Dream
Reach for Dream when the pile has stopped being trustworthy rather than when it has stopped growing.
- A quarter has passed since the last pass. 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 in front of a customer.
- You just bulk-imported. A transcript dump or a migrated wiki arrives with duplicates baked in.
- A health score has drifted down two passes in a row. That is a trend, not a bad week.
Do not reach for it to make a base bigger. Dream never adds knowledge. It only removes the parts that have stopped being true, and it asks first every time removal touches human words.
Consolidation is the habit most operators are missing, and it is not about adding anything. It is pruning. Which of these went stale, which two say the same thing, which one was written with total confidence and turned out to be wrong.
Dream runs that pass on a schedule across every note the system holds, flags what is stale, contradicted, or duplicated, fixes what it can fix safely, and asks before it touches anything else.
When I first ran it on my own accumulated intelligence, roughly three quarters of it was no longer accurate. Deleting that felt like destroying work. It was the most valuable pass I have ever run, because stale intelligence is worse than none. People trust it, and then they act on it.
The Four Passes
- Stale. Notes past their expiry date, or untouched long enough that the world they describe has moved.
- Contradicted. Two notes making claims that cannot both be true. Dream flags the pair and dates them. It does not pick a winner.
- Duplicated. Two records about the same person, account, or decision. Merge target proposed, never auto-merged.
- Dead references. Links pointing at files that no longer exist.
The Rule That Governs Every Fix
An early version had a rule to remove orphaned index entries, meaning entries pointing at a file that no longer existed. Reasonable. Then it met an index entry running 2,371 characters and linking ten files, one of which was dead. The naive rule would have deleted the whole entry, nine live references and a paragraph of history, to fix one broken pointer.
The rule that came out of it is procedural, not motivational:
Every fix that runs without approval has to be additive or markup-only. If a fix would remove words a human wrote, it goes in the ask-first column. No exceptions.
The fix for the dead link became de-link, never de-line. Strip the link markup around the dead reference and keep every word of the sentence it sits in. Applied correctly, that repair cost exactly 252 characters of markup and zero words.
Two Columns, Always
| Runs Without Asking | Asks First |
|---|---|
| Strip markup around a dead link | Delete any sentence or paragraph |
| Add an expiry date | Merge two notes into one |
| Add a tag or a cross-link | Rewrite a compiled-truth section |
| Append a dated timeline entry | Resolve a contradiction by picking a side |
| Flag a duplicate pair | Retire a note from the active set |
Do Not Trust A Clever Rule With The Record
A rule to auto-correct vague dates in notes looked useful until I measured it: 107 raw matches and about 5 real ones. A false-positive rate near 95%, most hits being quoted speech or phrases like "currently missing" that were correct as written. Run automatically, it would have silently falsified the record it was meant to clean. It is still in the skill. It still asks.
The lesson generalizes. Measure a rule against the real corpus before you let it run unattended, and the measurement is the count of true positives over raw matches, not how sensible the rule sounds.
The Health Score
The pass ends with a number so the next pass can be compared to the last one. It starts at 100 and takes deductions per stale note, per expired note, per dead reference, per duplicate, with a small bonus for notes carrying an expiry date.
| Band | Score | What It Means |
|---|---|---|
| Excellent | Above 90 | Pass on schedule, nothing urgent |
| Good | 70 to 90 | Clear the ask-first queue this week |
| Fair | 50 to 70 | Contradictions are probably in circulation |
| Poor | Below 50 | Stop citing the base until it is cleared |
The absolute number matters less than the direction across three passes. A base sliding from 88 to 74 to 61 is telling you the capture rate has outrun the pruning cadence.
Where It Sits In The Loop
Wrap-Up and Ingest create. The accumulation is the point. Dream is the counterweight, running over everything the first two piled up. Graph Health is the narrower diagnostic on structure rather than truth, and Weekly Review is the cadence that runs all of them and writes a dated file so health becomes a trend line instead of four checks nobody compares.