Stocks desk — 14 Oct 2026, 14:00 UTC

Phase 1 set the frame: cash $38,563.12, portfolio value $97,789.46, 0 of 3 buys used, nothing pending, 5%/position cap (≈$4,889). My book is concentrated in AI-Tech four-deep (ANET/NVDA/TSM/NOW) plus Energy (XOM/CVX), Healthcare (ABBV/ISRG/MRK), DIS, TSLA. The two carried-forward lessons both say the same two things: (10-13) the insulation test must gate on the THEMATIC/earnings driver, not the GICS label — AI-power/memory/neocloud names load on the same AI-capex factor I already own; and (10-12) the 50d vs stop buffer is computed off the research snapshot, so if the live limit prints materially above it the real geometry is thinner than the table — re-derive stop = live×0.92 vs the slow MA before trusting it.

The tape backs the insulation lesson hard today: memory/chip rout (MU 938.865 below its fast MA, WDC sell, SNDK/STX falling), Burry's short broadening to a chip ETF, D.A. Davidson's "CRWV could go to zero," hot CPI (+4.0% YoY, core 2.6% vs 2.5% est) lifting the 10Y to 5.24%. This is an AI-risk-off session, so adding same-factor AI exposure buys concentrated downside.

Phase 3(a): no discretionary exits. SYM was the only mechanical stop (hard 39.82 hit, −10.0%) — not my call. Of the rest, DIS alone shows a sell/thin-bearish crossover (fast 105.1719 < slow 105.8018, −0.6%) but price 108.2488 is above both MAs and it's +1.7% — a marginal crossover, not a broken thesis, so it stays on its ratcheting stop (hard 97.86) as in prior runs. MRK (−0.5%, the 10-12 thin-buffer name) holds its Healthcare-decorrelator thesis; stop 133.96. No trims.

Phase 3(b): screening the whole buy list through the lessons. GOOGL (+5.7% clean) and SPY (+8.0% clean) are geometrically fine but both ARE the AI/Mag-7 concentration I'm trying to decorrelate from (SPY is "the Last Men Standing" cap-weight), so the 10-13 rule rejects them. VST (+1.7%) is an AI-power proxy — explicitly the 10-13 failure case. META (−5.9%), MRNA (−24.1%), WBD (−0.5%), TLN (−4.2%), PLTR (−4.9%) all fail geometry (negative 50d vs stop). Held Energy/Healthcare buy names (XOM/CVX/ISRG/ABBV/NOW) are all ADD-RISK per check_stocks_stops. That left exactly one genuine decorrelator: V — Financials (I hold zero), payments demand independent of AI-capex, bullish (394.545 above both MAs), +2.3% geometry, no earnings blackout, analysts 12/33/6/0/0. The MRK 10-12 "act on the true decorrelator" case, not GEV's false one.

Held all 11 surviving positions. One new buy (1 of 3): V, intended as a half-size entry because its +2.3% snapshot buffer was already thin and V was "on pace for a record close" (extended intraday — exactly the 10-12 spike-entry risk). But the script dollar-sizes to the 5% cap regardless of the QTY I passed, so it filled a full 12.196262 @ limit 400.9 (≈$4,889). The live limit (400.9) printed +1.6% above the 394.545 snapshot, so the 10-12 phenomenon materialized: re-derived stop = 400.9×0.92 = 368.83 against a 50d MA of 371.30, collapsing the real buffer from the table's +2.3% to only ~+0.7%. The order cleared the coded 0.2% band (guardrail works off the live quote), so this is a thin-geometry full position I'd rather have sized down — a sizing-discipline miss, not a rule breach. Kept 2 buys in reserve; passed on every AI-complex name as same-factor concentration into an AI-risk-off tape.

The 10-12 thin-buffer rule got its first real test today and exposed a mechanical gap: I correctly identified V as the one true non-AI decorrelator (Financials, +2.3% 50d vs stop) but judged its snapshot buffer too thin and V too extended ("record close" pace) to deserve a full position, so I tried to size down by passing a share QTY to place_stocks_order (buy V 6). The script IGNORED it and dollar-sized to the 5% cap anyway — it filled 12.196262 @ 400.9 (≈$4,889, full cap). And the fill confirmed the 10-12 risk exactly: live limit 400.9 printed +1.6% above the 394.545 snapshot, so the re-derived stop 400.9×0.92 = 368.83 sits only ~+0.7% below the 50d MA 371.30 — the table's +2.3% buffer collapsed to razor-thin at the real entry. Generalized principle: "size smaller to respect thin/extended geometry" is NOT a lever this script gives you — a share-QTY argument does not reduce the dollar size; every buy fills at the 5% cap. So when the 10-12 re-derivation (stop = live×0.92 vs slow MA) shows the live buffer has collapsed toward the slow MA on an extended intraday entry, the only real risk controls are WAIT for a pullback entry or DON'T buy — a cap-sized fill is the ONLY outcome, and conviction must clear that full-size bar before you submit, not after. Falsifiable: if V holds above ~371 the thin buffer never gets tested and the Financials decorrelator plays out as the right "act" case (like MRK 10-12); but if V mean-reverts its intraday pop toward the 50d, the stop at 368.83 is far closer than +2.3% implied and I should have waited for a pullback rather than submitting at the record-close spike — watch whether entering extended, at full cap size, is what puts a geometry-thin decorrelator at risk.