# Cyclopes

> Cyclopes is Boltcliq observability and logging for decisions: business and AI observability, decision records and audit.

Canonical URL: https://www.boltcliq.com/products/cyclopes
Part of: https://www.boltcliq.com/llms.txt

See, explain and improve what your organization and its agents actually do. Cyclopes is observability and logging for modern decision making, covering business and AI observability, reported in business terms.
- Open: telemetry on open standards, starting with OpenTelemetry
- Every decision: recorded with its inputs, rule or model, outcome and approver
- Business terms: alerts tied to ontology objects, such as 'Lagos branch approvals slowed'
- Observes only: never blocks an action; Odin and Monolith do that
- When someone asks what happened and why, the answer should already be on record.
### Overview
- Traditional monitoring tells you that a server is slow. It does not tell you that loan approvals at one branch have slowed, or that a support agent gave a customer the wrong refund policy. Business teams look at one set of dashboards, engineers look at another, and neither can explain a single consequential decision from start to finish.
- AI makes the gap wider. A model call has a prompt, retrieved documents, tool calls, an answer, a latency and a cost. A good answer and a bad one can look identical in a standard log. If an agent acts, someone will eventually ask what it saw, what it chose and why. Without a record that was designed for the question, the only answer is a shrug.
- Cyclopes is built for that question. Applications, workflows and agents emit logs, metrics, traces and decision events to one append-only stream. Technical telemetry is tied to Zeus objects, so an alert reads in business terms. Every consequential decision, by a person or by an AI, is stored with its inputs, the rule or model used, the outcome and the approver. Cyclopes observes and never decides. Blocking an action is the job of Odin and Monolith.
### Cyclopes in one picture
- See, explain and improve what your organization and its agents actually do. Cyclopes is observability and logging for modern decision making, covering business and AI observability, reported in business terms.
- A decision record for every consequential action, by a person or an agent.
- Traces of what agents read, called and answered, with cost per task.
- Audit views and exports that an examiner or oversight body can read.
- Schematic of Cyclopes. Illustrative, not a product screenshot.
### How it works
**Collect from everywhere**
- Applications, workflows and agents send logs, metrics, traces and decision events to one stream using open standards. Existing tools can keep running beside it. The stream is append-only, so a record cannot be quietly changed afterwards.
- Every approval emits a decision record.
**Make agents inspectable**
- For each agent task Cyclopes keeps the prompt, the documents retrieved, the tools called, the output, the latency and the cost. You can open one task and see exactly what happened. Quality evaluations and drift checks run over the same data.
- Which documents did the agent read before it answered?
**Tie it to the business**
- KPIs are bound to Zeus objects, so a technical signal becomes a business statement. Alerts name the branch, the product line or the trade lane, not a host name. They can start a workflow in Zeus so the right person acts.
- Loan approvals at the Kano branch slowed this week.
**Prove it**
- Audit views show who or what did what, when and under which rule. Retention policies decide how long records are kept. Exports are designed to be regulator-ready, with formats agreed with the regulator in scope.
- All credit decisions at one branch for a quarter, with approvers.
**Replay and improve**
- Before you change an agent, test the new version on real past cases and compare the results with what happened. You see where it is better and where it regresses. Only then is it promoted.
- Run last month's 500 cases through v1.3 and compare.
### Capabilities
**Telemetry**
- Logs, metrics and traces from applications, workflows and agents, collected on open standards such as OpenTelemetry. Cyclopes can receive from existing monitoring tools as well as from Boltcliq products. A common format means you are not locked in.
- OTLP ingestion
- Logs, metrics and traces
- Works beside existing monitoring
- You should not have to rebuild your monitoring to start seeing decisions.
**AI observability**
- Prompts, retrieved context, tool calls, outputs, latency and cost per task, with quality evaluations and drift detection. You can follow one answer back to the documents behind it. Cost can be attributed per team, agent and task.
- Prompt, retrieval and tool-call traces
- Cost and latency per task
- Quality evaluation and drift
- An agent that is quietly getting worse, or quietly getting expensive, is a risk nobody sees.
**Decision records**
- A structured record for each consequential decision: the inputs, the rule or model that was used, the outcome and the approver. The same format covers a credit officer, an automated rule and an agent. A record can be opened later and read without special knowledge.
- Inputs, rule or model, outcome, approver
- One format for people and agents
- Linked to Zeus objects
- When a decision is questioned, the record is the evidence.
**Business observability**
- KPIs tied to ontology objects, so alerts speak the language of the business. A slowdown is reported against a branch, a depot or a district, with the supporting traces one click away. Operational teams and engineers can look at the same incident.
- KPIs bound to Zeus objects
- Alerts in business language
- Link from KPI to underlying traces
- Business people do not read host metrics. They read branches, shipments and customers.
**Audit and compliance**
- Immutable audit views, retention policies and exports for regulators and oversight bodies. Access to logs is controlled by Odin, since logs often hold sensitive content. Formats for exports are confirmed with design partners and the regulator in scope.
- Append-only audit views
- Configurable retention and erasure
- Regulator-ready export (design goal)
- Regulators ask for evidence, not for assurances.
**Evaluation and replay**
- Run a new agent version, prompt or policy against past cases before it goes live. Compare outcomes, cost and errors with what actually happened. Evaluation sets can be built from real decision records.
- Replay past cases
- Compare versions side by side
- Evaluation sets from real records
- It is cheaper to find a regression on last month than on a customer.
**Alerting and incident workflow**
- Alerts on thresholds, anomalies and quality drops. An alert can start a workflow in Zeus, so a named person is asked to act and the response is itself recorded. Cyclopes never blocks the underlying action. It raises the issue and keeps the evidence.
- Threshold and anomaly alerts
- Workflow routing through Zeus
- Response recorded with the alert
- An alert nobody owns is noise.
**Telemetry storage at scale**
- Log volume will exceed what a transactional database should carry. This is a design intention that will be confirmed as the product is built.
- Columnar and object storage
- Tiered retention
- Runs on Argus in your boundary
- Observability that is too expensive to keep on is not observability.
### A decision record, in plain terms
- A log line says that something happened. A decision record says that someone or something chose, and gives the evidence. It has the inputs the decider saw, the rule or model version they used, the outcome and the person who approved it. Because it is built on Zeus objects, it also says which loan, which branch and which customer it concerned.
- The same format covers a credit officer, an automated rule and an AI agent. That means a single review process can look at all three. It also means that when you change a rule or an agent, you can test the new version on the decisions already on record.
- Inputs: what the decider saw.
- Method: the rule, model or person, with version.
- Outcome and approver.
- Links to the business objects concerned.
### Use cases across Africa
- Banks
- A bank approves loans and flags suspicious transactions through a mix of staff, rules and a scoring model. Examiners ask for the basis of specific decisions. The answers sit in several systems and emails.
- Cyclopes stores a decision record for each credit and fraud action, with the inputs, the rule or model version and the approver. It watches approval times by branch and alerts when one slows. Records can be prepared for an examiner request.
- Risk and compliance decide which decisions must be recorded in detail and how long to keep them. Branch heads decide what to do about a slowdown.
- Telecoms
- An operator runs marketing campaigns, network operations and customer support, with AI assistants in each. Costs are growing and nobody can say which campaign or assistant is responsible.
- Cyclopes links network and customer events to business outcomes. It attributes AI cost per task, campaign and team, and shows where an assistant gives poor answers. Alerts name the campaign or region.
- Commercial leaders decide which campaigns justify their cost. Engineering decides whether to change an assistant or its model.
- Agriculture
- Field officers send records from phones with patchy connectivity. An advisory assistant answers farmers in several languages. Records go missing and answer quality varies by language.
- Cyclopes monitors the health of field-data sync, showing which areas have not reported. It measures the advisory assistant's answer quality by language and flags drift.
- Programme managers decide where to send support. Agronomists decide whether an assistant's answers are acceptable and when to retrain or restrict it.
- Government
- A ministry uses automated tools to triage applications. Oversight bodies and the public expect to know how decisions were made, and who was responsible.
- Cyclopes keeps an evidence trail for each automated step and each human decision, with the rule version. Audit views give oversight bodies read access under Odin policy. Replay can test a change to a triage rule on past cases.
- Officers decide each case and own the outcome. Oversight bodies decide what to examine and what changes to demand.
### Interfaces
- OpenTelemetry (OTLP) ingestion
- Event bus subscription
- Decision-record API
- Query API and SQL
- Audit export
- File formats to be confirmed.
- Alert routes into Zeus workflows
- Dashboards
- SDKs
- Existing monitoring tools
- Receive from or sit beside, via open standards. Specific tools confirmed with design partners.
- The lists describe target interfaces and are confirmed per release and per design partner. Nothing on this list is a claim that an integration with a named third party exists today.
### Security and sovereignty
**Sensitive logs, controlled**
- Logs can contain personal and commercial data. Access is decided by Odin and redaction rules apply.
**Append-only**
- The audit stream is designed so that records cannot be silently edited or removed.
**Inside your boundary**
- Telemetry is stored on Argus, in your boundary, in line with your residency rules.
**Retention and erasure**
- Retention is configurable. Erasure follows policy where the law requires it, with the effect on audit records made explicit.
### What Cyclopes does and does not do
- Record logs, metrics, traces and decision events.
- Trace agent prompts, retrieval, tool calls, cost and quality.
- Tie technical signals to business objects and KPIs.
- Provide audit views, exports, evaluation and replay.
- Alert and route incidents into workflows.
- Decide or block actions. Odin and Monolith do that.
- Define business meaning. Zeus does.
- Run agents or hold their memory. Argus and Monolith do.
- Replace your existing monitoring unless you choose to.
- Claim regulator acceptance of any export format today.
### Questions
- How is Cyclopes different from the business intelligence in Zeus?
- Zeus answers what the state of the business is and how it should work. Cyclopes answers what systems and agents actually did and whether it was any good. Zeus sends decision events to Cyclopes, and Cyclopes sends alerts back to Zeus workflows.
- Can Cyclopes stop a bad action?
- No, by design. Blocking belongs to Odin and Monolith. Cyclopes can alert, and it can start a workflow so that a named person responds.
- Does it replace our existing monitoring?
- It does not have to. It uses open standards, so it can sit beside your current tools or receive data from them. Which specific tools are supported will be confirmed with design partners.
- How are sensitive logs protected?
- Access is decided by Odin, redaction rules are applied, and the data is stored in your boundary on Argus. Retention and erasure are configurable.
- Can we replay old cases against a new agent version?
- That is the purpose of evaluation and replay: you test a change on real past cases, compare it with what happened and only then promote it.