a simulation
We will not show you a scripted demonstration of something that does not exist. What we publish instead is the workflow specification, held to the same standard as shipped work: actor, trigger, happy path, exceptions, terminal state, and the audit evidence it must leave. When it is built, this section becomes a demonstration.
Status and what promotion requires
Candidate next release — the highest-priority module after the M0/M1/M2 foundation. Promotion requires design partners, a named SIEM/EDR connector, and targets for citation correctness, query validity and analyst time saved.