smart_money · security master, live in production2026-09-07

CUSIP Succession, Live in Production

Six days ago this was five events, hand-verified, proven to work but not yet wired into the pipeline everything else reads from. It's now 260 events, individually cited, wired into the live join — and every number below is the current, automatic output of that pipeline, not a projection.

260
events curated & live
255
verified in one overnight pass
1.1%
residual false-signal rate on curated cusips
2
edge-case categories held out on purpose

Exhibit A — identity continuity, real filer

Brookfield Business Partners LP → Brookfield Business Corp · JPMorgan Chase & Co, live production data

Before the fix (how this read on 2026-09-01)
G16234109 (old CUSIP)
  classified asEXIT
  shares → 01,013,814 → 0
113006100 (new CUSIP)
  classified asNEW
  shares70,300
Read as: JPMorgan sold their entire Brookfield stake, then separately bought into an unrelated new security.
Today's live pipeline output
2025-12-31 shares1,013,814
2025-12-31 eventincrease
2026-03-31 shares70,300
2026-03-31 eventdecrease
Reads as: JPMorgan trimmed a continuing position by ~93% — a real, large trade, but one position, not two fabricated ones.

Nothing was hidden or smoothed over — JPMorgan really did cut this position hard. The fix just tells the true story: one continuing position that got smaller, not an impossible full-exit-then-brand-new-purchase in the same quarter.

Exhibit B — identity continuity and the split ratio

Amcor PLC, 1-for-5 reverse split · BlackRock, Inc., live production data

Before the fix (how this read on 2026-09-01)
G0250X107 (old CUSIP)
  classified asEXIT
  shares → 0158,434,896 → 0
G0250X149 (new CUSIP)
  classified asNEW
  value$1,321,347,037
Read as: BlackRock exited entirely, then made a fresh $1.32B purchase.
Today's live pipeline output
2025-12-31 shares31,686,979
2026-03-31 shares31,813,214
Change+126,235 sh (0.4%)
2026-03-31 eventincrease
Reads as: BlackRock's position was essentially flat through the split — the correct, unremarkable answer.

Without ratio adjustment, an owner who did nothing shows up as a $1.3B new buyer. The re-expression at the 5:1 ratio applies retroactively across this security's whole history in our data — the same way a split-adjusted close applies to price — so every quarter, not just the ones touching the split, compares correctly. Independently corroborated, not just by us: the SEC's own official 13(f) list marks Amcor's old CUSIP DELETED the following quarter.

Exhibit C — a chain, not just a pair

Menlo Therapeutics → VYNE Therapeutics, three separate events over three years

586858102Menlo, pre-2020
→ merger, ×1 →
92941V1002020 merger
→ split, ×4 →
92941V2092021 split
→ split, ×18 →
92941V3082023 split — current

Three unrelated-looking CUSIPs, three separate corporate actions, each independently curated. The chain-resolver composes the ratio across every hop: 1 × 4 × 18 = a single, correct 72:1 adjustment from the earliest CUSIP straight to today's. A holder who's been in this security since 2019 reads as one continuous position, not four disconnected fragments.

What curation correctly refused to force

the review discipline is the point — a wrong merge would be worse than no merge

SPAC units — excluded as a category, not case by case. A special-purpose acquisition company's units split into common stock plus a warrant. Whether that counts as “the same position changing identity” or “a genuine change in what's held” is a real, debatable question — not one to answer by pattern-matching. Held out entirely pending a deliberate decision, rather than guess.
Preferred-to-common conversions — same reasoning. A share class conversion is economically different from a rename or a split. Excluded from the 260 for the same reason as SPAC units: a plausible-looking default here is exactly the kind of thing that erodes trust in the whole dataset if it's wrong.
A high-visibility name got extra scrutiny, not a pass. A BlackRock Inc. holding-company reorganization looked, on its face, like it deserved a second look before curating — two CUSIPs, both showing years of history. Verified independently against our own filer-level exit/new counts (1,493 exits, 1,783 new positions, both in the exact reorg quarter) before it was added. Confirmed correct; the extra step is the point.

Where this leaves things

All 260 events are live in production. The comparisons above are the actual current output of holdings_events — the same table Qdrant, the MCP server, and the mars platform integration all read from, with zero changes needed on their end. What's left is scale, not mechanism: an explicit policy call on the two excluded categories, and curating the remaining ~9,500 lower-confidence candidates the detector has already flagged, at the same standard as the 260 already shipped.