AI · Berkshire Doctrine — Trades
Every virtual fill this book has made, with the model's reasoning — the fused signal, valuation and news case — behind each. Prices are model fills; quantities are never shown.
Fill history
1 recorded fill.
| Date | Ticker | Name | Side | Fill Px | Signal | Conv | Valuation | Why |
|---|---|---|---|---|---|---|---|---|
| 2026-10-31 | GOOGL | ALPHABET INC | BUY | 281.09 | ReversalUp_Oversold | — | GROWTH PRICED IN | rationaleQ ended 2026-06-30 (pub 2026-07-23): rev +24.2% YoY; margin elevated by one-time gains. No sustainable catalyst to add. Hold. The fund holds Alphabet at a 6.59% weight. After the -12.6% fall on the Q3 EPS miss the position has drifted to 5.45%, so holding at weight means buying back to target. Target Book 6.59% vs held 5.45% (gap +1.14pp). Under-weight by 1.14pp: market at next open. Technicals agree on direction (ReversalUp_Oversold conv 90, RSI 22). The 12wk projection 197.41 is a gap-extrapolation and the valuation is GROWTH_PRICED_IN; both are annotation only, and doctrine buys to target. The post-earnings gap is DOWN, so this is not an L6 chase. Per L8, reprice if the 10-30 open gaps >~1.5% from 283.82. GOOG (+0.38pp under) is sub-threshold: no card. · news decided @ 283.82 · stop 254.37 · t1w 297.86 · t4w 330.62 · t12w 197.41 |
Learnings in force
What this book's own record has taught it.
Last reviewed 2026-10-02
Rank buy candidates by hit_4w and stop_first_rate_1w; there is NO named favourite family. (REWRITTEN 2026-09-13, US-90/US-74.)
The former 'PullbackBuy_VWAP is the sole reliable buy cell (~0.77-0.86)' is retired as factually inverted: that cell decayed +8 (.90) -> +0 (.481) at n=28, hit_4w.227, hit_12w.059, stop_first_rate_1w.464. Prefer the non-extended entry on the family with the better hit_4w and lower stop_first_rate_1w. US-16's chase-family raw-conviction floors are UNCHANGED. US-2/US-10/US-24 cite the retired premise - historical only.
Trim on the stop-break + negative-12wk combo; don't wait for the SELL-REDUCE label.
A decisive close below the recorded stop AND a negative 12wk is itself a signal-based exit (CLAUDE.md §9), even while the engine still prints NO TRADE.
Execute the de-risk promptly.
A trim recommended but not executed keeps bleeding; act when a held name closes below stop with a negative 12wk, don't re-recommend the same trim for days.
The settled close is authoritative; don't act on weak opens.
Intraday/BOD prints whipsaw. Flag intraday reads provisional; reserve firm acquire/dispose for the settled-close EOD.
Use f=0.30 as the STANDING default for the 1wk estimate (calib_1w is structurally unstable on this book).
calib_1w oscillates across zero week to week (-0.045 -> +0.113 -> -0.025); f=0.30 is the default, not just the degenerate-case fallback, until n grows or the scorer winsorizes calib (US-15).
Don't chase post-earnings gaps.
A blowout (e.g. MU +14% AH) often round-trips within days; wait for the pullback to the model entry.
Same-pool/cross-book shares execute ONCE.
The parent book's decision governs; sub-book runs flag conflicts and avoid double- the shared cash. Tuned sub-book configs can diverge on the same share — surface the conflict, don't silently double-act.
EOD recommendations execute at the NEXT market open, not the decision close.
Persist actions as `execution_status: pending_next_open` with an `execute_on` date; action prices are close-refs — reprice/re-validate if the open gaps >~1.5%; the 1wk/4wk/12wk horizons and the scorecard grade from the execution date/price; live holdings/monitor stay unchanged until the JSON is updated post-fill; the next BOD reconciles pending proposals against the actual open.