SNAPSHOT — 21 NOV 2026 · LIVE FEED VERIFIED + TALLY
The PRISMA Ledger
Every item in the Mycelium, accounted for the way a systematic review accounts for every paper: identified → screened → flagged → assigned. You were right — it is over a thousand. 1,152 verdict chunks in the Master feed, 1,121 still open — now triaged by who can actually act on them, with the distillation gap shown in the same diagram.
1,152Total verdict chunks — exact match with the live Master feed, refreshed Nov 21
1,121Still open — the full backlog this page triages
248Look autonomous-doable on first pass — builds, fixes, verifications, filings
23,511 → 12Bronze → Silver. Gold 0 · Platinum 0 — the distillation gap
The verdict in one line: the architecture is sound — the problem is prioritisation and throughput. 248 items can move without waiting for a judgment call, 360 need a read before anyone can classify them, and almost nothing has distilled past silver. This ledger is the fix: verify each can-do item before touching it, work in queue order, report each as verified.
1 The flow — Dream Cycle subset
PRISMA 2020 structure. This is the 220-verdict Dream slice of the 1,121-item full queue in §2. Switch windows to see how the rate changed.
Identification
Claims captured & indexed
2,003
Entry counts across both windows (1,223 + 780). The audited ID-line basis below removes the counting change made on Oct 21.
↓
Screening — Dream Cycle, nightly, 5 checks
Claims screened
1,390
ID-line basis: 601 + 789. On this consistent basis the verdict rate rose 11.8% → 21.9% — driven by staleness, not sloppier capture.
↓
Eligibility — verdicts raised
Verdicts raised
244 → 220 unique
244 is the per-window sum. Oct 23–24 sits in both windows, contributing 24 verdicts twice — de-duplicated, the true queue is 220.
↓
Included — Master queue disposition
Awaiting disposition
210
Batch 1 (10 check-0 verdicts, Oct 21–Nov 5) was disposed Nov 21 on a voice-call “accept all ten,” closing 28 queue items in the ledger. App-side taps are still owed — the Master machine key can read and suggest, not close.
Clean — no verdict 1,146 claims (82.4% of screened) Passed all five checks. Only 2 fully clean days in the last 30, vs 9 in the first 30 — work velocity, not decay.
Duplicates removed 24 verdicts counted in both windows (Oct 23–24 overlap). Removed once, recorded here — PRISMA never lets a duplicate silently inflate the queue.
Disposed 10 verdicts · Batch 1 · Nov 21, ~21:00 CST 27 of its 28 underlying items were still open in the live feed at acceptance; 1 was already at “workshop.”
Not excluded — by design No claim is ever deleted from the index. A verdict attaches its correction to the claim itself (green-lit Nov 21), so retrieval surfaces the supersession alongside the stale claim.
How to read this like PRISMA: the left column must always sum. Screened = clean + verdicts. Verdicts (unique) = disposed + open. If a future tally breaks that arithmetic, the ledger — not the memory — is wrong, and it gets flagged as a verdict itself.
2 The full queue — 1,121 open, triaged
First pass over the entire Master feed, not just the Dream Cycle subset
The 220 Dream verdicts in §1 are one slice. The Master feed holds 1,152 chunks total, 1,121 open (31 already closed). First-pass triage of the 1,121 — every bucket sums to the total, nothing double-counted:
Can-do — autonomous
248 · 22%
Unclear — needs a read
360 · 32%
Dream verdicts
220 · 20%
Needs outward send
118 · 11%
Needs money
78 · 7%
Blocked on a person
64 · 6%
Needs your decision
33 · 3%
248 + 360 + 220 + 118 + 78 + 64 + 33 = 1,121. The biggest bucket is “unclear” — not blocked, just unread. That is where the next pass goes: classify first, then the can-do count will move.
How the can-do lane runs
Verify before touching — the first pass already misclassified a couple of sign-ups as doable, so no blind execution
Work in queue order, report each item as verified when done
App-side taps stay yours — the Master machine key can read and suggest, not close
This is days of work, not one night — triage + this ledger are tonight; the chew-through starts from the top of the 248
What stays blocked — honestly
293 items cannot be auto-closed: 118 need an outward send, 78 need money, 64 wait on a person, 33 need your decision
Those are not failures of the system — they are the boundary of autonomy, made visible instead of buried in one undifferentiated queue
3 The distillation funnel
Why the queue keeps growing — almost nothing moves up a level
Bronze — raw capture
Bronze
23,511
Everything captured and indexed. The intake works — this number proves it.
↓
Silver — normalised
Silver
12
Only 12 items have been normalised into silver docs. Conversion: 0.05%.
↓
Gold — synthesised
Gold
0
No gold synthesis yet. Demand-driven gold — distil only what gets revisited — is the proposed fix from the 300-workflow comparison.
↓
Platinum — verdict'd fact
Platinum
0
No platinum facts yet. The 10/day disposal cadence and confidence-graded promotion are how this level starts filling.
That gap — 23,511 → 12 → 0 → 0 — is the whole story in one diagram. Capture is not the bottleneck. Distillation is.
4 Why claims get flagged
Share of verdicts, first window vs last window
Staleness took over
Stale — first 30d
38%
Stale — last 30d
66%
Structural — first
31%
Structural — last
19%
Wrong numbers stayed flat: 6 vs 6 across both windows. The system isn’t getting less accurate — the world is moving faster than the claims. Of 44 classifiable receipt verdicts, 40 were superseded the same day they were made (median gap: 0 days).
The retrieval risk
The Nov 21 probe is why wiring verdicts into the index mattered:
3 of 8 test queries surfaced a stale claim with no marker that it had been corrected.
In only 4 of 8 did the correction rank #1.
Flagship find: the Sep 25 claim that the agent is still called “Muse” sat ACTIVE against the Oct 3 rename to Vesper — a name contradiction only the Dream Cycle caught.
Attaching the correction to the claim itself — rather than leaving it in the Master app alone — is the compounding move.
5 Autonomous assignment
Your instruction: run them all, see which ones Muse can do. Done — here is the actual split.
Counts below are the first-pass triage of the live 1,121 open items (§2), not a model. Each can-do item still gets verified before it is touched — the first pass already misclassified a couple of sign-ups as doable, so classification alone never triggers execution.
Lane A — Muse does it
Can-do
248 items
Builds, fixes, verifications, filings — no outward send, no money, no one else's decision required.
Worked in queue order, each reported as verified
Plus the Dream subset's mechanical work: same-day supersessions with corrections already indexed, check-0 wiring of corrections onto source claims
Exact duplicates from the window overlap
Lane B — Classify, then assign
Unclear + Dream
580 items
360 unclear need a read to classify — the single biggest bucket. 220 are Dream verdicts on their own track.
Unclear (360): read, then re-sort into can-do or blocked — expect the 248 to grow from here
Dream verdicts (220): your 10/day disposal cadence, “keep dreaming” — Batch 1 already disposed, 210 of this subset still open
Staleness confirmations pre-written as one-tap batches of 10; confidence-graded platinum promotions
Lane C — Blocked / Maxx only
Needs you or the world
293 items
These stay put no matter how fast Lanes A and B run.
118 need an outward send
78 need money
64 blocked on a person
33 need a decision from you — plus check-3 judgment calls and contradictions about people, money, identity, or names
Scenario
How it works
Clears in
Dream subset at 5 / day, no lanes
Below the 5.8/day arrival rate — queue grows forever
Never
Dream subset at 10 / day (current cadence)
Net +4.2/day after new arrivals — for the 210 open Dream items
~53 days
Dream subset at 20 / day
Double sessions, all manual
~16 days
Full queue: Lane A first
248 can-do items verified and worked in queue order; 360 unclear re-classified as they are read — no honest single date until that re-sort lands
Days of work
The honest constraint is unchanged: the Master key can read and suggest, not close — app-side taps are still yours. Lane A work is verified and reported per item; nothing is bulk-closed on a first-pass label.
6 Batch ledger
Two tracks: Dream disposal at 10/day + Lane A in queue order
Triage — full Master feed, first pass 1,152 chunks verified against the live feed Nov 21; 1,121 open sorted into 248 / 360 / 220 / 118 / 78 / 64 / 33. This ledger is the record.
DONE ✓
Batch 1 (Dream) — 10 check-0 verdicts · Oct 21–Nov 5 Accepted by voice call Nov 21, ~21:00 CST. Ledgered in dreams/disposals-2026-11-21.json — closed 28 queue items.
DISPOSED ✓
Lane A — 248 can-do items, from the top Verify each before touching it, work in queue order, report each as verified. No blind execution on a first-pass label.
STARTING
Batch 2 (Dream) — next 10, check-0 first Presented at 10/day under the Nov 21 cadence alongside Lane A. No batch skips Lane C items into an auto-close.
QUEUED
Nightly — Dream Cycle continues Cron runs ~02:31 CST, one finished day per run, skips above 90% of plan. New verdicts enter at Identification and flow through this same ledger.
RUNNING
7 What this ledger steals — on purpose
From PRISMA itself
The flow diagram was the #1 steal from the 300-workflow comparison for a reason: it makes double-counting and silent disappearance impossible to hide. Every future tally has to reconcile with the diagram above.
Still owed from the comparison
Confidence grades on platinum, not binary ratification
Retrieval-quality measurement, not just throughput
An explicit export guarantee for the ledger itself
Gold synthesis that preserves disagreement instead of smoothing it