Read The Nightly Scorecards

How to read the morning note and the nightly scorecard

Note

Goal: Read what last night did in the morning note, then read the corpus scorecard's numbers in the order that makes them mean something, and know which ones are safe to ignore. Prereqs: a running daemon (agentmd status answers), and at least one night since the morning-note and corpus-scorecard jobs were registered. Both reports are written into the vault.

Where they are

Both land under diagnostics/ in the vault, one file per day plus a copy of the newest at a stable name. The morning note is diagnostics/morning/YYYY-MM-DD.md, and the corpus scorecard is diagnostics/health/YYYY-MM-DD-health-scorecard.md. Diagnostics has been its own vault space since filing-v2 part 2a; the old desk/diagnostics/ location is gone.

ls ~/Vault/Agent/diagnostics/morning/latest_morning_note.md ~/Vault/Agent/diagnostics/health/latest_health_scorecard.md

latest_morning_note.md is about last night: what ran, what needs you, and what it cost. latest_health_scorecard.md is about the corpus: what is in it and whether it is degrading. The dreaming scorecard is gone, and the morning note carries what it used to report.

Steps

The morning note

The morning-note job writes it as the last step of the night, order 5 in the 02:00-06:00 window, after enrichment, the dreaming binary, the Python cycle and the corpus scorecard. Each of its four sections appears only when it has something to say.

1. Read the last line of What ran first. Did not run last night: names each nightly step that did not run, with the reason from the runner's last cycle: disabled, dry run, not registered, budget-ceiling, and the rest. When the runner starts the note, no cycle in the window has finished yet, so a step that did not run reads no reason on record for the night unless it is disabled or held at watchdog-stop. The reasons the night's cycles gave are in ~/.cache/agentm/telemetry/runner-launchd.log. A step you ran by hand inside the window counts as having run. When the line is missing, every step ran.

2. Read the enrichment line left to right. The format is N judged · N filed active · N below the floor · N sank · N calls · N tokens against the line · <model>, with N project records · inserted just before the call count on a night that merged into any (agentm-vault plan 09) — a project's charter, or a note under its decisions/, designs/ or research/, is merged into rather than filed, so it sits beside the filing verdicts rather than inside them. Judged is how many notes the batch sent to the model, and the next three counts are what their verdicts decided. A note that sank is counted below the floor as well. N failed and Stopped by <reason> follow when a run had failures or reached one of its limits.

3. Check the dreaming binary's gate before its table. The binary's line names the pass's mode, its outcome and the gate's reason, and a table gives one row per job: lifecycle, copies, refile, promote, calendar, mocs, dates. When the runner started the binary and its gate held, the line reads ran, and its gate held; the last pass was N ago and there is no table. Nothing was due that night: too little time had passed since the last applying pass, or nothing had happened since.

4. Fix filing is halted before anything else. The Python cycle's line counts possible twins, shared keys, proposed facets, and three lint findings. When it reads filing is halted instead, the filing contract did not parse and every stage after that check stopped. The line carries the parse error.

5. Work What needs you from the top. Each list gives a count and its first five items: unfiled notes the batch judged below the floor, possible twins with their similarity, shared keys, proposed facets, and what sank in the last seven days. Then three lists about what is coming or has just landed: sinking within 30 days and archiving within 30 days, each note with how long it has been silent — a threshold is easier to argue with while it is still thirty days off — and lessons this week, every note the weekly crystallize phase wrote, with how many sources taught each one. Read those last. A lesson never decays and nothing ages it out, so the morning it is written is the cheapest morning to disagree with it; if you would not keep one, that is the phase's own re-audit trigger firing. The next line links the day's dreaming facet, which lists every move the night made with the manifest behind any deletion, and the last links the needs-review map, which holds the full lists. See Review flagged memories for working it.

6. Read The corpus as one sentence, then follow its link. It gives the class populations, N awaiting a judgment, the oldest <age>, and coverage N of M stamped at this pass, then links the day's corpus scorecard. When the daemon does not answer, the line says not measured and why. It never prints a zero for a number it could not read. When coverage reads not measured (the ledger answered 0 eligible twice over a corpus of N cards — read it again), run agentmd ledger --pending --limit 0 yourself a minute later. An empty corpus still reads coverage 0 of 0.

7. Check Spend against the operator's lines. Last night: gives the tokens the night added against the line, each tier as <tier> N of <line> (2,000,000 a night on either tier), then the tokens it processed, calls against the 250-call guard, and dollars. The line counts what a call adds: input, cache writes and output. The processed figure also counts the cached prefix every call re-reads, so at a card's measured cost it runs about three times larger. The note holds the night to these numbers even when a run by hand lowered its own. Under Last night: sits a line per tier-table job — enrichment deep, enrichment light, crystallize — with that job's own tokens, calls and dollars. That is a finer reading of the same spend, not a second accounting of it: a run recorded before the meter carried job names has no per-job lines and still counts in the totals. The weekly phase: appears on a night the crystallize phase ran, with the lessons it wrote, the recurrences that cleared the bar, its calls and its cost — the design re-audits that phase's cadence on this line after three runs. Seven days: gives both token figures and dollars across the week's paid runs, the weekly phase included. Sessions, the last day: appears when the observability rollup recorded session spend.

8. If the session-start line warns that notes stopped, look at the runner. The line normally reads [agentm] Morning — <headline> (written <age>). The headline is What ran in one line, with a count of the lists that need you. Once the newest note is two days old, the line reads ⚠ Morning note — none in N days (last: <date>); the night has stopped finishing — see runner. Set AGENTM_DIGEST_DEADMAN_DAYS to change the two-day threshold.

The corpus scorecard

1. Read the corpus scorecard top to bottom, not by hunting for red. The sections are ordered so each one gives the next its context: how much is in the corpus, whether enrichment kept what the sources said, whether the writing is converging, and what the graph looks like. The first line is class populations — flat memory counts per class directory (semantic, procedural, episodic, entities, crystallized, mocs), the accumulate loop's supplement lanes reported apart so a class holding only lanes doesn't read as populated. It exists because the six class directories once sat built and empty for months while the corpus lived in a staging area, with nothing counting them — filing-v2 part 3 populated them, and this is the line that keeps that failure from going invisible again. entities and mocs commonly read zero: they're derived classes nothing routes to directly yet, not a sign anything is missing. episodic holds one session-trace note per session, written at session end — it populates as soon as a session runs on this vault. A diversity number means something different over 500 notes than over 5.

2. Read "needs review" and "writes per day" right after class populations. Both replace what the old staging directory used to make visible just by its own size.

   Needs review counts notes still waiting for a judgment — filed at low confidence, still unfiled, or flagged as a probable duplicate or a same-key update — broken down by reason, with a link to the generated page. See Review flagged memories for how to work it.

   Writes per day is today's count against the volume gate's cap: the 7-day mean, this week against last week, the fortnight's peak, and the headroom left before the gate refuses the next write.

3. Check what is unavailable before believing what is green. Every reading is either measured directly or marked unavailable-with-a-reason. There is no third state where a number is invented to fill a row. A section reporting the sampled grading pass is not built yet is telling you the truth; a section reporting 0.00 would not be.

4. Read the completeness number with its sample size. claim-level coverage is the headline, and the cell beside it says how many notes and how many replicates it came from. Five notes at three replicates is a smoke test, not a corpus measurement, and the row states that rather than leaving you to assume.

5. For the diversity meters, read the direction, not the value. There is no good absolute number for trigram concentration. What matters is which way it moved since the last card:

   Meter	Worry when
   trigram concentration	rising — the same phrases recurring
   lexical diversity	falling — vocabulary narrowing
   pairwise similarity	rising — every note drifting toward every other
   nearest-neighbour dispersion	falling — clusters tightening first

   The last one moves earliest. Convergence starts locally, so a few notes tighten around each other while the corpus-wide average is still flat.

6. If the two embedding meters say they refused, that is correct behaviour. They will not run without the dense arm rather than returning zero. The safeguard exists because zero dispersion is what a perfectly converged corpus looks like and zero similarity is what a perfectly diverse one looks like, so a missing embedder returning zero would report either "everything is fine" or "the corpus has collapsed" depending which row you read.

7. Regenerate either report on demand rather than waiting for the next night:

   python3 harness/skills/memory/scripts/morning_note.py
   python3 harness/skills/memory/scripts/corpus_scorecard.py

   A morning note run by hand covers everything since the most recent 02:00, and it replaces that day's note and the latest_morning_note.md copy.

What these reports will not tell you

Whether a specific note is good. Every number here is over a sample or the whole corpus. To ask about one note, read it.

Whether retrieval is working. That is the pinned evaluation (scripts/check-retrieval-regression.sh), which scores a frozen gold set and is the only thing that speaks to ranking. A healthy-looking corpus scorecard says nothing about whether the right note comes back.

Whether the numbers moved because the corpus did. A meter can move because the sample window shifted or because a migration rewrote frontmatter. Before reading a change as a trend, check whether anything ran between the two cards.

Related

• Memory daemon reference § the morning note — every line the note can carry, where each is read from, and who reads the note.
• Memory daemon reference — the subcommands each section reads from, and what the meters refuse to do.
• Enable the daily email — get the morning note by mail.
• Review flagged memories — working the needs-review page this scorecard's line counts.
• CI gates reference — the deterministic checks, including the pinned retrieval evaluation these cards deliberately say nothing about.
• Audit the vault — the per-note lint, which asks the opposite question to a corpus-wide meter.