Dream
Reach for Dream every two weeks, or whenever the base has stopped being trustworthy rather than when it has stopped growing.
- The two-week pass is due, or the content count has slipped. Notes do not announce decay. A wrong number looks exactly like a right one.
- Two notes disagree and both look current. Contradiction is the expensive kind of stale, because somebody acts on the wrong one.
- You just bulk-imported. A transcript dump or a migrated wiki arrives with duplicates baked in.
Do not reach for it to make a base bigger. Dream does not add knowledge. It removes the parts that stopped being true, and it asks first whenever a removal touches words a person wrote.
When I first ran a pass like this on my own accumulated notes, roughly three quarters of them were no longer accurate. Deleting that felt like destroying work, and it turned out to be the most valuable pass I had run. Stale notes are worse than none, because we trust them, and then we act on them in front of a customer.
Dream is that pass, and it runs every two weeks. It reads every item and opens each suspect before it reports it, so nothing gets flagged on a hunch. It fixes only what can be fixed without losing a word, and it asks about everything else. Then it writes a content count and a dated log, so the next run has a number to beat.
Three Modes
Every run names its mode on the first line of the report, so you'll always know what it touched.
- Gentle reports only. It changes no file and writes no log, so you can look without touching anything.
- Standard is the default, and it's the one to run. It applies the safe fixes after a dry run, writes the log, and queues every removal for the approver.
- Deep is standard plus a restructuring pass. That pass proposes moves and rewrites in the report, and it doesn't change anything.
What The Scan Finds
Stale items are past their expiry date, or they have gone more than 90 days without an update while still marked as emerging. Expired items are past the expiry date they carry. Dead references point at a name no file answers to, and duplicates say the same thing about the same subject. Contradictions are pairs that can't both be true. A dated note beside an undated one on the same subject is the likely superseded fact, and the owner decides.
The Rule That Governs Every Fix
Only three fixes run without asking, and none of them removes a word.
- De-link a dead reference. Strip the link markup and keep every word of the sentence, so you won't lose a word. De-link, never de-line.
- Fill a missing field, but only when the first sentence of the note already states the value.
- Suggest an expiry date for a time-bound item. The suggestion waits in the approval list.
The de-link rule came from a real mess, when an index entry that ran 2,371 characters and linked ten files turned up one dead link, and a naive rule would have deleted the whole entry, nine live references and a paragraph of history, to fix one pointer. The repair that shipped cost 252 characters of markup and zero words. Every fix runs as a dry run first, and a word count before and after decides it. If the count moves, it's rejected.
The Approval List
Nothing on this list gets written until the approver answers, and we set a default on each item so one word settles it. Expired items get an extension offer first. Archiving follows only on a yes, and it sets the status and keeps the file, so this skill doesn't delete anything. Stale items default to keep, with the question "still true?" attached. Contradictions default to hold, and duplicates default to keep both. Standing decisions get no expiry and are never archived for age, because a decision still in force isn't stale.
The Content Count
The content count starts at 100 and takes deductions for confirmed findings only, with a cap on each class. Items that carry an expiry date earn a small bonus. The count is separate from the structure score that Graph Health reports, so one number doesn't hide the other, and the structure score tells you which skill to run when a band drops below Good.
| Class | Deduction | Cap |
|---|---|---|
| Stale | -3 each | -15 |
| Expired | -4 each | -12 |
| Expiring soon | -1 each | -3 |
| Dead reference | -5 each | -15 |
| Contradiction | -4 each | -12 |
| Duplicate | -2 each | -10 |
| Missing field | -1 each | -5 |
| Restated rule | -3 each | -9 |
| Bonus, item carries an expiry date | +1 each | +5 |
The bonus rewards dates, since each item with an expiry date adds one point, up to five, and a base full of dates can correct itself.
| Band | Score | What It Means |
|---|---|---|
| Excellent | 90 to 100 | Pass on schedule, nothing urgent |
| Good | 70 to 89 | Clear the approval list this week |
| Fair | 50 to 69 | Contradictions are probably in circulation |
| Poor | Below 50 | Stop citing the base until it is cleared |
The Two-Week Rhythm
Dream checks whether a run is due before it scans, because a run is due when no log exists, when the next run date in the last log has arrived, or when the structure score has fallen into Fair or Poor. Every log sets the next run date fourteen days out, and Weekly Review reads the latest log, so a skipped pass shows up as a gap in the trend.
Compare each run with the one before it, not with an abstract standard, because a single score tells you little. Three in a row show which way the base is moving, and that's the trend worth watching. Run it on the cadence, and the base stays something people can trust.