Nightly QA findings and incoming change requests. Ash proposes fixes overnight in one pull request per repo; nothing ships until you merge it. A finding stays here until a later review sees the merged fix running. Ash's weekly deep reviews now live in their own Weekly Review tab.
Loading review status...
Nightly QA Findings 0
What could be broken. Ash works this list every night: he reads each finding, confirms it is real, writes the fix and runs the tests, then opens ONE pull request per repo and stops. He does not deploy, and he does not close the finding: it stays open until a later review sees the merged fix running. Where he stopped without a fix, either because it sits in a protected area (money, consent, secrets, infrastructure) or because it needs a decision only you can make, his reasoning is attached to the finding and repeated in the pull request.
Fixed this week 0
Changes Ash's nightly loop made, in plain language. The full history is kept permanently; this shows the last 7 days.
Change requests 0
Weekly Review
Ash's weekly deep reviews (Sundays): continuous improvement, scale & security, autonomy, handoff readiness, donor experience, code optimization, model fit, and SOC 2 posture. Most are advisory (Ash reports, you weigh); code optimization feeds the same overnight fixer as the nightly QA list.
Loading weekly review status...
Continuous Improvement Recommendations 0
Ash's weekly pass (Sundays): what could be better, for our team and for the people we serve. Claire especially. One button hands the whole list to Claude Code to weigh and implement.
Scale & Security 0
Ash's weekly pass (Sundays): barriers to scaling the outward-facing surfaces, and where the stack is open to penetration. Posture read from the code, not live-attack detection. One button hands the whole list to Claude Code to harden.
Autonomy & Automation 0
Ash's weekly pass (Sundays): every point that still forces the SysOp to act by hand, each with a workflow Ash could own to remove it, walking toward near-100% agentic IT support. Items tagged keep-human are deliberate safety gates (funding, award, money, irreversible actions) and stay human by design. One button hands the automatable list to Claude Code to build.
Handoff Readiness 0
Ash's weekly pass (Sundays): what a future outside team (AE Studio) would have to wade through. Code bulkier than the behavior needs (dead paths, duplication, over-built abstraction) and functional details missing from the stack docs. Every shrink is behavior-preserving and every doc gap names the exact detail to add. One button hands the whole list to Claude Code to shrink and document.
Donor Experience 0
Ash's weekly pass (Sundays): does each donor-facing app serve donors and the mission as well as it could, and is the experience shaped so the model could stand on its own as a philanthropic-SaaS product (Claire, the donor surfaces, the IMAD)? Donor-facing apps only. Advisory: Ash reports, you weigh.
Code Optimization 0
Ash's weekly pass (Sundays): where the code does more work than the behavior needs (wasted model calls, recomputation, N+1 queries, blocking work on hot paths). Every change is behavior-preserving and never trades away a safety check. Like QA, these feed the overnight fixer.
Model Fit 0
Ash's weekly pass (Sundays): is each AI agent on the right Anthropic model for its use case and workload? Judged on the experience for donors, recipients, and providers, not on cost. A model change is a one-line config edit per agent. Advisory: Ash recommends, you decide.
SOC 2 Posture 0
Ash's weekly pass (Sundays): the process and infrastructure controls that move the stack toward SOC 2 certification (access, change management, monitoring, retention, vendors, incident response, evidence). Inheris handles PII, never PHI. Complements the code-level Scale & Security pass.
Weekly Studies 0
Focus areas you add on top of the built-in weekly passes, each on its own cadence (ongoing, monthly, one time, or a date window). Use Manage studies to add, edit, pause, or remove them. Findings appear here after a study next runs on a Sunday.
Stack KPIs
The global scorecard: eight metrics that measure Inheris end to end, traced in 7-day increments. A metric whose owning app is not yet instrumented shows Awaiting, never a false zero. Snapshots capture automatically on the weekly run; Refresh now captures one on demand.
Loading KPIs...
Longitudinal trend
One KPI at a time, traced week over week across a rolling 52-week window as data allows. Set an upper and lower control limit to draw the in-control band; any weekly point outside it is flagged. Rates read 0–100%; counts scale to their own range.
Compute
What Inheris pays to run the fleet: AWS to keep it standing, Anthropic tokens and ElevenLabs voice to run its agentic workflows, and GitHub Actions minutes to build and ship it. Fleet-operation cost only, not the SysOp's Claude or Claude Code subscription. Snapshots capture nightly; AWS is the account's real billed cost, Anthropic is priced from published rates, ElevenLabs from an operator-set rate.
Loading compute...
Total spend $0
AWSAnthropicElevenLabsGitHub Actions
AWS $0
The account's own billed cost, this month to date.
GitHub Actions $0
Measured from the runs themselves, so it is this month to date rather than an invoice.
By app
Which app or agent is spending. Ash's own review passes are usually the largest of the agents, since they read the whole fleet's source. AWS is one shared line rather than a share per app: the VPC, the load balancers and the EC2 hosts are shared (one box carries RDS, IME, IOC and Ash), and no cost allocation tag would split them honestly.
HIL Gate Register
Every point in the stack where a human must act. Column A is the standing inventory. Column B is a measurable pathway to remove the human; Column C is the real risk of doing so. Ash drafts B and C; you curate them. The goal is near-total autonomy without exposing donors, recipients, or providers to AI slop, an irreversible mistake, or needless friction — so money, award, publish, delete, and access gates stay human by design.
Questions where the library came up short. This is the backlog of docs to write next.
Loading...
Review scheduling
Pause the automatic review for its next run or a stretch of them. The manual "Run review now" button still works while paused, and a skip clears itself once its dates pass.
Loading...
Weekly studies
Add a focus area for Ash to review each week. Give it a name and say what to look at, then pick how long it runs. Ash reviews it per app on the weekly (Sunday) gate and files what it finds under Weekly Studies. Studies are advisory: Ash reports, you weigh.
Name
What should Ash look at?
Cadence
Apps
From
To
Loading...
Agent Admin
Every Inheris agent's specification, objectives, and scorecard in one place. Ash is editable here; the others mirror their home app (read-only for now).
Things not to say
One global list that applies to every agent. Each phrase is injected at the very top of every agent's prompt and overrides all other wording guidance. Only add a phrase when it is 100% wrong for every agent (for example, “If I’m being honest”).
Loading…
Specification
Ash's role and how he works. This is layered under his built-in persona and safety rules; it refines him and never overrides them.
Performance scorecard
Measured from Ash's real work. Capture a snapshot to also run the model review of his support quality.
Objectives (KPIs / OKRs)
The dimensions Ash is measured on and steers toward. These fold into his live prompt as objectives.
Automated run focus areas
Ash's nightly and weekly reviews. Each pass runs on the rubric below, pre-populated with the current specification. Edit it to steer a pass, or reset to restore the shipped default. Edits change the live prompt for that pass.
One-time and time-limited checks
Add an extra instruction for the next run or for a date window, folded into a pass on top of its rubric. One-time checks are consumed after they fire; windowed checks expire on their end date.
IT skill library
Upload IT knowledge (PDF, Word, PowerPoint, Excel, Markdown, or text, up to 15 MB) to add to Ash's agentic skill set. The text is folded into his live answers; toggle a file out of the prompt without deleting it.
Audits
Every app's append-only audit trail in one place, pulled live from each system. Read-only. Pick an app below.
