← AEI Scan on Prosights.ai. This line is not part of the report: the website adds it.

Kyberna: AEI 79, Level 1

This is the report AEI Scan 0.1.0 wrote on 2026-11-16 for the repository of Kyberna, the product Prosights builds. The coding agent is Claude Code. Prosights ran the scan and gave the answers. On this page the names of single rule, skill and agent files, of the worktree folders and of 1 tool are replaced by a pattern and a count; every other line is as the tool wrote it. 81 of the 97 checks come from the process used to build Kyberna, so read this as an example of a report, not as validation of the index.

Summary

Your result: AEI 79, Level 1.

AEI Level 1, WrittenOverall level
78.7 of 100Overall score
95 of 97 checks (98%)Coverage
0Claim gap

What the level means: The practice is written down; nothing checks that the agent follows it.

What to do next

#DimensionPut in placeWhere, or who confirms it
1Safety controlsIn every hook on PreToolUse, set onFailure to block..claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json
2Review and mergeAdd a CODEOWNERS file.1 of: .github/CODEOWNERS, CODEOWNERS, docs/CODEOWNERS
3Review and mergeAdd an automated review that runs on every change before it merges and cannot be skipped.Then show it to the assessor
4Memory and continuityHave each agent record that it read the carried-forward lessons before its work.Then show it to the assessor
5MeasurementMeasure the merge rate and the revert rate of agent changes each quarter, against your own baseline.Then show it to the assessor

These are the first 5 of 10 actions. The appendix lists all of them, with the check behind each.

What the result rests on

95 of 97 checks were graded: 33 by the scan, 16 by your survey answers and 46 by what a person was shown. The answers are your own, and nobody checked them unless an assessor asked for the artifact. The scan shows that a control is present, not that it works.

2 checks could not be graded, because the question was not answered or the scan could not read the evidence: 12.1, 12.4. Answer the questions and run the tool again to raise the coverage.

The score is the share of the weighted evidence that was found; it is not a percentile. The claim gap counts the checks where your answer was yes and the scan found nothing.


The 16 dimensions

#DimensionLevelHighest level possibleScoreChecks graded
1Requirements and recording3, Enforced31004 of 4
2Decisions2, Checked21003 of 3
3Expected behavior and design2, Checked21002 of 2
4Agent briefing and model policy3, Enforced31006 of 6
5Build3, Enforced31004 of 4
6Test and evidence2, Checked4566 of 6
7Review and merge1, Written4477 of 7
8Release4, Measured41003 of 3
9Safety controls4, Measured4868 of 8
10Memory and continuity1, Written2786 of 6
11Measurement1, Written4535 of 5
12Governance1, Written3252 of 4
13Token and cost3, Enforced48812 of 12
14Harness depth3, Enforced31007 of 7
15Process guards3, Enforced310011 of 11
16Model conduct1, Written4279 of 9

Select the title of a part of the appendix to open it. Every part prints open.

At most 10, in this order: safety controls first, then whatever holds each dimension at its level, starting with the lowest. An action a scan can confirm comes before one a person must confirm.

#CheckPut in placeWhere, or who confirms it
19.5 (level 3)In every hook on PreToolUse, set onFailure to block..claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json
27.2 (level 2)Add a CODEOWNERS file.1 of: .github/CODEOWNERS, CODEOWNERS, docs/CODEOWNERS
37.3 (level 2)Add an automated review that runs on every change before it merges and cannot be skipped.Then show it to the assessor
410.6 (level 2)Have each agent record that it read the carried-forward lessons before its work.Then show it to the assessor
511.2 (level 4)Measure the merge rate and the revert rate of agent changes each quarter, against your own baseline.Then show it to the assessor
611.4 (level 4)Keep 1 quality trend (warnings, complexity or failures) that covers at least 3 months.Then show it to the assessor
712.2 (level 3)Put an organization-level policy for the agent in force on developers' machines.1 of: /Library/Application Support/ClaudeCode/managed-settings.json, /Library/Application Support/ClaudeCode/managed-settings.d/*.json
816.4 (level 2)Register a hook on the event PermissionDenied..claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json
96.3 (level 3)Register a hook on the event Stop..claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json
1013.12 (level 4)Set env.CLAUDE_CODE_ENABLE_TELEMETRY to 1.~/.claude/settings.json

Other open scan checks on Claude Code

The scan grades these checks too, and they are open. Every file name, key and hook event is the one the vendor documents for Claude Code. Where the vendor documents none, the row names no file.

CheckLevelResultPut in placeWhere
16.73FailRegister a hook on the event SubagentStop or TaskCompleted..claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json

Every practice that passed, sorted by the kind of control behind it. A written rule depends on the agent choosing to follow it; a checked rule is caught after the agent has acted; an enforced rule is stopped before.

KindCheckWhat is in placeGraded byWhere it was found
Written1.1Work items are written down before an agent starts on them.survey answer
Written2.1Decision records are kept in a folder of the repository.scandocs/decisions/
Written2.3It is written which choices a person makes and which an agent may make.survey answer
Written3.1The expected behaviors of a change are written before the build.shown to a person
Written4.1The agent has an instruction file.scanCLAUDE.md; ~/.claude/CLAUDE.md
Written4.2The instruction files cover at least 4 of these 6 areas: commands, testing, project structure, code style, git workflow, boundaries.scanCLAUDE.md; ~/.claude/CLAUDE.md
Written4.3Rule files apply only to some paths.scan.claude/rules/*.md (17 files)
Written4.4Written procedures or agent definitions exist for the agent to load.scan.claude/skills/*/SKILL.md (2 files); .claude/agents/*.md (2 files)
Written5.1Each agent works in its own checkout or environment.survey answer
Written5.4The size of 1 agent task has a written limit.survey answer
Written6.1The instruction files say how tests are run.scanCLAUDE.md; ~/.claude/CLAUDE.md
Written7.1A reviewer other than its author reviews every agent change before it merges: a person, or a model run as a separate reviewer.survey answer
Written7.5It is written who may start or assign an agent.survey answer
Written8.1A person confirms each release.survey answer
Written9.7The instruction files say what the agent must never do.scanCLAUDE.md; ~/.claude/CLAUDE.md
Written10.1The agent keeps memory files for this project.scan~/.claude/projects/<project>/memory/
Written10.3The start-of-session procedure is written.survey answer
Written10.4Each agent writes down what it learned from its work, in a place the next agent reads.survey answer
Written11.5Agent work is measured by a named outcome.survey answer
Written12.3The rules have stable identifiers and a record of why each exists.shown to a person
Written13.1It is written which model and which reasoning effort each kind of task uses.survey answer
Written13.2Every instruction file has fewer than 200 lines.scanCLAUDE.md; ~/.claude/CLAUDE.md
Written13.9It is written that long or verbose work goes to its own agent, and that agents wait without polling.survey answer
Written14.2MCP servers are declared in a file or named in a permission rule.scan.claude/settings.json
Written14.4Agents work in worktrees the tool sets up.scan.claude/worktrees/* (2 folders)
Written15.4It is written that a new integration or tool starts with no access until it is listed.survey answer
Written15.8It is written that a change touches only what its own work item needs.survey answer
Written16.1A written record shows how each model in use has failed in the team's own work.survey answer
Written16.2Every brief to an agent states what done means.survey answer
Checked1.2A check refuses a change that is tied to no work item.shown to a person
Checked1.3A script regenerates the list of open work.shown to a person
Checked2.2A check reads the decision records at commit.shown to a person
Checked3.2A check ties each expected behavior to a test.shown to a person
Checked5.3A limit stops a build or test run that hangs.shown to a person
Checked6.2A git hook or a CI workflow is in the repository.scan.git/hooks/pre-commit; .git/hooks/pre-merge-commit
Checked6.6A gate fails on new warnings in changed files.shown to a person
Checked7.4A merge step builds and checks after each merge and stops on failure.shown to a person
Checked8.2A release gate refuses when a required item is missing.shown to a person
Checked9.6A secret scan runs before code leaves the machine.shown to a person
Checked10.2An end-of-session step refuses to finish with failing tests or unpushed work.shown to a person
Checked10.5A mechanism refuses a change whose agent recorded no lesson entry.shown to a person
Checked13.8A mechanism shortens long command output before the agent reads it.shown to a person
Checked14.5A command shows the state of each background or unattended run.shown to a person
Checked14.7The team's own hooks and guard scripts have tests.shown to a person
Checked15.3A check keeps credentials and real messages out of test fixtures.shown to a person
Checked15.5A mechanism stops a test from reaching a live account unless a person turns that on.shown to a person
Checked15.6A commit check refuses personal identifiers and authorship marks.shown to a person
Checked15.7A set of injection attempts is run against a model before it is given a channel.shown to a person
Checked15.9A guard has a single switch, and its off state is tested.shown to a person
Checked16.3Someone other than the agent checks the claims in its reports.shown to a person
Checked16.5Refused calls are logged, each with the model that made the call.shown to a person
Enforced1.4A hook runs when a prompt is submitted to the agent.scan.claude/settings.json: UserPromptSubmit
Enforced4.5The agent's model is named with a version number.scan.claude/settings.json (env); .claude/agents/*.md (2 files)
Enforced4.6A hook runs when an agent starts another agent.scan.claude/settings.json: PreToolUse on Agent
Enforced5.2A hook runs before the agent's shell commands.scan.claude/settings.json: PreToolUse on Bash
Enforced7.6A written rule says which changes need a person's approval before they merge or ship, by risk.shown to a person
Enforced9.1The agent's sandbox is on.scan~/.claude/settings.json
Enforced9.2The agent's network access is limited to a list.scan~/.claude/settings.json
Enforced9.3Permission rules exist, and no setting approves everything.scan.claude/settings.json; .claude/settings.local.json; ~/.claude/settings.json
Enforced9.4A hook can deny a tool call.scan.claude/settings.json: PreToolUse on Agent; .claude/settings.json: PreToolUse on Bash; .claude/settings.json: PreToolUse on an MCP tool; .claude/settings.json: PreToolUse on Write|Edit|NotebookEdit
Enforced13.3A threshold is set at which the conversation is compacted.scan.claude/settings.json
Enforced13.4The reasoning effort is set in settings or in an agent definition.scan~/.claude/settings.json
Enforced13.5Agents that another agent starts are given their own model.scan.claude/settings.json; .claude/agents/*.md (2 files)
Enforced13.6A prompt cache lifetime is set, and caching is on.scan.claude/settings.json
Enforced13.10A spend or usage limit is set at the provider.shown to a person
Enforced14.1A hook runs before the agent writes or edits a file.scan.claude/settings.json: PreToolUse on Write|Edit|NotebookEdit
Enforced14.3An agent definition limits the tools its agent may use.scan.claude/agents/*.md (1 file)
Enforced14.6An agent runs commands on another machine only from an allow list.shown to a person
Enforced15.1Every delete goes through 1 script, and a hook refuses a delete made any other way.shown to a person
Enforced15.2A hook refuses a build or test command run outside the team's wrapper script.shown to a person
Enforced15.10A script removes an agent's workspace and refuses when its work is not merged.shown to a person
Enforced15.11No command can run outside the sandbox.scan~/.claude/settings.json
Measured6.4The test results of each build are kept as files.shown to a person
Measured8.3The gate result of each release is recorded.shown to a person
Measured9.8A list marks each of the team's rules as enforced, checked or written only.shown to a person
Measured11.1Session records of the last 30 days are kept.scan~/.claude/projects/<project>/<session>.jsonl
Measured11.3Usage or cost is recorded per agent or task.shown to a person
Measured13.7A report separates input tokens read from the cache from input tokens sent uncached.shown to a person
Measured13.11Tokens or cost per model are reviewed over time, and a change was made because of a review.shown to a person

Said but not found

None.

In place but not known to the team

None.

Dimension 16, model conduct: your counts for each model

You gave no counts. To fill this table, count from your own records, for each model your agents run on and for 1 period, how often it did each of these: rules broken; stopped early or asked for a permission it had; reported done when it was not; invented a reason; did work nobody asked for. Put 1 row for each model and period under model_conduct in the answers file (SCHEMA.md). Check 16.8 passes when an assessor is shown these counts for 2 periods.

The scan looked at these locations and at nothing else. It read settings, hook registrations, front matter and the headings and command lines of instruction files. It copied no file contents and no setting values into this report.

Questions answered: 16 survey, 0 on what you believe is in place, 47 shown to an assessor.

CheckLevelGraded byResultWhat the check readsWhat was found
1.11survey answerPassYour answer to the survey question "Are work items written down before an agent starts on them?"the answer was yes
1.22shown to a personPassThe check that refuses a change tied to no itemthe artifact was shown
1.32shown to a personPassA list of open work that a script regeneratesthe artifact was shown
1.43scanPassA hook on the prompt-submit event (Claude Code and Codex CLI UserPromptSubmit; VS Code userPromptSubmitted; on other platforms, shown to a person).claude/settings.json: UserPromptSubmit
2.11scanPassThe decision-record folder you namedocs/decisions/
2.22shown to a personPassThe check that reads decision records at committhe artifact was shown
2.31survey answerPassYour answer to the survey question "Is it written which choices a person makes and which an agent may make?"the answer was yes
3.11shown to a personPassFor 1 recent change: the expected behaviors written before the buildthe artifact was shown
3.22shown to a personPassThe check that ties each expected behavior to a testthe artifact was shown
4.11scanPassThe agent's instruction filesCLAUDE.md; ~/.claude/CLAUDE.md
4.21scanPassHeadings and command lines of the instruction filesCLAUDE.md; ~/.claude/CLAUDE.md; 5 of 6 areas found: commands, testing, code style, git workflow, boundaries
4.31scanPassRule files that apply only to some paths.claude/rules/*.md (17 files)
4.41scanPassProcedure files and agent definitions.claude/skills/*/SKILL.md (2 files); .claude/agents/*.md (2 files)
4.53scanPassThe model named in settings and in agent definitions.claude/settings.json (env); .claude/agents/*.md (2 files)
4.63scanPassA hook on the tool that starts another agent (Claude Code PreToolUse on Agent; on other platforms, shown to a person).claude/settings.json: PreToolUse on Agent
5.11survey answerPassYour answer to the survey question "Does each agent work in its own checkout or environment?"the answer was yes
5.23scanPassA hook before shell commands (Claude Code PreToolUse on Bash; Cursor beforeShellExecution; on other platforms, shown to a person).claude/settings.json: PreToolUse on Bash
5.32shown to a personPassThe limit that stops a build or test run that hangsthe artifact was shown
5.41survey answerPassYour answer to the survey question "Is there a written limit on the size of 1 agent task?"the answer was yes
6.11scanPassCommand lines of the instruction filesCLAUDE.md; ~/.claude/CLAUDE.md
6.22scanPassA file in .git/hooks/ that is not a sample, or a workflow file in .github/workflows/.git/hooks/pre-commit; .git/hooks/pre-merge-commit
6.33scanFailA hook on the end of the agent's turn (Claude Code and Codex CLI Stop; Cursor stop; VS Code agentStop; on other platforms, shown to a person)looked in: .claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json; no hook registered on Stop
6.44shown to a personPassTest results kept as files for each buildthe artifact was shown
6.54shown to a personFailFor each function you name as crucial: a test set run at every releasethere is no artifact
6.62shown to a personPassThe gate that fails on new warnings in changed filesthe artifact was shown
7.11survey answerPassYour answer to the survey question "Is every agent change reviewed before it merges by a reviewer other than its author: a person, or a model run as a separate reviewer?"the answer was yes
7.22shown to a personFailA CODEOWNERS file; when none is found, branch protection shown to a personlooked in: .github/CODEOWNERS, CODEOWNERS, docs/CODEOWNERS; no CODEOWNERS file; branch protection is not on disk and is shown to a person; there is no artifact
7.32shown to a personFailAn automated review that runs on every change before it merges and cannot be skippedthere is no artifact
7.42shown to a personPassThe merge step that builds and checks after each merge and stops on failurethe artifact was shown
7.51survey answerPassYour answer to the survey question "Is it written who may start or assign an agent?"the answer was yes
7.63shown to a personPassThe written rule that says which changes need a person's approval before they merge or ship, by riskthe artifact was shown
7.74shown to a personFailThe record of automated approvals, and the sample of them a person reviewed in the last quarterthere is no artifact
8.11survey answerPassYour answer to the survey question "Does a person confirm each release?"the answer was yes
8.22shown to a personPassThe release gate that refuses when a required item is missingthe artifact was shown
8.34shown to a personPassThe record of each release's gate resultthe artifact was shown
9.13scanPassThe sandbox setting~/.claude/settings.json
9.23scanPassThe setting that limits network access~/.claude/settings.json
9.33scanPassThe permission rules in the agent's settings.claude/settings.json; .claude/settings.local.json; ~/.claude/settings.json
9.43scanPassA hook that can deny a tool call (Claude Code and Codex CLI PreToolUse; Cursor preToolUse, beforeShellExecution, beforeMCPExecution, beforeReadFile; VS Code preToolUse; on other platforms, shown to a person).claude/settings.json: PreToolUse on Agent; .claude/settings.json: PreToolUse on Bash; .claude/settings.json: PreToolUse on an MCP tool; .claude/settings.json: PreToolUse on Write|Edit|NotebookEdit
9.53scanFailWhat happens when the hook itself failslooked in: .claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json; 0 of 4 hook entries have onFailure set to block
9.62shown to a personPassThe secret scan that runs before code leaves the machinethe artifact was shown
9.71scanPassLines of the instruction files that hold "never", "must not" or "do not"CLAUDE.md; ~/.claude/CLAUDE.md; 35 boundary lines found
9.84shown to a personPassA list of the team's rules marked enforced, checked or written onlythe artifact was shown
10.11scanPassHow many files the agent's memory folder holds, never their contents~/.claude/projects/<project>/memory/; 63 files in the folder
10.22shown to a personPassThe end-of-session step that refuses to finish with failing tests or unpushed workthe artifact was shown
10.31survey answerPassYour answer to the survey question "Is the start-of-session procedure written?"the answer was yes
10.41survey answerPassYour answer to the survey question "Does each agent write down what it learned from its work, in a place the next agent reads?"the answer was yes
10.52shown to a personPassThe mechanism that refuses a change whose agent recorded no lesson entrythe artifact was shown
10.62shown to a personFailFor the last 5 agent changes, the record that the agent read the carried-forward lessons before its workthere is no artifact
11.14scanPassThe count and dates of the agent's session records, never their text, and only when you turn the count on~/.claude/projects/<project>/<session>.jsonl; 16 of 22 records are dated in the last 30 days
11.24shown to a personFailMerge rate and revert rate of agent changes against the team's own baselinethere is no artifact
11.34shown to a personPassUsage or cost recorded per agent or taskthe artifact was shown
11.44shown to a personFailA quality trend: warnings, complexity or failures over timethere is no artifact
11.51survey answerPassYour answer to the survey question "Do you measure agent work by a named outcome?"the answer was yes
12.11survey answerUnknownYour answer to the survey question "Is there a written statement of which AI tools are used, on which data, and who approves?"the question was not asked
12.23shown to a personFailThe organization's policy file for the agentlooked in: /Library/Application Support/ClaudeCode/managed-settings.json, /Library/Application Support/ClaudeCode/managed-settings.d/*.json; no file at the documented locations; a policy delivered by a macOS configuration profile, the Windows registry or the vendor's server is not a file and is shown to an assessor; there is no artifact
12.31shown to a personPassRules with stable identifiers and a record of why each existsthe artifact was shown
12.41survey answerUnknownYour answer to the survey question "Is agent use covered by your organization's AI risk process?"the question was not asked
13.11survey answerPassYour answer to the survey question "Is it written which model and which reasoning effort each kind of task uses?"the answer was yes
13.21scanPassThe number of lines in each instruction file, never the linesCLAUDE.md; ~/.claude/CLAUDE.md; the longest instruction file has 194 lines
13.33scanPassThe setting for compacting the conversation.claude/settings.json
13.43scanPassThe reasoning-effort setting~/.claude/settings.json
13.53scanPassThe model set for agents that another agent starts.claude/settings.json; .claude/agents/*.md (2 files)
13.63scanPassThe prompt-cache settings.claude/settings.json
13.74shown to a personPassA report that separates input tokens read from the cache from input tokens sent uncachedthe artifact was shown
13.82shown to a personPassA mechanism that shortens long command output before the agent reads itthe artifact was shown
13.91survey answerPassYour answer to the survey question "Is it written that long or verbose work goes to its own agent, and that agents wait without polling?"the answer was yes
13.103shown to a personPassA spend or usage limit set at the provider, for the workspace or for each personthe artifact was shown
13.114shown to a personPassA review of tokens or cost per model over time, and 1 change made because of itthe artifact was shown
13.124scanFailThe usage-telemetry settinglooked in: ~/.claude/settings.json; the setting is not set
14.13scanPassA hook that runs before the agent writes or edits a file (Claude Code PreToolUse on Write or Edit).claude/settings.json: PreToolUse on Write|Edit|NotebookEdit
14.21scanPassThe MCP server declarations and the permission rules.claude/settings.json
14.33scanPassThe tool limits in agent definitions.claude/agents/*.md (1 file)
14.41scanPassWorktree folders and the worktree setup file.claude/worktrees/* (2 folders)
14.52shown to a personPassThe command that shows the state of each background or unattended runthe artifact was shown
14.63shown to a personPassThe channel through which an agent runs a command on another machinethe artifact was shown
14.72shown to a personPassThe tests of the team's own hooks and guard scriptsthe artifact was shown
15.13shown to a personPassThe script every delete goes through, and the hook that refuses a delete made any other waythe artifact was shown
15.23shown to a personPassThe hook that refuses a build or test command run outside the team's wrapper scriptthe artifact was shown
15.32shown to a personPassThe check that keeps credentials and real messages out of test fixturesthe artifact was shown
15.41survey answerPassYour answer to the survey question "Is it written that a new integration or tool starts with no access until it is listed?"the answer was yes
15.52shown to a personPassWhat stops a test from reaching a live account unless a person turns that onthe artifact was shown
15.62shown to a personPassThe commit check that refuses personal identifiers and authorship marksthe artifact was shown
15.72shown to a personPassThe set of injection attempts run against a model before it is given a channel, and the step that runs itthe artifact was shown
15.81survey answerPassYour answer to the survey question "Is it written that a change touches only what its own work item needs?"the answer was yes
15.92shown to a personPassFor 1 guard: its single switch, and the test of the off statethe artifact was shown
15.103shown to a personPassThe script that removes an agent's workspace and refuses when its work is not mergedthe artifact was shown
15.113scanPassThe setting that lets a command run outside the sandbox~/.claude/settings.json
16.11survey answerPassYour answer to the survey question "Is there a written record of how each model you use has failed in your own work?"the answer was yes
16.21survey answerPassYour answer to the survey question "Does every brief to an agent state what done means?"the answer was yes
16.32shown to a personPassFor the last 2 agent reports: who other than the agent checked the claims, and what they checkedthe artifact was shown
16.42scanFailA hook on a denied tool call (Claude Code PermissionDenied; Cursor postToolUseFailure; on other platforms, shown to a person)looked in: .claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json; no hook registered on PermissionDenied
16.52shown to a personPassThe log of refused calls, each with the model that made the callthe artifact was shown
16.63shown to a personFailThe mechanism that refuses the end of a turn while tasks are open, and 1 refusal it madethere is no artifact
16.73scanFailA hook on the end of a started agent's work or on a task being marked complete (Claude Code SubagentStop, TaskCompleted; Codex CLI SubagentStop; Cursor subagentStop; VS Code subagentStop; on other platforms, shown to a person)looked in: .claude/settings.json, .claude/settings.local.json, ~/.claude/settings.json; no hook registered on SubagentStop, TaskCompleted
16.84shown to a personFailCounts per model for 2 periods (rule breaks, early stops, false "done" reports) and the schedule on which they are readthere is no artifact
16.94shown to a personFailThe measurement made before the last new model version was adopted, and the line of the model policy it changed or confirmedthere is no artifact

LevelNameWhat it takes
AEI Level 0UnrecordedNo evidence for the practice
AEI Level 1WrittenThe practice is written where the agent or the team reads it
AEI Level 2CheckedA mechanism checks the result at commit, merge or release, and refuses
AEI Level 3EnforcedA mechanism refuses the action before it happens
AEI Level 4MeasuredRecords of what happened are kept and read