The Strategy Report

Definitions, in plain language, for every term the rest of the document uses. Skim it now; refer back as needed.

Quality vs. price — two different questions

The framework keeps two questions strictly apart. “Is this a good business?” is about quality — profits, debt, growth durability. “Is the share a good price right now?” is about valuation. A wonderful business can be a poor buy if it’s expensive, and a weak business can look cheap and still be a trap. Mixing the two is the classic mistake; we never do.

Deep Score (0–100)

Our single quality grade for a business, built from a 16-point value-investing checklist: profitability (return on equity/capital, margins), growth (revenue and earnings over years), balance-sheet strength (debt, free-cash-flow consistency), valuation discipline, and a moat/quality factor. It deliberately ignores the share price — it grades the company, not the stock chart — so price and momentum can sharpen a decision without contaminating the quality read. Bands: ≥85 exceptional, 70–84 high quality, 55–69 acceptable, 40–54 middling, below 40 “avoid.”

Fair value — a range, not a number

Our estimate of what a share is actually worth, produced by blending several independent valuation models (see §6). We never publish a single fake-precise figure; we publish a low–high range and a confidence level (e.g. “C$240–C$320, medium confidence”). A price inside the range is fairly valued; above the top is genuinely expensive; below the bottom is a genuine discount.

Concentration, weight & “look-through”

Weight is how big a position is. Per-account weight is its size inside one account; aggregate (look-through) weight is its total size across every account you own. The aggregate is the number that matters — the same stock held in a TFSA and an RRSP is one bet, and per-account numbers hide that. Concentration is having too much in one stock, one industry, or one currency.

Market regime

A plain label for “what kind of market are we in right now,” read from interest rates, the yield curve, credit spreads, volatility (VIX), and whether the market is above its long-term trend line. The five regimes are expansion (calm, risk-on), mid-cycle (neither stressed nor calm, with no downturn behind it — no regime tilt), recovery (climbing out of trouble: the S&P 500 closed below its 200-day average within the past 12 months), late-cycle (getting toppy, caution), and contraction (stress/downturn). We never predict a price; we just describe the weather. If any of those five readings could not be fetched, the regime reads Unknown: we name the missing reading, list no regime actions, and fill nothing in with an assumed value.

Canadian tax words you need

• Capital gain, 50% inclusion — if you sell something for more than you paid in a taxable account, half of the profit is added to your income and taxed. Selling a winner is therefore a real, immediate cost — not free.
• ACB (adjusted cost base) — your true average cost per share, used to work out the gain.
• Registered vs. taxable accounts — inside a TFSA, RRSP, or FHSA, gains aren’t taxed as you trade; in a non-registered/joint account they are. So we prefer to do any necessary trimming inside registered accounts first.
• US dividend withholding (15%) & the treaty — the US withholds 15% of dividends paid to Canadians — except in an RRSP/RRIF, where the Canada-US treaty waives it. In a TFSA or FHSA the 15% is simply lost (the treaty covers retirement plans only, not the TFSA or FHSA); in a taxable account you can usually recover it via the foreign tax credit.
• Tax-loss harvesting & the superficial-loss rule — deliberately selling a loser to offset gains — but if you, your spouse, or one of your own registered accounts buys the identical security in the 30 days before or the 30 days after the sale, the CRA denies the loss (the “superficial loss” rule). The safe move is to wait a full 30 clear days before rebuying. The report treats every listing of the company as that security — the same share on another exchange or in another account is held too.

A few more terms you’ll meet

• Beta — how much a stock tends to move relative to the whole market. Beta 1 moves with the market; above 1 is more volatile (amplifies both up and down); below 1 is steadier.
• Sequence-of-returns risk — the danger that a big loss arrives at the worst possible time — right before or early in retirement, or just before a goal — when there’s little time left to recover. The same average return can hurt far more if the bad years come first.
• Dispersion / confidence — how much our valuation models disagree with each other. Tight agreement = high confidence; wide disagreement = we say “unsure” instead of guessing.

A single generation pass executes a deterministic pipeline; the LLM is invoked once, late, with a fully-resolved context object.

1. Ingestion & enrichment. Holdings are joined to live prices, the latest fundamentals snapshot, account-type registry (capital-gains treatment, US-withholding flags), and the investor profile / IPS inputs. Per-lot FX conversion yields both price_native and price_cad; aggregate single-name weights are computed across accounts.
2. Quality & valuation. Each name carries a fundamental Deep Score and an ensemble fair-value point + range + confidence tier (§3–4).
3. Signal engine. ~28 deterministic rules evaluate every holding, emitting typed signals with strength scores; de-duplicated and precedence-ranked (§7). Concentration ceilings are tightened in hostile regimes (§8).
4. Portfolio analytics. Concentration (single-name, GICS sector, industry cluster, currency), beta-adjusted drawdown stress tests, and contribution-aware goal solve (§8–7, §11).
5. LLM narration. The resolved context (slimmed holdings, concentration block, macro, grouped signals, goal, profile, engine flags) is passed to the model, which returns structured JSON: verdict, ranked priority actions, diversifiers, simulation, scenario notes (§15).
6. Post-processing & render. Engine post-checks (look-through fund overlap; backfill of the contribution-dependency note) run on the model output; the designed multi-page artifact is assembled and served as printable HTML/PDF.

A 0–100 composite of profitability, balance-sheet, growth-durability and capital-efficiency factors, descended from a 16-point value-investing checklist. It is deliberately price-blind — it grades the business, not the security — so valuation and momentum can sharpen it without contaminating it.

A cross-sectional rank of the entire universe on the traits that recurred in history’s long-run compounders. Where the Deep Score grades quality today, Winner Odds asks which names have the makings of a business that keeps compounding and does not fail — and answers with a rank and a tier, never a probability. It is not a multi-bagger finder: on the point-in-time population test, top-tier names rose fourfold in five years about as often as any other name.

The factor set

Eight curated factors in two families, ranked against every other name using the same peer-relative robust-z statistics core as the Deep Score:

• Core quality — return on invested capital, gross profit over assets (Novy-Marx), the nine-signal Piotroski F-score, low accruals (Sloan), and margin stability.
• Core growth — growth acceleration, the Rule of 40, and reinvestment runway.

Value and momentum are computed as a side overlay only — never blended into the quality-and-growth rank, so a name cannot climb on cheapness or price action alone.

Percentile → tier, with hard gates

Two hard gates override the tier regardless of the factor score: financial distress (a weak Altman-Z) caps a name at Unlikely, and a cyclical business at a likely earnings peak caps it at Watch.

A forward-looking companion to the Deep Score. It reads the rate of change — is the business getting better, faster — to catch businesses that are turning a corner before the trailing averages the Deep Score relies on can show it.

The inflection signals (engine 3.0.0, 2026-09-28)

Eleven signals, each measured as a trend (recent vs. earlier), not a level. Six carry weight: each is the name’s percentile among its sector peers that night, and the score is the weight-averaged percentile, re-based to the published bands. The weights are each signal’s measured rank correlation with the three-year outcome on every US 10-K filer FY2009–2022, scaled to sum to 100. The other five are shown as facts with no weight.

Guards

• Cyclicality guard — commodity / cyclical names whose “acceleration” is just the up-leg of a cycle are capped at 50, so a rebound isn’t mistaken for a structural inflection. Heavy dilution caps a name at 50 the same way.
• Confidence band — names with less financial history score fewer signals; the band discloses how much data stood behind the number, and fewer than four of the six weighted signals gets no number at all.

Intrinsic value is estimated by a multi-model ensemble (two DCF variants, EV/EBITDA, Graham-style, dividend models, etc.), blended with dispersion-aware confidence. The output is a point estimate, a low–high range, a coefficient-of-variation, and a confidence tier — never a single fake-precise figure.

1. Applicability filter. Keep models with a positive fair value applicable to the sector (e.g. exclude FCF/EBITDA constructs for financials via sector policy).
2. Price-anchored sanity gate. Drop any model outside [0.33× , 3×] of the current price — these reflect broken inputs (e.g. a normal multiple on peak-cycle EPS).
3. Median outlier trim. Drop values outside [0.40× , 2.50×] of the median.
4. Family reduction. Average within each model family → one representative per family (prevents a single methodology from dominating).
5. Blend & dispersion. Point = mean of family representatives; range = their min/max; dispersion = coefficient of variation (how widely the models disagree, σ/μ).
6. Confidence tiering. CoV < 0.15 → HIGH, < 0.30 → MEDIUM, < 0.50 → LOW, > 0.50 → NOISE (suppressed — no trustworthy estimate; caller skips the name rather than print a fantasy number). Requires ≥ 2 corroborating families.

Acknowledged bias: the ensemble is conservative on steady mega-caps; in a genuine melt-up it will correctly flag names well above range — that is signal, not error.

Rules evaluate per holding and emit typed signals with 0–100 strength. Precedence: SELL / ROTATE (exits) → TRIM → HOLD → ADD; TAX fires independently. De-duplication keeps the strongest signal per (ticker, account, type); a TRIM rounding to < 1 share is suppressed as a no-op.

NEW What a holding is, and which of its numbers are current (22 October 2026)

• One company is one position — a Canadian depositary receipt and its US share, held in two accounts, were two companies to every grouping, so one report could ADD to the receipt while it TRIMMED the share, and size each against half the company. Weights, trims, room to add, the diversification count and the plan projection are now keyed by company (issuer.py), and a TSX fund’s US-dollar units (XUS-U.TO beside XUS.TO) are the same fund — one weight, one position in the count, one fund to the held set, the chat and a loss harvest (the stock rules neither size nor trim a fund). A loss harvest holds back the sizing panel’s room to add as it does the engine’s ADDs.
• A score comes only from the ranked batch — a Deep Score is a percentile of one night’s ranked universe, so a row the scan stopped writing carries no score, and a price-anchored number (close, fair value, trend) is current only from a row the scan still covers. A receipt is researched as its underlying at the receipt’s own price; a fund never carries a stock-table score. Every surface that rates a holding reads its research through one decider (holding_research.HeldResearch), and an architecture test fails on a new per-ticker read of the stock table that is not registered.
• Every engine action reaches the plan — the model chooses which actions to feature; every action the rules raised that it left out is added under the verb the rule asks for, and an action the before/after projection cannot model is listed as unmodelled rather than skipped. Measured on production before the fix: 4 of the latest 5 reports left engine actions out of the plan.
• A failed read is stated — the portfolio loader raises on a database error instead of returning an empty book.
• The CAD/USD indicator is the quote it names (23 October 2026) — the US dollars one Canadian dollar buys, with the loonie’s own direction. Since June 2026 the macro feed had stored the inverse quote (Canadian dollars per US dollar, the Bank of Canada’s FXUSDCAD) under that label and called a rise in it the loonie strengthening — the opposite of what a rise means. The two currency rules, Market Pulse, the Telegram briefing, the report’s macro cell and the AI’s narrative read that direction backwards; a report stored before the fix keeps its narrative, and its FX cell reads “—”.

NEW The allocation policy, and the trade a breach implies

Where a client has declared an investment policy (target weights for equity, fixed income and cash, plus a drift band), that policy is the first lens and a breach is the first action in the plan — ahead of any single-name idea, enforced by the engine after the narrative is written rather than requested of the model.

The policy is compared with the client’s own risk profile. The risk questionnaire scores risk tolerance (the crash answer, experience and purpose) and risk capacity (horizon and income) separately, and the profile is the lower of the two (Companion Policy 31-103CP s.13.2; Joint CSA/CIRO Staff Notice 31-368 — reference norms, not a compliance claim). A saved policy holding more than 5 points more equity than that profile, or than the risk tolerance the client declared, is stated as a recorded departure on every surface that carries the policy; the policy is never changed for the client. The band edges, the caps and the 5-point threshold are house decisions recorded in the model register. A plan is a review draft until the client explicitly confirms the six minimum profile fields are current; saving a form does not confirm them. riskprofile.py · advisor_engine._ips_drift · portfolio_db.confirm_suitability_review

When a breach has already been shown, the report asks about the policy. CFA Standard III(C) says an investment policy is reviewed at least annually, and a client holding an off-policy book is discussed with: the trade is made or the policy updated. The report asks that question when the book was outside the same policy, in the same class and direction, in the report before (and every report back to the first such one), and the policy has not been saved since; or when the policy was last saved more than 365 days ago. Re-saving the policy restarts the clock. A report inside the band, or one whose lens failed, ends the chain. An unknown save date triggers nothing. The question never proposes moving the policy to match the book.

A declared cash reserve is outside the policy. Liquidity needs are the first constraint a policy is written under (CFA Standard III(C)). The client declares a reserve in dollars; the mix removes min(reserve, cash) from the cash class and from the base every percentage is a share of, so drift, the policy trade and the contribution plan are computed on deployable money only. A reserve larger than the cash on file is reported as a shortfall and is never made up from a GIC or a holding; new contributions rebuild it first. With no reserve declared the value is unknown, not zero: no number moves, and where a cash-over breach draws on savings the client entered, the report states how much of the cash that is and asks. The reserve is never inferred from an item's name or from months of expenses.

Each fund has one asset class, decided in one place. The policy mix and the fund score used separate keyword tables and disagreed on 17 of 247 funds, one of them held: a gold-bullion fund the mix booked as equity. One decider (etf_scoring.etf_asset_class) now serves both, following the CIFSC categories: a money-market or treasury-bill fund is cash, an ultra-short or preferred-share fund is fixed income, and a commodity fund is its own class. A commodity or alternative fund is reported on its own named line outside the equity / fixed income / cash policy — not as equity, and not as unclassified. A CIFSC Fixed Income Balanced fund (more than 5% and under 40% in equity) is a balanced mix, not a bond fund: “balanced” in the category outranks the bond words beside it. The all-in-one check starts from this decider, so a fund it classes as fixed income is never treated as an all-in-one portfolio because its name says “conservative” — a risk level, not an asset class.

• Every breached class, not only the largest — the over-weighted class and the under-weighted one are a single trade. Naming only the bigger side leaves half an instruction: a book 24pp over in cash and 20pp under in fixed income was previously told “reduce cash toward 5%”, and the destination was the second breach.
• In dollars, and the legs balance — every class moves to the target the client declared, so what is added equals what is reduced by construction. A class inside its band still appears, because it is where money comes from or goes; a rebalance naming only breached classes does not add up.
• The funding source, because it decides the cost — money moved out of cash is not a disposition and realises nothing; selling equity or bonds realises a gain or loss in a non-registered account and no taxable gain inside a registered plan (the plan’s trust is not taxable: ITA s.146(4) RRSP, s.146.2(6) TFSA, s.146.3(3) RRIF, s.146.6(3) FHSA). The sale is split by account type — registered, non-registered, and any account whose type is not on file, said to be unknown — and registered dollars are drawn first. New contributions are offered first because they move the mix without a sale. NEW What the classes over their targets do not hold is not booked to an account of unknown type: it is named as new contributions or a sale of holdings outside the policy, which the plan does not size.
• NEW The destination is named (22 October 2026) — each fixed-income and cash leg that takes money names the top screened fund in each listing currency (the fund score’s top tier, the client’s exclusions applied), or says that none reached it — never a fund from memory. A sale’s prose may not send its proceeds into the class that is already over, or into anything the suitability check would hold; under a breach of the sold class the engine writes the destination itself. A tax-loss replacement that would deepen an overweight sector, or that the policy would hold, is withdrawn and the withdrawal is stated. Measured on production before the fix: of 23 sales in reports with a class over policy, 16 named a redeploy ticker and 15 of those named none in the under class. A fund the plan is selling for a tax loss is dropped from that screen, so it is never named as the place for the money (ITA s.54); a buy the model writes is checked against the whole company, so an ADD on a name harvested or reduced under another listing or in another account becomes a hold.
• The tax itself is not estimated. It depends on which lots in which accounts are sold, and the trade builder is handed neither — so it reports how much requires a disposition and stops there rather than publishing a plausible figure.
• Unclassified money is named, never absorbed — holdings the look-through cannot split into an asset class are excluded from the trade and disclosed, so the class figures are not quietly inflated.

Concentration (the first lens)

• Single-name — measured on the aggregated weight across accounts (the same issuer in a TFSA and an RRSP is one position); honours a client-specified max_single_position_pct, else a 25% ceiling, trimming toward ~70% of the ceiling (18% at 25%).
• Sector & industry — a GICS sector ceiling (see regime-aware tightening below); a finer industry cluster lens (e.g. semiconductors as one cyclical complex) is surfaced because it is often the real risk even when the “Technology” sector merely looks large.
• Currency — a USD-exposure ceiling; trimming is registered-account-only and never applied to a taxable winner for a soft FX reason.

NEW Regime-aware concentration ceilings

Concentration hurts more when markets are fragile, because in a downturn correlated names fall together. So the sector and currency ceilings are now a function of the market regime rather than fixed:

The tightening is deliberately modest (≤7 points) and the rationale string names the regime and the active ceiling, so it is transparent. When the macro feed is unavailable, the ceiling falls back to the base 45% — missing data can never tighten anything.

Drawdown & horizon

• Beta-adjusted stress tests — dollar-denominated downside under semiconductor-downcycle, broad-bear and severe-recession scenarios. The client’s stated max-drawdown tolerance is tested against the worst severe case modelled — the severe-recession row or the book’s own crisis replay (a window with at least 50% coverage) — and the note names which and over how much of the book; once it is breached, any addition with market exposure is held. With no case modelled the check is listed as not judged, never as passed. Each holding falls by its own beta (a CIBC CDR by its US company’s); a holding with no beta on file is left out and named, never given one, and the book’s value after a market-wide fall is then not stated (the Goal Risk line then says the shortfall cannot be stated). With no beta anywhere nothing can be modelled, and the rows state no loss and no value after.
• Dual horizon — the engine reconciles risk capacity from the retirement horizon with the nearest funded goal: where a major drawdown landing near the goal year would cause a shortfall, it surfaces an explicit Goal Risk and a light glide-path — without forcing the whole book conservative. Sequence-of-returns risk dominates inside ~7 years.
• Deterioration sells fire BEFORE price — note that several SELL rules (dangerous leverage, shrinking revenue, an AVOID-rated score the business itself confirms) trigger on the business deteriorating, independent of any price drawdown. The 20%-drawdown “Crash Watch” (§9) is a separate, last-line protective trim, not the only defence.

The governing policy — chosen for an after-tax, buy-and-hold personal book — is that valuation alone does not trigger a SELL of a quality compounder: a rich price routes to HOLD (stop adding), preserving compounding and avoiding a realised capital gain. SELL is reserved for genuine deterioration (§7). Extreme valuation is handled not by a valuation-SELL but by the concentration ceiling (no name runs unbounded as a share of the book, §8) and the goal-return lens (which can redeploy on forward-return grounds, §11). We accept the trade-off openly: a pure valuation discipline would occasionally sell a very expensive great business; we weigh that against tax friction and the cost of interrupting compounding.

The single sanctioned price-driven protective action is the crash-risk overlay: the engine does not trim a winner for being up or richly valued, but when a quality holding has rolled over — drawdown ≤ −20% from its 52-week high plus a confirmed trend break (sub-200-DMA, a sharp 3-month decline, or a realized-volatility expansion) — it raises a “Crash Watch” and proposes a ~25% protective trim or a stop, the client’s discretion. Sizing is tax-tiered (realise in registered first; quantify the gain in non-registered). A normal pullback with the trend intact is tolerated and produces no signal.

NEW Tax-neutral rebalancing — fund changes with new money first

The framework’s standing preference is to reach a target allocation without realising a taxable gain wherever possible. Two mechanisms enforce this:

• Contributions as the primary de-risking lever — when the contribution stream can pull an overweight industry to target within ~3 years, the engine computes that path and prefers directing new cash over selling winners; it states the dependency (and a fallback floor if contributions fall short). This logic predates this revision and lives in the Strategy report’s action plan.
• Tax-Neutral Check in the manual Rebalance tool — when a rebalance plan would realise a capital-gains tax, the preview now shows a Tax-Neutral Check: it quantifies the estimated capital-gains tax the plan triggers and, where the increases could instead be funded from idle cash or upcoming contributions, says so explicitly — and confirms when the buys fit entirely within idle cash and need no sale at all. The reflex becomes “add with new money before trimming a taxable winner.”

A separate Tax & Structure Audit report handles full contribution-room / placement modelling; the Strategy report keeps tax to a single room-nudge and per-action consequence.

Regime & forward outlook

Current regime is classified from VIX, the 10Y–2Y curve, the high-yield spread and the S&P 500 against its 200-day average, into contraction, late-cycle, expansion, mid-cycle or recovery. A recovery needs a downturn behind it: the S&P 500 must have closed below its 200-day average at some point in the past 12 months; otherwise a market that is neither stressed nor calm is mid-cycle, with no regime actions. NBER dates only expansions and recessions and speaks of “the recovery” as what follows a trough; the 12-month window is a house convention. Until 22 October 2026 every such market read “Recovery / Stabilising” and fired the recovery ADD.

A forward outlook ensemble — yield-curve slope (10Y–2Y), HY credit-spread trajectory, index vs 200-DMA, VIX trend, and CPI direction — produces an outlook_pressure on a −1 to +1 scale, a heading and a confidence. The pressure is a weighted score of indicator trends, not a probability: nothing in it is fitted to recession outcomes (the New York Fed’s yield-curve model, by contrast, is a probit fitted to NBER-dated recessions). Until 22 October 2026 it was printed as a “transition probability”. Deteriorating readings bias defensive timing toward act-now / within-30-days; the regime itself (§8) tightens concentration ceilings. It is not a price or index-level forecast.

Goal-based framing

• Contribution-aware required return — the required CAGR is solved (binary search) including a level annual contribution, since the balance-only figure materially overstates difficulty when contributions are large; on-track is benchmarked at ≤ 8%.
• Goal-return lens — idle cash and any holding whose realistic forward return sits below the required rate are framed as opportunity-cost drags — candidates to redeploy into higher-expected-return diversifiers (tax-aware), not a reason to sell a quality compounder.

• Objective-aware — growth / aggressive or 12+ yr horizon keeps a growth tilt (quality compounders in different growth industries; broad/growth funds), explicitly not defensive sectors or bonds; income / capital-preservation / < 7 yr tilts to dividend, low-beta, broad-market or fixed income.
• Grounded single-name candidates — sourced from the client’s own scored universe via a GARP screen (deep-score ≥ 70, D/E < 1.0, ROE ≥ 15%, earnings/revenue CAGR ≥ 12%, valuation gate), with multi-year downtrends and currently-held companies excluded under any listing (a receipt held as ASML.TO excludes ASML); the LLM may not invent a ticker — a single name off the screened menu is removed from the plan and recorded. A candidate priced above its fair value ranks behind one priced below it, whatever their scores — a house ordering, not an external standard.
• Phase-in vs deploy-now — each candidate carries an engine-decided deploy_mode: at/below fair value → deploy idle cash; above fair value → accumulate via contributions / on a pullback toward the 200-DMA.
• Currency-matched — deploy each account’s idle cash in its own currency to avoid an FX spread and a taxable conversion event.
• Look-through overlap — a recommended broad fund is checked against the client’s overweight names; material single-name/mega-cap overlap is disclosed rather than silently recommended.
• NEW Where the next contribution goes — for a client still saving, what they are about to add outweighs what they hold. The report splits the annual contribution across asset classes in dollars: after a contribution the book is its present value plus that amount, each class’s target is its policy weight times that, and the contribution is shared in proportion to each class’s shortfall. A class already above its target receives nothing; nothing is sold; growth is ignored, so every horizon is conservative and no return is assumed. Where a year of contributions reaches the policy the report says so beside the rebalancing trade; where it would take longer it says how many years. It also states what the contribution does to a concentrated sector. It needs a declared policy, and it does not say which account the money goes into.
• NEW Funds are screened and measured, not recalled — a recommended fund comes from the scanned fund universe, never from the model’s memory. To be offered it must carry known sector weights, classify as equity, sit in the fund score’s top tier, and hold no more than 40% in any one sector — the same threshold at which this report calls a client’s own book concentrated. Funds are ranked by their weight in the client’s overweight sector, and each recommendation states that weight and what the book’s sector weight becomes if new money equal to a tenth of the book goes into the fund. That is arithmetic on holdings, not a forecast of return. Where a sector exceeds 40% of the book and the recommendations name no screened fund, the engine places the first one at the top of the list, with no position size. The quality floor inherits the fund score’s trailing-return component, so it leans toward funds that have done well; sector weights are on file for 182 of 245 scanned funds, and the other 63 cannot be offered.
• NEW What “diversifies” rests on — the screen selects candidates outside the client’s overweight sectors, and until now the word “diversifies” stood on that alone: the candidate record carries ticker, name, sector, Deep Score, the growth-pillar fields and valuation fields, and nothing about how the name moves with the book. A different sector is a real basis and a weak one — two GICS sectors can move together closely, which is precisely why the Risk X-Ray measures effective bets rather than counting names. Each single name now states that basis, and where price history allows it carries the measured change in effective bets, computed through the same risk_data / risk_concentration pair the X-Ray uses so the two cannot disagree about one book. Coverage is a minority and is named as such: measured on 2026-09-24, 4 of 10 live candidates carry the ≥60 days of history a correlation needs, and the other six read NOT measured rather than staying silent — silence beside a diversification claim reads as assent. A candidate that measurably does not diversify is disclosed as such, with the recommendation left standing and the contradiction stated beside it.
• NEW What each pick did in past crises (15 October 2026) — every single-name diversifier, and every name on the report’s watch-list, carries a crisis record: for the 2008 GFC, 2020 COVID and 2022 rate-shock windows, the largest peak-to-trough fall of its own stored total-return closes from the window’s start to that crisis’s market low (the S&P 500’s closing low; since 16 October 2026 — a later fall in the recovery years is not that crisis’s), the dates, and whether it regained the peak by the window’s end. A window it cannot measure says why — no price on record before its first recorded close, price history not stored that far back, or could not be read — and carries no number. It is never an estimate: a beta-times-market figure was considered and declined, because it would put a model’s guess beside measured falls. The rule for which windows a stock was in is the one the portfolio crisis replay and the stock page use (stress_replay.window_coverage), and the record is written by the engine (attach_crisis_profiles), not the model; the model is given it and forbidden to state, estimate or imply a fall for a window it was not given. After every stock scan a refresh keeps the latest ranked scan’s BUY-or-better names and every held name stored over their whole history on one adjustment basis. A past fall is not a forecast.
• NEW The screen can suspend itself — if graded cohorts persistently underperform (three, a recorded decision rather than a fitted number), the report stops recommending single names entirely and reverts to allocation-level advice; ETF recommendations continue, because allocation rests on a declared policy, a tax rule and an arithmetic identity and is not what failed. The rule cannot fire yet — the grader is network-bound and not in CI, so no cohort verdict is stored where the report can read one — and it returns a state with its basis (“no graded cohort yet”) rather than a boolean, because “not suspended” and “not yet judgeable” are different facts and a boolean collapses them into the reassuring one.
• NEW A single name is research, never sized — an individual company arrives labelled Research with no suggested position size, and states how much graded forward evidence stands behind the screen. A size is the sentence that turns research into advice: “consider this name” and “put 3–5% into this name” are different acts, and the second implies a basis this screen does not have — no pick has a graded forward return yet. Measured before the change: 180 of 180 single-name recommendations across 48 stored reports carried a weight. An ETF keeps its weight, because an allocation decision rests on the client’s declared policy, a tax rule or an arithmetic identity. Stripped at the engine choke point rather than requested of the model (the prompt already asked and produced 180 of 180 the other way), and guarded in both renderers so reports written earlier stop showing a size too.
• NEW Growth-pillar dependence, disclosed by the engine — where a recommended single name owes more of its Deep Score to the Growth pillar than an evenly-scoring name would — above the pillar’s own weight share of 25 of 100 points — that share is stated in the recommendation itself. Both of the pillar’s inputs (revenue_growth, earnings_growth) are historical, and the literature is against extrapolating them: Chan, Karceski & Lakonishok (Journal of Finance, 2003) find no persistence in long-term earnings growth beyond chance, and Lakonishok, Shleifer & Vishny (Journal of Finance, 1994) find the fastest-growing names underperform over five years. Measured on the 2026-09-24 production scan of 2,577 scored names: the rank correlation between the shipped score and the score with the pillar removed is 0.97, yet 45 of the top 50 names are in that set only because of it, and 6 of the 8 names the live screen surfaced were Growth-lifted (mean 20.6 of 25 against a universe mean of 12.5). The pillar is disclosed, not re-weighted — its sign has not been tested on its own. The 1,059 names scored on 8 June 2026 were graded at 126 days on 14 October 2026: over the 998 that could be graded, the pillar showed no detectable sign either way against returns relative to each name's own peer group and size tier (rank correlation +0.04, 95% interval −0.02 to +0.11), a reading the 14 names that could not be priced could have tipped. That is one window in one market regime — a diagnostic, not the test — so it moves no weight, and re-weighting on the literature alone would replace weights the population test measured together with weights nothing has measured. The note says screen, not forecast: no forward claim is made, because none has been validated (CFA Standard V.A).
• NEW Single-name concentration, stated against a budget — the share of the book held as individual securities rather than funds, measured by value and shown whenever it exceeds the 20% default budget, with its basis: only 42.6% of roughly 26,000 US stocks since 1926 beat one-month Treasury bills over their lifetime and the best 4% of firms account for all net wealth creation (Bessembinder, Journal of Financial Economics, 2018), and removing 90% of diversifiable risk takes 40–50 names rather than 15–20 (Statman 1987; Domian, Louton & Racine 2007). Disclosed, not enforced — no trade is blocked and no plan rejected for exceeding it. Measured 2026-09-24: all six live books exceed it, from 62.9% to 100.0%.
• The engine writes these, not the model — both notes are appended after the narrative has been written, at one choke point (_apply_engine_disclosures). A field placed in a prompt is a request; the model writes its own reasons and is free to leave the inconvenient half out. Neither pass changes a score, a rank, a pick or a position size.
• Contributions as the primary de-risking lever — see §10; directing new cash is preferred to realising gains, with a computed dependency and fallback floor.

The best frameworks make room for genuine human knowledge without letting it quietly corrupt a deterministic engine. The conviction override is that bridge.

• Structured override — on any holding, the investor can overrule the engine’s recommendation (set it to ADD / HOLD / TRIM / SELL). The engine’s own verdict always remains visible beside the override, so the disagreement is explicit, never hidden.
• Documentation is compelled on a contrarian call — when the override softens a protective verdict — the engine said SELL or TRIM and the investor overrides to HOLD or ADD — the system blocks the save until a written thesis is entered (“why, and what would prove you wrong?”). Agreeing with the engine, going more cautious, or clearing the override needs no thesis. So conviction is allowed, but never undocumented.
• Append-only audit trail — every change is written to a conviction log — a dated, append-only history showing the stance change (e.g. Auto → ADD) and the thesis. Nothing is overwritten; the record of why a position was held against the engine survives for later review. This mirrors the Sector Lab conviction journal already used for sector theses.

A good framework keeps score of itself — honestly — so its parameters can be improved by human judgement over time. It does not secretly rewrite its own weights.

What is measured

• Fair-value accuracy — stored fair values are graded forward against the realized price weeks later: a directional hit-rate (did the under/overvalued call match the later move?) and a median absolute % error, broken out by confidence tier and sector.
• Verdict forward grade — because every scheduled scan appends a dated snapshot of each name’s quality verdict (STRONG BUY … AVOID), the engine can grade its own calls as they age: for each verdict it measures the realized forward return 126 and 252 calendar days later (about four and eight months), bucketed by band, and checks the basic monotonicity that STRONG BUY/BUY should beat AVOID. Each band states how many of the names it labelled were graded, and every name left out is named with the reason, so coverage is never silently overstated. This is a forward grade, not a back-test, and nothing has been graded yet: no gradable cohort has matured (the 8 June 2026 scan stored no closing prices, so the first gradable cohort is 15 June 2026, which reaches 126 days on 2026-10-19 and is graded once a close dated on or after 2026-10-24 is stored), so no hit rate is published and the labels have not been tested against returns. It reads each name’s own stored dividend- and split-adjusted closes, so a split inside the window is not a crash; a name with no stored close at the scan or at maturity — a delisting, or a name whose stored closes stop before maturity, among them — is left out and named with the reason. A separate point-in-time population test (2026-10-06, US filers FY2009–2022) found the Deep Score’s fundamentals ranked later peer-relative returns modestly well; its label cut-offs, valuation inputs and the sizing bands were not tested. Even when graded, this measures signal accuracy on accrued history — not the investor’s realized portfolio return.

PRINCIPLE Measurement informs people — it never auto-tunes the engine

This track record is read by a human to decide whether to recalibrate — exactly as the sector multiples and normalized-earnings model were deliberately re-tuned in June 2026. No code path adjusts the Deep Score weights or fair-value parameters automatically from outcomes. We reject self-tuning on principle (§16): on a single portfolio of a few dozen names over a couple of years it would overfit noise, chase recent performance — the very behaviour value investing exists to resist — and quietly convert a transparent, deterministic calculator into a self-modifying advice algorithm, which is precisely the NI 31-103 line the product guards.

• Constrained role — the model (Claude, structured-JSON output) validates signals, ranks ~8 priority actions, adds timing/framework/tax narration and a simulation. It cannot fabricate a price, fair value, SELL, or out-of-universe ticker.
• Flag demotion — deterministic decisions (deploy-mode, contribution-dependency, fund-overlap) are computed in the engine and merely narrated, with engine-side backfill so a flag cannot be silently dropped under prompt load.
• Audit trail — wherever the model’s final action differs from the deterministic rule signal for a name, the override is disclosed with its rationale (“rule said TRIM → advisor chose HOLD: …”).
• Test & ops — the rule base and renderers are covered by an automated suite in which every signal rule carries at least one threshold test; the data layer self-heals its schema across deploys; the report degrades gracefully on missing inputs.

Common questions a reviewer asks, and the deliberate reasoning behind each answer.

Why a semiconductor-cluster heuristic instead of a live correlation matrix?

For a personal book of ~10–30 names, a rolling correlation matrix is fragile: correlations are noisiest exactly when you would act on them (in a crash everything correlates toward 1), so it would generate the most sell pressure at the worst time. The honest cluster heuristic (treating, e.g., semiconductors as one cyclical complex) captures the concentration that actually matters, and the new regime-aware ceilings (§8) supply the “tighten when fragile” behaviour without the instability or the false precision.

Is the tax handling reactive or projective?

Both, by design. Per-action tax consequences are reactive (they price the cost of a trade you’re considering). But the plan is projective: contributions are modelled as the primary de-risking lever, and the Rebalance tool’s Tax-Neutral Check (§10) steers you to reach targets with new money before triggering a taxable disposition. Full contribution-room/placement projection lives in the separate Tax & Structure Audit, so the Strategy report stays focused.

How does human intuition enter without introducing model bias?

Through the conviction override (§13): a structured, documented, append-only override that sits beside the deterministic engine rather than inside it. The investor may overrule a rating, but overruling a caution compels a written thesis, and every change is logged with a date. The engine stays deterministic; the human judgement is explicit and on the record.

The 20%-drawdown trigger is “late by design” — isn’t that dangerous?

The Crash Watch is intentionally late because it protects a winner you want to keep after a tolerated pullback — trimming early at every top would forfeit the compounding that is the whole point. Crucially it is not the only defence: fundamental-deterioration SELLs (dangerous leverage, structural revenue decline, an AVOID-rated score the business itself confirms — §7/§8) fire on the business breaking, ahead of price. So catastrophic-failure protection comes from the fundamentals path; Crash Watch only guards a still-good name whose chart has clearly rolled over.

Why not let the engine self-optimise its weights from outcomes?

Because a self-tuning loop on one small portfolio over a short history overfits noise and chases recent performance — the opposite of disciplined value investing — and it would turn a transparent, auditable calculator into an opaque, self-modifying advice algorithm (an NI 31-103 concern). We build the measurement (§14) and let a human recalibrate deliberately, which is both safer and more honest.

• Rules-based signals, not personalized advice. The report applies published rules to the investor’s own inputs; its SELL/TRIM/HOLD/ADD outputs — some with a suggested share count — are general information and starting points for the investor’s own judgement, not personalized advice from a registered adviser. The investor makes every decision, and nothing executes automatically.
• No price/level forecasts. Forward statements are a pressure reading on the regime’s direction (−1 to +1, not a probability), never targets.
• Fair-value conservatism. The ensemble is conservative on steady non-mania names; the range + confidence tier is the explicit mitigant against false precision, and NOISE-tier names are suppressed.
• Crash Watch is late by design. The trend-rollover overlay protects the residual after a tolerated drawdown, not the peak; the fundamental-deterioration SELLs are the earlier line of defence.
• No company-news trigger. Confirmation is price-and-fundamentals based; idiosyncratic event risk (earnings, litigation, M&A) is not modelled in the deterministic engine.
• Validation is measurement, not control. The track record informs human recalibration; it never auto-tunes engine parameters.
• Data dependency. Outputs are only as current as the latest scan; the forward track record accrues forward history over time and states its own coverage.