SCINTILLA · ALLOCATION — LAYER THREE IN THE OPEN

23 SEP 2026 · LANE M22

You said layer three was jumping the gun, and that the other tool had a process of filtering between the options. Layer three is now that process: a funnel you watch, one step at a time, where every name that leaves is named and given the reason it left.

1 · WHAT THE OTHER TOOL'S FILTERING ACTUALLY DID

I read both of them before touching anything: the older allocation module at allocation.scintillahub.ai (its source is _ALLOC_DEPLOY on this Mac) and the sector rotation tool (SECTOR_ROTATION.html, which publishes the table sector_rankings).

The allocation module's filtering ran like this, and its own notes record each stage:

STAGE	WHAT IT DID
Universe	Every membership row — cohort, industry, size, sector — plus your favourites. Its own measured note: 387 tickers in, 265 of them candidates, 182 of those able to be ranked.
Candidates	A name qualified by sitting in one of the pick cohorts (growth, AI hardware, AI software, crypto, megacap, blue chip, thematic, international) or by being a favourite.
Geiger cover	Names with no Geiger were dropped. The tool's own audit flags this as the weak point: “83 of the 265 candidates have no geiger… they are invisible rather than flagged.”
Style re-rank	Geiger plus half the average style spread, so a cold QQQ literally pushes growth names up the list.
Sleeve competition	The coldest three in each sleeve were tagged ADD, the hottest TRIM.
Your hand	Drop greyed a row and pulled it out of the statistics without deleting it; pick marked a keeper, capped by your total-names limit. The desk page says it plainly: “this is a ranked shortlist, you still choose”.
Comparables	Peers by industry, with outliers removed by the standard 1.5×IQR fences — and the number removed was printed, not hidden.

The sector rotation tool is a different animal: it computes RS-Ratio and RS-Momentum on weekly bars and puts each sector in Leading, Improving, Weakening or Lagging. It publishes a rank per sector, which is the thing worth borrowing.

WHAT I TOOK FROM IT, AND WHAT I LEFT Taken: counts at every stage, a reason attached to every removal, names kept visible after they are removed, a cap the page admits to out loud, and the stance that the page hands you a shortlist rather than a verdict. Left for later: the drop and pick buttons that remember your choices, and the comparables sheet. Those are interactions with saved state, and I would rather build them once, properly, than bolt them on tonight.

2 · THE PAGE NOW READS AS THE DECISION

Top to bottom: how much to own, where it goes, which names, why not the others, and then the picks. Layers one and two each carry a strip saying what it read and what it produced, so there is no step where a number appears without a source.

The top of the allocation page at 1680 wide — The top, at 1680. The conclusion in words, then the four bullets underneath it — the last one now says how many names the funnel removed and at which step.

Layer two with the rotation column — Layer two. The **ROTATION** column is new: it is the rank the sector rotation tool published, read as it stands and never recomputed here. The sentence above the table says what that leg changed — and when it changes only the order, it says exactly that rather than claiming it changed the choice.

3 · THE FUNNEL

Every name the Geiger covers inside the chosen sectors enters at the top. Then, one step at a time:

STEP	ITS LINE	RECOMMENDED	WHAT IT READS
The Geiger	at or above the middle of its own sector	the sector middle	the published Geiger composite, chart API `/geiger`
The balance sheet	debt no more than 2.0× equity	2.0×	`ratios_history`, last full year, with the book date shown
Earnings too close	nothing reporting inside 7 days	7 days	`earnings_events`, from today forward
Liquidity	at least $25m traded a day	$25m	the median of twenty sessions of dollars traded, from `/candles`
Trend	closing above its own 50-day average	on	the same daily bars
Distance from a level	no more than 12% above that average	12%	the same daily bars

Each step shows how many went in, how many came out, a bar that narrows as the funnel narrows, its line as a slider you can move with the recommended value printed beside it, and then every name that left, with the number that removed it.

The funnel at 1680 wide, all six steps — The funnel at 1680, read live at 3:40pm New York on 23 September — these counts move with the market. At that moment: 130 names in the four chosen sectors, 66 past the Geiger, 61 past the balance sheet, 61 past earnings, 60 past liquidity, 58 past trend, 39 past distance — then the best three in each sector are held. The drop lists scroll, and say how many are in them.

It ends with the picks, each with the one line that earned it, and before them the names that did not make it — both the ones a step removed and the ones that cleared everything but ranked below your cut.

The picks with one-line reasons — The picks. One line each: its reading against the sector middle, its multiple against the sector's, its debt, what it trades a day, how far it sits from its 50-day average, and when it next reports.

Why not the others. First the near misses — “cleared every step, then ranked #4 of 6, behind MPC” — then everyone a step removed, each tagged with the step that did it.

4 · ON A PHONE

The funnel at 390 wide — The funnel at 390. The counts move under the step name, the sliders run the full width, and the reasons wrap. Nothing runs off the right edge.

The picks at 390 wide — The picks at 390, with the weight under the ticker.

5 · WHERE EVERY NUMBER COMES FROM

NUMBER	SOURCE
The reading on any name	chart API `/geiger` — the published composite across your eight saved timeframes. Nothing is recomputed here.
Which sector a name is in	`company_profile` on scintilla-live
Multiples	`fundamentals`, with the age of the row shown
Debt against equity	`ratios_history`, last full year, book date printed next to the name
When a name reports	`earnings_events`, and whether the date is confirmed
Sector against the market	chart API `/candles`, the eleven sector funds against SPY
The rotation rank	`sector_rankings`, published by the sector rotation tool, with its date on screen
Dollars traded, the 50-day average	chart API `/candles`, 120 daily bars per name
The live price	chart API `/quotes`, asked for after the page has drawn

6 · WHAT COULD BE WRONG

The Geiger step is a sector line, not a zero line. In a weak sector a name can clear it while its own reading is still negative — today AT&T and Alphabet both do, in a sector whose middle is −0.34. The page now says so on the name's own line, and the slider lets you demand a positive reading instead.
The tape is read for sixty names, not all of them. Reading 120 daily bars for every name would be hundreds of requests. The names past that cap are listed as unread rather than dropped quietly, and the cap is a slider.
“Distance from a level” means distance from the 50-day average, not from a support or a base you would draw by eye. It is the honest version of the idea, not the full one.
Trend and distance read the same average, so they are not independent tests.
The debt figure is the last full financial year — it can be nine months old. The book date is printed beside every name it removes.
The rotation rank is a day old (published 22 September, read 23 September). The date is on screen, and the page says what that leg changed.
Per-timeframe Geiger is not available. The chart API publishes one composite per name across the eight rungs, not a value per rung per name. So the funnel uses the composite. I did not invent a per-timeframe number to fill the gap.
Three names per sector regardless of the sector's weight. Communications carries 6% today and still takes three names at 0.9% each. That is your perSector slider's doing, and it may want a rule.
Two greys on this page sit outside the house rule (the ink tones inherited from the page as it was built this morning). I left the palette alone rather than restyle a page you said you liked.

7 · WHAT I DID NOT DO

I did not push and did not deploy. The branch is local, and you merge it.
I did not change how much to own or how the sector weights are worked out, beyond adding the rotation leg at 25% — and the page states what that leg changed, so it can be set to 0 and checked.
I did not build the drop and pick buttons with saved state. That is the one piece of the other tool I would add next.
I did not touch the Indicator Lab, the prototypes index, or any other room.
I did not compute a Geiger, an RSI or any indicator of my own. Every reading on the page is one the chart API already publishes.
I could not test this against the live Hub origin: the chart API only accepts requests from scintillahub.ai, so for these screenshots I served the page locally and passed the API calls through a local proxy. The data in every screenshot is real and live — only the route to it was local.