MIS v2 — Trade-Email Ingestion + Behavioral/Crowd-Fade Layer

Captured: 2026-05-27 (Wed) Session 39 cont. Sam dictated both of these as vision while reviewing the Test Entry tickets. Status: DESIGN CAPTURED, not built (per "classify before building" — feedback_classifier_first_doctrine). This doc is the contract so the logic isn't lost.

PART 1 — Trade-Email Ingestion (the holdings/balances brain)

What Sam asked (2026-05-27, verbatim intent)

"I make trades, my emails come in — figure out where they're going so they can be properly mapped, and whether I should change which email/folder they go to. Pick them up. In the future, follow up with an API pull for the actual price per share / amount sold and how that changes my balance, totals, holdings, emails, and briefings. It should figure itself out based on certain keywords / certain locations — parsed through specific words in specific trade emails, only then it goes into that row. And that row is checked every single time before the briefing. A brain and a skill we need to learn, commit to, update, and learn from itself. And we should know about it already."

DISCOVERY (done this session — live Gmail scan, both accounts)

All broker email lands in the BUSINESS inbox (sam@hookstreetcapital.com), INBOX, unfoldered/unlabeled. Personal inbox (ztreitel@gmail.com) had ZERO broker mail (only a Bloomberg newsletter). So: no routing change strictly needed, but a Gmail filter+label would make parsing bulletproof (see "Routing rec").

Two email classes already arriving (no API required to start):

1. TRADE EXECUTIONS — the fills.
   - Fidelity: sender Fidelity.Alerts@fidelity.com · subject "Fidelity Alerts: Order Execution" (and "Order Partially Filled"). Body pattern:
   Account *****<last4> · Your order to <buy|sell> <qty> shares of <TICKER> was filled · Filled <qty> shares @ $<price> · Execution time <h:mm am/pm ET> · Order number <id>
   - Schwab: sender donotreply@mail.schwab.com · subject "Your trade was executed in your Schwab account ending in ".
   - Example caught (TODAY 5/27): sold ORCL — acct 0600: 0.847 sh @ $188.44 (1:09 pm); acct 5692: 1.00 + 0.267 sh @ $188.56 (1:11 pm). (Answers Sam's "I'm not sure if it traded today" — it did.)

2. POSITION SUMMARIES — a holdings snapshot already emailed daily ~5:30 PM ET, one email per account.
   - Fidelity: sender Fidelity.Alerts@fidelity.com · subject "Fidelity Alerts: Position Summary" · body: Account: XXXXX<last4> then a Symbol / Acct Type / Quantity / Price / Value table.
   - This is a full holdings feed sitting in email — HOLDINGS_CLEAN can be seeded from these WITHOUT the Schwab API. Schwab also sends "My Closing Summary" (daily) + "My Weekly Summary".

Account map (corrected from the scan — MORE accounts than the plan assumed)

Open item for Sam: what are 5378 and 4073? (And is there a 2nd Schwab?) The account label drives which bucket/owner a fill maps to.

NOISE to exclude (so the parser doesn't false-positive)

Fidelity "Fixed Income Offerings" (muni bond blasts), "Daily Alert Summary (AM)" (news), EFT/transfer alerts, Schwab workshop/marketing, SchwabAlerts@markit.schwab.com Price Alerts. None are trades. This is exactly the false-positive risk Sam flagged — the parser must key on the execution subjects + body fields, not just sender.

Routing recommendation (the "should they go to a different folder" answer)

Keep delivery to the business inbox, but add a Gmail filter: from:(fidelity.com OR schwab.com) (subject:"Order Execution" OR subject:"Partially Filled" OR subject:"trade was executed") → apply label Trades (+ optionally skip inbox). Then the parser targets label:Trades — deterministic, immune to inbox clutter. A second filter → label Positions for the Position Summary emails. (Mildred boundary: these labels live in the business account; fine.)

The pipeline (phased — NOT all now)

• Phase A (read-only seed, no API): parse label:Positions (or the Position Summary subjects) → write HOLDINGS_CLEAN (Ticker · Account · Qty · Price · Value · as-of). Gives real holdings immediately. Checked before every briefing (Sam's "that row is checked every single time").
• Phase B (trade log): parse label:Trades executions → append to a TRADE_LOG (Date · Account · Side · Ticker · Qty · Price · Order# · source-msgId). Dedup by Order#. Feeds realized P&L + Ticker_Memory ("you sold ORCL @ $188 on 5/27").
• Phase C (reconcile + balances): TRADE_LOG Δ against prior HOLDINGS → updates balances/totals; flags drift between what email says and what we computed (integrity, like the Data_Crosscheck guard).
• Phase D (API, later): Schwab API for authoritative fills/positions (the 7-day re-auth cost) — only once A–C prove the email feed isn't enough.
• Self-learning ("learn from itself"): unknown sender/subject that looks like a trade → quarantine row + ask Sam via the bot ("is this a trade? which account?"); the answer becomes a new parse rule. New tickers seen in fills auto-add to the universe watch. This is the skill Sam wants codified (a trade-ingest skill in hookstreet-skills, updated as patterns evolve).

"We should know about it already"

We do now — this doc + reference_broker_trade_emails memory are the institutional knowledge. Senders, subjects, body fields, account map, noise list are all captured.

PART 2 — Behavioral / Crowd-Fade / Mean-Reversion Layer

What Sam asked (2026-05-27, verbatim intent)

"It needs to pick up some behavioral psychology — after a couple of strong days in specific companies or the overall market, [people] follow the crowd. It has to fade and go somewhere; there's selling and profit-taking. It has to think about that and grow in that way, always — be willing [to play] the downside of its own theory, especially if that's the way to make money on it. Not [profiting] on someone's pain, just a way of moving — making money from the way the market is moving."

Translation into MIS logic (Phase 3 forward layer + FSE awareness — NOT a new engine)

• Over-extension / crowd-chase flag: after a strong run (e.g., N up-days in a row, or 30D% above a band, or sitting in the top X% of its 52-week range), raise an "extended / fade risk" awareness flag — the move "has to fade." (The new range bar + EXTREME-move read shipped this session is the visible front end of this — INTC at 92% of range after +90%/30d is exactly the case.)
• Profit-taking prompt on holdings: if a held name is extended + showing distribution (e.g., RS rolling over, volume on down days), surface a TRIM/take-profit nudge — consistent with project_mis_cashflow_thesis (extract, don't round-trip).
• Two-sided willingness (the new part): the engine should be willing to express the downside of its own thesis when that's where the edge is — i.e., a fade/mean-reversion setup (or simply not chasing, and standing aside / trimming), not only long ADDs. Frame = "make money from how the market moves," risk-first, never reckless. This is a bias toward not being trapped long at the top, and (later, optional) explicit fade expressions (inverse/short or hedge via the leveraged-ETF universe already in the plan).
• "Grow / learn from itself": track whether fade flags were right (did extended names actually pull back +1/+3/+5d?) and tune the bands — ties into issue #9 (recommendation scorecard / self-backtest).
• Discipline guardrails (already captured, reinforced here): never add to a loser, post-loss pause, sunk-cost/hot-hand nudges (feedback_capture_is_not_closure, the SmartHub behavioral synthesis). The crowd-fade layer is the market-side complement to those self-side guardrails.

Where it lives

Phase 3 (forward layer), surfaced in the briefing + the entry ticket's "Position read" (already started). FSE stays the single classifier — the fade/extension signals are inputs/flags, not a parallel scorer (per MIS_FSE_ARCHITECTURE.md — no surface classifies independently of FSE).

PART 3 — Holdings cadence / notify / settlement + folder routing + "Aladdin" probing (2026-05-27 cont.)

DONE this session (real, verified)

• Holdings cleaned + loaded NOW: parsed the 4 Fidelity Position Summary emails → 30 positions, $33,156.65, written to HOLDINGS_CLEAN (verified via fn=holdings). Artifact: outputs/2026-05-27_14-40_holdings_clean.html. Accounts: 0600 Margin $26,737 · 5692 Cash $4,668 · 4073 $1,245 · 5378 $506.
• Repeatable auto-importer SHIPPED (code): misV2ImportHoldings_() (GmailApp search → parse → write → MailApp notify) + setupHoldingsTrigger() (daily ~6:10 PM, after the 5:30 PM summaries). Gated: needs gmail.readonly (+ script.send_mail for the email ping) added to appsscript.json + a one-time re-authorize. NOT added yet on purpose — adding a scope forces a reauth that would break the live run-endpoint mid-session. Turn on when Sam's ready.

Sam's cadence/notify/settlement asks (2026-05-27, verbatim intent)

"Clean my holdings right now, but it has to be consistently done and I'm notified somewhere it was done — to me or Mildred or tracked — so I'm up to date and it's fully compounded. If there's a day or two for things to settle, they have to match. Based on certain emails, organize my folders. Maybe request different searches/alerts from IBKR/Schwab/Fidelity to skip the inbox and go somewhere else — instead of you doing APIs."

• Consistently done: the 6:10 PM trigger (above). Compounded: the artifact already shows the cross-account aggregate.
• Notified / tracked: three options — (1) RUN_LOG row every import (tracked, zero scope); (2) MailApp one-liner to Sam (needs send scope); (3) Telegram ping via the bot (cleanest — the bot already has notify/pulse; have it read RUN_LOG or have the importer hit the notify webhook). Decision pending Sam: who gets pinged (him? Mildred? both?).
• Settlement / T+1 match: positions are stamped with the email date; a fill today appears in the next evening's summary. Build a reconciliation check: TRADE_LOG fills since the last Position Summary should explain the delta between two summaries; if they don't match after T+1/T+2, flag it (integrity, same spirit as Data_Crosscheck). (Live example captured: today's ORCL sale isn't in last night's holdings — it should clear tonight.)
• Folder routing (the "skip the inbox" path): the better answer than APIs for now — set up broker-side alerts/delivery + a Gmail filter so trade/position emails auto-label (Trades / Positions) and optionally skip the inbox, then the importer reads the label. Brokers (Fidelity/Schwab) let you configure which alerts fire; pair that with the Gmail filter. IBKR can be a new alert source here too (email alerts), without the IBKR API/gateway problem. → action: Sam sets broker alert prefs; I set the Gmail filters (needs Gmail label scope — confirm).
• Schwab 898: not covered by Fidelity's email. Need Schwab's position email (or its "My Closing Summary" parsed) — add a Schwab parser branch + confirm the 2nd Schwab account.

"Real market snapshot" on the Test Entry (2026-05-27 ask) — STARTED

Sam: "On the snapshot test entry I need a real market snapshot — micro & macro in 2 seconds: S&P, Dow, VIX, overall, sector. My last info, my last trades, any warnings, watch-losses, the things I had in my old test entry (Momentum + snapshot info). And ask me different things — psychologically see what else I want — like an Aladdin of my own."
- SHIPPED today: a Market now (live) block on the entry ticket = S&P (SPY) · Nasdaq (QQQ) · VIX + regime, all LIVE-tagged. Plus the Scenarios block (Base / Contrarian-fade / Macro / Micro).
- TODO (bring back from the old test entry): Dow + Russell; my last trades (from TRADE_LOG once wired); watch-losses / warnings (positions breaching stop, near-earnings, do-not-buy, wash-window); last info / Momentum snapshot rows he used to have. Consider a separate one-screen MARKET tab (the "2-second micro+macro" view) distinct from the per-ticker entry.
- The "Aladdin that asks me" layer: the system should prompt Sam ("does this still make sense? still believe the thesis? would you change it? what else do you want to see here?") and learn from the answers — surfaced through the bot, tuned over time. This is the self-prompting/self-learning cadence (ties issue #9 + project_self_learning_loop_gap). Capture his answers → adjust what each ticket shows. Design item, not built.

PART 4 — Portfolio cockpit vision + design language (2026-05-27 cont.)

Design language (researched, now the standard for MIS surfaces)

From Aladdin + 2026 fintech UX research (synthesized this session): glanceable (key state in 2 seconds), coherent (no contradictory signals — one story), self-explaining (every number says what it means + how to act), contextual behavioral nudges (ethical, Kahneman/Tversky), high-contrast color + dark-mode aware, adaptive/asks-and-learns. Applied to the entry ticket v3: THE READ with inline colored badges, GUT CHECK (behavioral), HOW IT'S PROTECTED + plain how-to steps, ★-marked self-explaining number chips, reworded scenarios, smart (red-if-stale) data-freshness, commas on ≥$1,000.

DONE this session

• Entry ticket v3 (deploy @17): all of the above. NVDA contradiction fixed (REJECTED because quality, even though with-trend). Real Market tab (S&P/Dow/Nasdaq/VIX/10Y/Gold live levels via buildMarketTab_, fn=market).
• Holdings split by owner: Sam's portfolio (0600 Margin + 5692 Cash = $31,405) front-and-center; children's accounts (5378 + 4073 = $1,751) segregated to a quiet view-only line — 5378/4073 are his kids' long-term accounts, not his own. (outputs/2026-05-27_14-40_holdings_clean.html.)

Sam's portfolio-cockpit asks (2026-05-27, verbatim intent) — DESIGN, not built

• Per-lot / per-trade P&L: watch-list per trade, exact per-lot gain/loss in % and $, performance. (Schwab is better at this but doesn't surface it well.) → needs cost basis → comes from the TRADE_LOG (parse execution emails) matched to lots.
• "Always-on" P&L tracker: a persistent running view — "you're up X% across the portfolio on $Y current value; realized + unrealized here; target = Z; you're this much above/below it." Plus a monthly view. (Ties project_mis_cashflow_thesis 2.5%/mo target — show progress to it.)
• Sell-scenario simulator (what-if): "if I sell this and that — it keeps this, stops the bleeding, locks some profit, allows some downside, but frees a bigger punch on the upside." A before/after on cash freed, risk reduced, P&L locked, exposure changed.
• Children's accounts: keep separate / mostly hidden, still updated; long-term moves only. (Done in the holdings view; carry the owner split into HOLDINGS_CLEAN + any P&L view.)
• Access everywhere: all of this requestable + showable through Telegram — links to a fresh build / his own PWA; one day a native app if warranted (discuss security/protection then).

Build order implied (when we leave the 2-ticker brain)

1. TRADE_LOG (parse executions) → cost basis → per-lot + realized/unrealized P&L.
2. Always-on portfolio P&L + target-progress view (his accounts only; kids segregated).
3. Sell-simulator.
4. Telegram surfacing of holdings/P&L (Phase 2 bot wiring) → then a PWA.

   Discipline: none of 1–4 starts until the 2-ticker brain is locked + Sam says go (feedback_classifier_first_doctrine, feedback_capture_is_not_closure).

PART 5 — Live-read punch-list + the intent/horizon insight (2026-05-27, "Stop Building")

Sam live-read the v3 ticket and gave detailed notes, then said "Stop Building." Build paused at deploy @17 (the v3 ticket he's reading). Source matches deploy (no half-shipped drift). Everything below is captured, not built — run when Sam says go. Discipline: he's right that I was spiraling on re-renders; the thinking is now logged so the next pass is fast.

THE BIG reframe — buy INTENT / hold HORIZON changes the verdict (NVDA-for-his-daughter)

Sam: "What if I want to buy NVDA for my daughter — long-term, 5+ years? It's recently down, the space should go up, it's the biggest one, it'll eventually go up even if earnings disappoint. So is it still set up for me? Is it protecting me or mothering me? I'm NOT giving it full control to trade — but the thinking has to be done already so when I look, it's all there (XYZ), not too much thinking."
- The same ticker is a different decision by horizon. The FSE today is a swing classifier — REJECTED-for-swing ≠ REJECTED-for-a-5-year-core. The engine must take intent/horizon as an input (Swing / Position / Long-term core) and judge against the matching rubric. A REJECTED swing can be a fine long-term accumulate. → wire Position_Intent (why am I buying: trade vs core; target hold) into the verdict + the ticket ("for a 5-yr hold, this looks like X; for a swing, it's REJECTED because Y").
- Advisory, NOT autonomous (hard boundary). It does the thinking and lays it out; it does not place trades. "Is it mothering me?" → right level = decision-ready, low-cognitive-load ("XYZ, not too much thinking"), Sam decides. → memory feedback_mis_advisory_not_autonomous.

Punch-list from the live read (concrete ticket fixes)

1. Explain the target. Show % upside to target, the R:R, the $ reward, time expectation (how long it should take), and a time-stop ("if not ~halfway in N trading days, the system flags it to re-decide"). He didn't understand "target $143.81" alone.
2. Real market data, not proxies. Gold = actual spot $/oz (not GLD ETF) and real Bitcoin price (he stressed this). GOOGLEFINANCE may not do spot gold/BTC → likely Finnhub (OANDA:XAU_USD, BINANCE:BTCUSDT) or another feed. (S&P/Dow/Nasdaq/VIX/10Y read fine.)
3. Track record / learn-from-itself (he stressed twice). On every ticket show what MIS said before on this ticker (from FINAL_STATE_HISTORY) + whether it was right (forward-return scoring +1/3/5/30d) + did Sam act (tie to TRADE_LOG / Position_Intent "why I bought"). Pull up the same ticket tomorrow / a month from now and see the evolution. = the self-learning loop (issue #9, project_self_learning_loop_gap).
4. Trading-calendar awareness (system-wide rule). Never brief/report/pull on a closed day (weekend/holiday); when closed, label values "as of last close"; express horizons in business/trading days. Applies to ALL surfaces + the briefing cadence.
5. Two sides — clearer wording. "The other side" → say the action it implies (wait / fade / pass). Sam: "if you go with it / the other side — what does that mean? I'm not sure."
6. Macro deeper on VIX (level + what it means: <20 calm / 20-30 nervous / 30+ fear + recent trend). Micro: expected swings over the next ~7/10/14 business days ("what to be ready for, what kind of swings").
7. Crowded / herd / hedge-fund read. Is this move crowded — "the herd doing what someone else is doing — could I be getting played by a hedge-fund move?" Flag crowded momentum + unwind risk.
8. Deeper behavioral gut-check (go further than the one-liner).
9. Polish: checklist could look even better; 10× the look/usefulness/coloring; more compact / interactive; charts (future). Timestamps felt "a little too late" — keep the live dot prominent.
10. Dark-mode toggle (default light) — portal layer.

Future (explicitly "not for this")

A universe-wide view (toward 200+ tickers) that's more than a heatmap/grade — tickers cross-referencing each other, the sheet vis-à-vis its own history, learning across names. North-star, after the 2-ticker brain + the above.

PART 6 — Cost basis, MARGIN (levered/unlevered), and the one-login hub (2026-05-27 cont.)

Cost-basis sources found in the repo (all STALE — do not trust against current positions)

• MIS/data/holdings_reconciled.csv — has avg_cost + total_invested per ticker/account, but as of ~May 1 (Schwab 898 shows AFRM/DAL/MRVL — since sold; now UNG). Trustworthy ONLY where current share count == reconciled count.
• MIS/History_for_Account_Z29720600.csv (0600) + Z29835692.csv (5692) — full Fidelity transaction history but March only (latest row 03/31/2026); current positions reflect April/May trades not in it. Format is clean (Run Date, Action, Symbol, Type Cash/Margin, Price, signed Quantity, Amount, Cash Balance) → replayable for cost basis IF a current export is pulled.
• MIS/ZT_Acct_XXX898_Transactions_*.csv + SCHWAB_898_TRADE_LOG_SEED.csv — Schwab 898 history.
• Clean fix: Sam pulls a current Fidelity "Positions" CSV (Fidelity app/site → Positions → Download) — it carries Cost Basis Total + Average Cost Basis columns AND the account summary (cash/margin). Drop it in → real cost basis + margin for every position, instantly. Same for a current Schwab positions export.
• Trustworthy NOW (exact share-count matches, reconciled avg_cost vs 5/26 price): 0600 SPY 10 @ $647.63→$745.64 +15.1% (+$980), ASML 4 @ $1,309→$1,633 +24.7% (+$1,294); 5692 AAPL 1.338 @ $277.56→$308.82 +11.3%, AMZN 1.19 @ $210.46→$266.32 +26.5%, EQIX 0.406 @ $984.50→$1,079.79 +9.7%, LLY 0.781 @ $967.57→$1,065 +10.1%, SPY 0.954 @ $682.69→$745.64 +9.2%, ORCL 1.267 @ $205.72→$192.08 −6.6%. (~+$2,540 unrealized on these 8.)

MARGIN — the critical gap Sam flagged

• 0600 is a Margin account; the March history shows running negative cash balances (margin loan). Schwab is now margin-approved too.
• The Position Summary email total is GROSS position value — it does NOT show the margin loan / negative cash. So the $26,737 in 0600 is NOT Sam's equity; his real (unlevered) equity = gross − margin borrowed. Need the current margin/cash balance (on the Positions export or account summary) to compute it.
• Build: a levered/unlevered toggle — show holdings (a) gross, and (b) net of margin loan = true equity; plus the margin balance + buying power. Sam wants to judge "with or without margin."

Sam's portfolio asks (this round) — DESIGN, sequence after cost basis is real

• Daily performance: how each position + the whole did today (and didn't) — needs day-change per holding.
• Performance over time: am I up for the month / quarter; vs my own target; why I'm holding (thesis per position).
• Fresh-eyes, bias-free recommendation per holding: run the FSE on what he HOLDS → REDUCE/EXIT/HOLD/ADD as if seen fresh today — explicitly fighting endowment bias (don't keep it just because you own it) and loss aversion (don't let it bleed because you won't sell). Tax-aware (when a sale makes sense after-tax / wash-sale), and context-aware: season/time-of-year, volume, where he's up to, other obligations, longevity, short- vs long-term, whether he needs the cash.
• Toggle Schwab/Fidelity on/off in the view; charts.
• One-login HUB (NOT the ops portal): Sam's own main page that routes to all his things/services — LevSMS, marketing-package builder, MIS (stocks), obligations, tracking, start-here — one interactive front door. North-star; its own design later.

Cross-references

• Plan: ~/.claude/plans/valiant-frolicking-trinket.md (Phases 2/3, HOLDINGS_CLEAN, Ticker_Memory, behavioral guardrail)
• docs/MIS_V2_DEFINITION_OF_DONE.md · docs/MIS_V2_STATUS.md
• Memory: reference_broker_trade_emails · project_mis_cashflow_thesis · feedback_classifier_first_doctrine · feedback_no_monday_morning_qb · user_sam_investment_philosophy
• GitHub MIS issues: #8 (holdings source) · #9 (backtest/learn-from-itself)