audit

Files (1)

• SKILL.md

SKILL.md (first lines)

# Audit Godspeed Mission Control

This is a structural review of the selected installation: whether the instructions, shared memory, connections, reusable methods and self-running work are actually in good shape. It is not proof that every product capability works. Name missing capabilities separately and keep their behavioral acceptance open.

The four-layer scoring structure is adapted, with credit, from Nate Herk's Four-Cs framework. Each layer has 25 points. The first run is a baseline; later runs compare the same criteria and disclose changes in scope.

## Discover the real installation

Read the workspace operating manual and authorization rules. Confirm the chosen workspace, account, schedule owner, software version and physical host before service operations. A configured flag is not evidence of completed onboarding. Never inspect or operate another installation merely because an SSH alias or old URL resolves.

Inventory equivalent intent, not particular names: substantive operating manual, current profile/voice/rules, memory index and structured claims, decisions, reference guides, skills with frontmatter, configured assistant/provider, peer-agent registry, connection registry, source receipts, actual jobs/receipts, procedures register and kept prompts/templates.

Use the notebook's normal Settings/Recovery controls and scoped MCP reads. The installed file CLI offers validate, record list/get and memory search with GODSPEED_WORKSPACE set to this installation. Inspect tool schemas rather than inventing commands. Query settings, jobs, job_receipts, goals, work_items, deadlines, claims and connection receipts as needed. Inspect skill frontmatter for the inventory; do not read every recipe unless a specific claim depends on it.

Read configured connection names and credential presence only. Never dump keys, environment files, browser sessions or private tokens. A key plus documentation establishes a possible mechanism, not a successful live read. Prefer actual last successful receipts.

## Score the four layers

Record criterion, evidence, points earned and points lost. Missing evidence is unknown, not generous credit. Equivalent storage and names count. A structural score never substitutes for behavioral completion.

### Context: 25

| Criterion | Points | Evidence |
|---|---:|---|
| Substantive operating manual | 5 | Read it; more than 200 words |
| User identity, purpose and voice captured | 5 | Actual shared instructions/profile |
| Persistent memory with multiple entries | 5 | More than three meaningful entries, visible and retrievable |
| At least one reference guide | 5 | Readable source/API/procedure reference |
| At least one recorded decision | 5 | An actual retained decision and reason |

### Connections: 25

The seven universal domains are financial/revenue data, customer interactions, calendar, communication, project/tasks, meeting intelligence, and knowledge/files. MCP, API scripts, documented exports and other tested mechanisms are equivalent. AI-provider access, content publishing and decision history are additional context; they do not invent coverage of a missing universal domain.

| Criterion | Points | Evidence |
|---|---:|---|
| Universal domain coverage | 10 | 1.4 per reachable domain, rounded to nearest 0.5, capped at 10 |
| Connected tools have usable reference guides | 5 | Start at five; subtract one per tool without a guide, floor zero |
| Authentication and pipeline freshness | 5 | Subtract one per broken, needs-login, not-set-up or last-success older than 30 days; one per checker older than three days; floor zero |
| Complete connections registry | 3 | Zero absent, one sparse, two most, three all reachable mechanisms |
| Read and write balance | 2 | An actual authorized write-capable connection; do not execute outward work during an audit |

A registry description date is not an authentication check. A connected row without a corresponding current checker is unchecked. Private credentials are not audit output. Account restrictions and a missing product adapter are different findings.

### Capabilities: 25

| Criterion | Points | Evidence |
|---|---:|---|
| Three or more installed skills | 10 | Actual installed frontmatter and entry points |
| At least one user-authored or customized skill | 10 | Maintained method beyond the shipped audit/design/skill-creation examples |
| At least one reachable peer agent | 5 | Current registry plus verified reachability |

A selected model/provider alone is not a peer-agent mesh. If this installation lacks a peer capability, lose those points honestly. Do not create agents or call peers without applicable authorization. A skill file's presence earns structural inclusion only; unbundled helpers and failed deliverables remain capability gaps.

### Cadence: 25

| Criterion | Points | Evidence |
|---|---:|---|
| At least one recurring trigger | 10 | Actual configured schedule/hook, owner and enabled state |
| Recent activity within 30 days | 10 | Actual decision or completed routine receipt; timestamped modification alone is weaker evidence |
| Kept reusable material | 5 | At least one retained useful prompt/template in the workspace |

Read the procedures register and compare it with actual enabled jobs, next run and latest receipt. More than 30 days since verification, a stopped checker, missing owner or contradictory claims are cadence gaps. A recipe named daily-* is only a discoverable method, not proof its routine runs. Deliberately paused routines stay paused.

## Rank three structural gaps

For each lost criterion compute lost points times impact multiplier:

- Zero universal data domains: 4.
- Missing/thin operating manual: 3.
- At most two universal domains: 3.
- No skills, no recurring trigger, or all connections read-only: 2.
- No connection guides or no decisions log: 1.5.
- Other gaps: 1.

Sort descending; choose at most three actual gaps. Explain the concrete correction and its verification. Prefer the already documented mechanism; do not suggest an unavailable tool. Do not hand the user busywork or quietly change policy during a review. Authorized reversible implementation work is separate from the score; read-only audits leave project files unchanged except an authorized saved report.

## Rules review

When rules exist, read the compiled block and each relevant owning rule before judging overlap. Report count and actual size against the configured limit, if one exists. Never invent a limit. Flag IDs as restated elsewhere, mechanically enforced, frozen incident detail, or over the per-rule cap. Give a reason and proposed surviving wording/home.

Only propose subtraction or compression. Do not retire, renumber, rewrite or delete the user's rules during an audit. Historical explanation stays with its owning rule. If no compiled-rule mechanism exists, report that scope explicitly rather than assuming the personal engine is installed.

## One shared memory

Check each configured assistant's actual loaded memory/profile and write tools. Separate caches and file-backed assistant history from an independent private fact store. Report whether a second memory can silently contradict the shared notebook. Do not sweep another assistant home or company bot.

If a second store is found, preserve its contents, identify the current facts and propose their correct shared home. Disable/reset it only with the user's applicable authorization and its documented command after preservation. Verify configuration, retained files and live prompt together; an empty file today is not proof the model cannot write it tomorrow.

## Do messages explain themselves?

Read the selected installation's last seven days of visible user/assistant conversations and delivery receipts. Count user replies expressing confusion and answers exposing internal paths/decision codes as unnecessary explanation. Report numerator, denominator, date range and missing channels. Compare with the previous same-scope review.

When a count rises, read the actual confusing message and identify its producer and a plain-language correction. A saved draft is not evidence of receipt. A substring counter is a triage aid, not a judgment of meaning; read the message before attributing failure. Never import another account's transcripts to fill the denominator.

## Comparable-system check

Use an existing configured watch topic with dated primary-source evidence. Select one named comparator per ISO week using week number modulo the number of comparator entries, so the rotation needs no independent state. Read its requirements and latest evidence, then compare one capability this installation lacks and one it does better, or say no material difference was established.

An empty comparator log is a watch evidence gap. Do not invent a comparison or silently do a broad new research job. State whether the comparison changes the three ranked gaps.

## Report and verification

Show total /100 and four /25 scores with evidence. Stage ranges: 0-39 foundation, 40-69 built, 70-89 compounding, 90-100 autonomous structure. Labels: at least 20 strong, 15-19 solid, 8-14 thin, below eight missing. These labels describe structure and never claim reliable autonomy or completed product acceptance.

Include one to three evidenced strengths, the ranked gaps with lost points/multiplier, rules review, shared-memory result, message-readability result and comparable-system result. State inaccessible sources and scope. Avoid fictional counters and screenshots as substitutes for execution.

When saving is authorized, retain one dated notebook audit note and open/read it back. Compare to the previous baseline without replacing it. Scheduled runs already authorized to save do so; a read-only manual review asks only when saving falls outside its authorization. Publishing, sending or changing external accounts always keeps its actual approval boundary.

The user gets a short explanation of the practical result and any action only they can take. No internal machinery tour, blanket completeness claim or worksheet.


## Installed personal workspace

Use the current user workspace and its configured providers. Keep original workflow, command contracts, scripts and verification criteria. Read the workspace authorization rules before sends, sign-ins, payments or publishing. Search existing device-private credentials before asking for configuration. Saved output and a passing screen are not evidence that the full requested result happened.

From MichaelZelbel/godspeed-mission-control, MIT.

More from MichaelZelbel/godspeed-mission-control