The value report is switched on one org at a time and is not available to customer orgs yet. Until it is on for your org, the report you get is the email from the command-line scan. See Run Org Scan yourself.
What you are asked for
Two things, both in plain terms:- Install the GitHub Scan app on the repositories you want scanned. It is read-only. It asks for repository metadata, pull requests and checks. It does not ask for your code.
- Connect your Anthropic usage, or type a monthly figure. Connecting lets the report measure your API spend. With no API usage, you can type your approximate monthly AI spend instead. Both are optional. Without either, the report has no cost row.
The period it covers
The report covers the 13 most recent full weeks, in UTC, each starting on a Monday. The week in progress is never included. Trends compare the first four weeks of that window with the last four. The top of the report shows:How to read a row
The report has at most one row for each of five themes: cost, velocity, rework, follow-up and time. A theme with nothing to show has no row. Each row has:- A severity:
high,mediumorwatch. - A headline: one sentence naming the problem.
- A figure: the number behind the headline, marked
measuredorestimate. - Evidence: the counts the figure comes from, each with its label.
- A target, on some rows. See Targets.
- A recommended fix: a Mnemom capability, what it does, the command that starts it, and a link to its docs.
high first. Two rows with the same severity keep the theme order above. The first row is marked as the largest.
Measured and estimate
Every line that holds a dollar amount, a percentage or a number of hours says whether it ismeasured or an estimate.
- Measured figures are counted from GitHub or read from your Anthropic usage reports.
- Estimates are worked out from those counts. A yearly cost appears only on the cost row, as an estimate at the pace of the last four weeks.
- A range is written “between X% and Y%” and is a 95% confidence interval.
- AI-assisted counts are a floor, written “at least”. The scan only knows a pull request is AI-assisted when it carries a mark of an AI tool, so some AI-assisted work is counted as other work.
Minimum sample
A figure for the whole org needs at least 20 pull requests. The one exception is the count of merges with no approving review when every merged pull request was read. A comparison between AI-assisted work and other work needs at least 20 in each group. Below that, the figure is not shown. A median time to merge is shown only for a group with at least 20 merged pull requests.The review figures
Two figures come from pull-request reviews read on a sample. See What the scan reads.- Sent back for changes. A floor, written “at least … sent back for changes on GitHub before merging”. Changes asked for outside GitHub are not seen.
- Merged with no approving review. Written “merged with no approving review from a person on GitHub”. An approval from a bot or app does not count. When every merged pull request was read, it is a count. Otherwise it is a share with a 95% range, “measured on a sample of N”.
A monthly figure you typed
With no Anthropic usage connected, a cost row can come from the monthly figure you typed. It is shown under the labelReported monthly and the headline says the team reported it. It is your figure, not a measurement, and the report marks it as an estimate.
Targets
A target is a direction for a row, not a promise. Every target is labelled as an estimate and has no dollar amount. There are two kinds:- The published target. Only on cost rows where spend is rising: long sessions cost 20% to 35% less than stock Claude Code in Mnemom datasets. This is an estimate. Its source is Agent savings.
- Your own baseline. Some rows compare you with yourself, and the line ends with
Basis:and what it was drawn from. For example, the AI-assisted sent-back share is compared with your own rate for work with no AI mark.
spend_rising_fast and spend_rising (the published target), and ai_up_output_flat, repos_not_ready, ai_sent_back_more, other_work_waits_longer and reverts_present (your own baseline). See Which rule picks each row.
Which rule picks each row
Rule table2026-11.2 has 14 rules. In each theme, the first rule that matches picks the row.
The recommended fixes
Each theme has one fix, with one exception. A follow-up row picked bymerged_without_approval gets its own fix, because that row is about work nothing on GitHub checked. Every fix links to Mnemom Agent.
What the scan reads and does not cover
From GitHub, through the Scan app, the scan reads pull-request metadata: size, check results, the merge day, revert markers, marks of AI tools, and review state on a sample. It never reads code, diffs or file names, and it does not keep who wrote or reviewed anything. The details are in What the scan reads. From Anthropic, when you connect it, the scan reads your usage and cost reports. The report does not cover:- Subscription seats, such as Claude Team or Enterprise plans. Measured spend is Anthropic API usage only.
- Other AI providers. Their spend is not read.
- AI-assisted work with no mark of an AI tool. It is counted as other work.
- The quality of your code. The scan does not read it.
What Mnemom keeps
Mnemom keeps no copy of the report. The report lives on the page while it is open, and in any export you save. At every share level, Mnemom keeps these records of a scan:- A run record: who started it, their role and how they signed in, the share level, the status and any failure code, the times it was created, started and finished, the GitHub numbers of the repositories (not their names), and the Anthropic Console ids you connected, if any. These are ids, not secrets. An email address you gave for the summary is cleared when the scan ends.
- An audit entry when the scan starts: who started it, the org id, the scan id, the share level and the number of repositories.
- An audit entry each time Mnemom issues a workload identity token for the scan to read your Anthropic usage: who started the scan, the org id and the scan id.
none, Mnemom also keeps:
- A run log entry: the scan id, who ran it, the time, the share level and whether the email was delivered.
- A send log: one entry for each email, with the recipient address and the subject line.
The email summary
When the scan finishes, the email summary lists up to five row headlines, the window and the as-of time. It links to the Org Scan page and to the method. It never names a repository. The full report, with repository names, is on the page and in your export.Export the report
Under Export this report, there are two buttons:- Save data file saves
mnemom-org-scan-report-YYYY-MM-DD.json. It holds the signed report. - Save page saves a single
.htmlfile that opens in any browser, offline. The signed data file is inside it.
Open it again
No Mnemom link opens a past report, because Mnemom has no copy. To see it again:- open the saved page in a browser, or
- load the saved data file on the Verify a saved report page in your organization’s Org Scan settings. It checks the signature in your browser and sends nothing to Mnemom.
Verify the signature
The saved page does not check its own signature. To check it, load the data file on the Verify a saved report page, or verify it yourself. The data file is JSON with ajws field. That is a compact JWS signed by Mnemom with Ed25519. To verify it yourself:
- Read the JWS header. Check that
algisEdDSAandtypisOrgScanReport/v1. - Fetch the signing keys from the Mnemom API host your org uses, at
/v1/.well-known/jwks.json. For example,https://api.mnemom.ai/v1/.well-known/jwks.json. Find the key whosekidmatches the header. Akidthat is not there means the report was signed by a different Mnemom host. - Verify the signature with that key.
- In the payload, check that
issishttps://mnemom.aiandsubis your org id. The token also carriesiat(when it was signed) andexp(by default 365 days later). The Verify a saved report page does not check either time, so a saved report still verifies there afterexp. Checkexpyourself if your policy needs it. - The report is the
reportclaim. Itsschema_versionisorg-scan-report/2.
exported_at time in the data file is outside the signature. These are the same keys and steps as for an AAP attestation.
Related
- Run Org Scan yourself: the command-line scan, its email, and what it reads.
- Mnemom Agent: the capabilities the fixes point to.
- Agent savings: the source of the published target.