Recommendation: Rebuild. Salvage the knowledge, not the machinery.
Handholding Ratio: not measurable from a capture alone — requires the
interview tier to convert stalls and interventions into hours per completed item.
(Reported basis: structural evidence only.)
The three biggest causes
1 · Authority system answers "may I?" — nothing answers "what is done?"Deny-by-default rules with tamper-evident hashes exist; nothing defines a finished task.
Stopping early violates no rule — so the safe behaviour is to stall, and a person pushes
everything through.
2 · No machine-checkable definition of "finished" anywhereNo acceptance criteria, no test gates, no named-output contracts. "Done" is decided by
a person looking at things — so the system can stop early and nothing detects it.
3 · Scheduled jobs failing repeatedlyWork the system scheduled for itself is erroring on schedule, not completing —
and nothing escalates it.
Scores — the ten checks
Completion0/10
Handholding5/10
Idle Burn5/10
Approval Drag5/10
Definition of Done2/10
Invented Work7/10
Visibility6/10
Editability7/10
Resilience0/10
Lock-in & Exposure0/10
Where the work stops
The same task under both regimes. The difference is not intelligence —
it is a checkable definition of finished.
Selected findings (of 20)
Severity
Finding
CRITICAL
Unattributed service listening on all interfaces
CRITICAL
No offsite backup tooling evidenced anywhere in the capture
CRITICAL
Acceptance criteria verify shape, not substance — owners must supply the rest
HIGH
Scheduled jobs failing repeatedly in the last 48 hours
HIGH
Five persistent agent personas with independent state
HIGH
Single alerting and control channel
HIGH
Source projects not under version control
HIGH
Architecture document describes design but no operational metrics
Every finding is evidenced
The full report traces each finding to a specific file in the captured system,
marks everything Verified (measured) vs Reported (described), and states plainly
what this access tier could not assess. Reproducible: the same capture re-scored yields the
same report. Read the raw generated output →