SentraMind
← All crew Available for POC M0 Platform Core

Core

Policy, sealed evidence and the tamper-evident record the other capabilities stand on, with local accounts standing in for Active Directory in this POC.

Conceptual cutaway of the locally governed decision chamber and append-only evidence ledger.
Concept artwork, not a product interface or deployment claim. The mechanism and evidence flow below describe the current or specified workflow.

What changes for the organisation

A deterministic policy engine, sealed evidence and a hash-chained record for later capabilities to build on. The logic is real. The identities, endpoints and traffic it runs on in this POC are seeded or synthetic.

How it works

A local account table supplies identity and roles in this POC. One deterministic evaluator resolves subject, detector class, destination, action, override cost and evidence capture. Captured content is sealed under an evidence key whose secret half is split among officers, and each event is hashed over the previous hash and its own stored bytes.

What it watches

  • Local accounts: password sign-in and the roles stored on each account. Active Directory users, groups and OUs are the design target, and no forest is connected in this POC.
  • Chained product events: verdict decisions, user overrides and disputes, false-positive reviews, exception grants and revocations, policy publishes and rollbacks, account and role changes, each step of an evidence reveal, licence state changes and checkpoints.
  • Signed content packs: detector patterns, destination catalogue, report packs, policy templates, threat intelligence and knowledge content. Signed update bundles and a removable-media install path are not built.

How it actually works

The mechanism, not the marketing. If a step below would not survive an engineer's scrutiny, it should not be on this page.

  1. 1
    Identity is a local account in this POC
    An operator signs in with a password checked against an Argon2id hash in a local account table, and roles are read from that account. Sign-in is single-factor, and five consecutive failures lock the account for 15 minutes. Actors on events carry a Windows-shaped SID and account name from a seeded development roster. Active Directory with Kerberos sign-in is the design target, and it is not built.
  2. 2
    A published policy compiles to a snapshot
    A rule selects on subject (group SIDs, user SIDs, OUs), detector class and destination, and sets the action, fail mode, evidence capture and an optional override cost. One evaluator applies a fixed precedence (exception, then rollout stage, then rule action) for enforcement, dry-run and the effective-policy viewer, so the three cannot disagree on the same inputs.
  3. 3
    Publishing takes two people, and the dry run is a choice
    A dry run replays recorded events from a chosen window under the candidate version and the active one, and counts the actions that would change. A publish does not require one. A publish or rollback is proposed by a maker, approved by a distinct checker, and moves the active pointer. The runtime database credential can still write a version's rule document, so this binds operators, not a compromised service.
  4. 4
    Evidence is sealed against the appliance
    Each sealed record gets its own AES-256-GCM data key, bound to its event by additional authenticated data. The data key is wrapped to the public half of an X25519 evidence key, and the secret half is split among officers with Shamir sharing. A reveal records an investigation purpose, needs at least two officers, and chains each step. Officer keys are software keys in this POC and are labelled as such. Hardware tokens are absent.
  5. 5
    The record chains, and checkpoints sign the head
    Each event is sequenced under an advisory lock and hashed with SHA-256 over the previous hash and its RFC 8785 canonical bytes. The runtime database role has no UPDATE on the event tables. Ed25519 checkpoints sign the chain head. The verifier recomputes from the stored bytes and stops at the first bad row. The next checkpoint waits for an acknowledgement, and the only acknowledger here is a stand-in tool, so without it the cadence stops after one.
  6. 6
    Packs are signed, and update bundles are not built
    A content pack is checked against a trust store, then its Ed25519 signature, then its digest, and a refusal names the failure. The appliance holds its own trust root in this POC, so the check does not stop someone with root from swapping that root. Signed update bundles, A/B activation and a removable-media install path do not exist yet.

Operating path

From signal to auditable evidence

Available for POC

01Observed inputs

  • Local accounts: password sign-in and the roles stored on each account. Active Directory users, groups and OUs are the design target, and no forest is connected in this POC.
  • Chained product events: verdict decisions, user overrides and disputes, false-positive reviews, exception grants and revocations, policy publishes and rollbacks, account and role changes, each step of an evidence reveal, licence state changes and checkpoints.
  • Signed content packs: detector patterns, destination catalogue, report packs, policy templates, threat intelligence and knowledge content. Signed update bundles and a removable-media install path are not built.

02Deterministic workflow

  1. 1 Identity is a local account in this POC An operator signs in with a password checked against an Argon2id hash in a local account table, and roles are read from that account. Sign-in is single-factor, and five consecutive failures lock the account for 15 minutes. Actors on events carry a Windows-shaped SID and account name from a seeded development roster. Active Directory with Kerberos sign-in is the design target, and it is not built.
  2. 2 A published policy compiles to a snapshot A rule selects on subject (group SIDs, user SIDs, OUs), detector class and destination, and sets the action, fail mode, evidence capture and an optional override cost. One evaluator applies a fixed precedence (exception, then rollout stage, then rule action) for enforcement, dry-run and the effective-policy viewer, so the three cannot disagree on the same inputs.
  3. 3 Publishing takes two people, and the dry run is a choice A dry run replays recorded events from a chosen window under the candidate version and the active one, and counts the actions that would change. A publish does not require one. A publish or rollback is proposed by a maker, approved by a distinct checker, and moves the active pointer. The runtime database credential can still write a version's rule document, so this binds operators, not a compromised service.
  4. 4 Evidence is sealed against the appliance Each sealed record gets its own AES-256-GCM data key, bound to its event by additional authenticated data. The data key is wrapped to the public half of an X25519 evidence key, and the secret half is split among officers with Shamir sharing. A reveal records an investigation purpose, needs at least two officers, and chains each step. Officer keys are software keys in this POC and are labelled as such. Hardware tokens are absent.
  5. 5 The record chains, and checkpoints sign the head Each event is sequenced under an advisory lock and hashed with SHA-256 over the previous hash and its RFC 8785 canonical bytes. The runtime database role has no UPDATE on the event tables. Ed25519 checkpoints sign the chain head. The verifier recomputes from the stored bytes and stops at the first bad row. The next checkpoint waits for an acknowledgement, and the only acknowledger here is a stand-in tool, so without it the cadence stops after one.
  6. 6 Packs are signed, and update bundles are not built A content pack is checked against a trust store, then its Ed25519 signature, then its digest, and a refusal names the failure. The appliance holds its own trust root in this POC, so the check does not stop someone with root from swapping that root. Signed update bundles, A/B activation and a removable-media install path do not exist yet.

03Evidence outputs

  • A hash-chained event with its stored canonical bytes, an actor SID and account name where one was resolved, the policy version in force, and an evidence reference where content was sealed. The SIDs are seeded in this POC.
  • A verifier result naming the first failing sequence number and its kind, from the aidom-verify-chain CLI, or from bounded verification over the API for auditors and security admins.
  • Signed checkpoints, exported to file and pushed as CEF over syslog by one declared forwarder. No vendor SIEM has been tested, and the only acknowledger is a stand-in tool.
This diagram describes the specified workflow; the maturity label states its current delivery status.

A case

“Prove to me that nobody altered the record.”

A synthetic run of the real chain and verifier code, on two events built in memory. It shows what the answer to an auditor's question can say, and what it cannot.

Tamper

A short synthetic chain holds two events and a signed checkpoint, and the verifier passes. Then one stored byte of sequence 1 changes on purpose, at offset 288, so “Synthetic” reads “Rynthetic”. Nothing else moves.

Recompute

The verifier runs again. It computes 61fd4504… from the stored bytes, finds 22d2868d… stored beside them, and stops: verified false, at sequence 1, kind hash-mismatch.

Two answers

The auditor compares that with the checkpoint. Its signature still verifies, because it signs a body that names a hash, and that hash has not changed. The auditor goes by the verifier, not the signature.

Localisation

The break is pinned to one sequence number. The output also reads checkpointsChecked 0, because the verifier stopped before the head the checkpoint names. That zero does not mean a clean checkpoint.

Limits

The verifier does not say who changed the byte or when, and it cannot restore it. This run used a throwaway signing key and did not touch the database path that sequences events under a lock.

What it came to

The auditor can state one thing: the stored bytes at sequence 1 no longer match the hash written for that row. Showing that nobody rewrote the whole chain takes a copy of the head held off the box, which this POC does not have.

How a decision gets made

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.

  1. 01
    Identity
    In this POC, a password against a local account, with roles read from that account. Active Directory, with no user store in the product, is the design target and is not built.
  2. 02
    Policy
    One deterministic evaluator with a fixed precedence: exception > rollout stage > policy rule action.
  3. 03
    Reveal
    Opening sealed evidence takes at least two distinct enrolled officers, and neither can be the person who asked. Their keys are software keys in this POC.
  4. 04
    Record
    Append-only and hash-chained, with signed checkpoints. The verifier names the first row that fails.

What the model adds

Nothing. Policy evaluation, sealing and the chain run no model. A detection from a model-backed tier can reach the evaluator as an input, and the evaluator itself does no inference.

What it cannot claim

It cannot claim directory identity, hardware custody or third-party anchoring. Someone with root on the appliance could reach the signing key, rewrite the chain and sign a new checkpoint. Only a copy of the head held off the box would expose that, and no external system holds one in this POC.

What it will never do

  • Never stores evidence as plaintext, and never offers a decrypt call on the appliance. It seals, and it relays ciphertext and wrapped keys. Decryption happens client-side, after officers re-wrap the data key.
  • Never opens evidence on one person's approval. A reveal needs at least two distinct enrolled officers, and the analyst who requested it cannot approve it.
  • Never treats a role as a key. Granting security-officer decrypts nothing without an enrolled key holding a share, and a security admin cannot read the officer-key list.
  • Never publishes or rolls back a policy on one person's word. Both are proposed by a maker and approved by a distinct checker under a signed separation-of-duties matrix.
  • Never stops enforcing because a licence lapsed. The verdict path does not read the licence state. The state change is chained.
  • Never claims zero egress at runtime. A gate fails on an undeclared outbound sink, and the console and portal reference no external origin, but those are source checks. Runtime egress has not been verified, and a configured POC model provider calls an external API with synthetic fixtures only.

What evidence it leaves

  • A hash-chained event with its stored canonical bytes, an actor SID and account name where one was resolved, the policy version in force, and an evidence reference where content was sealed. The SIDs are seeded in this POC.
  • A verifier result naming the first failing sequence number and its kind, from the aidom-verify-chain CLI, or from bounded verification over the API for auditors and security admins.
  • Signed checkpoints, exported to file and pushed as CEF over syslog by one declared forwarder. No vendor SIEM has been tested, and the only acknowledger is a stand-in tool.
  • An event export in JSONL or CEF, which can be filtered to one actor, for auditors and security admins. It is a summary without the canonical bytes, so it supports review and not independent recomputation.

See it work

Status and what promotion requires

Available for POC on seeded and synthetic inputs. Not shipping. M0 is committed v1 scope, and production still needs Active Directory and Kerberos sign-in, hardware officer tokens, a vendor SIEM as the off-box anchor, the HA pair, and the appliance image with signed update bundles.

Roles that touch it

Policy AdminSOC AnalystAuditor / CompliancePlatform SysadminSecurity Officer / Key Custodian