How it works
The whole chain, what each step claims, and what it hasn’t proven.
- 1Macrothe weather
- 2Screenernarrows the market to a list
- 3Analystreads each name
- 4Gateearlier analyst · stopped
- 5Follow-upwatches daily · being rebuilt
- 6Gradingscores the record
Welcome — what this site is, and what it is not
Stockpeak is a research system published as a website. Every day a computer reads the financial reports and price history of roughly 7,000 US-listed stocks and narrows them to a short list. The AI analyst, a cloud AI model, then reads the filings of the names on that list, one company at a time, and its analysis appears on each stock page. This handbook explains how each step works, what every label on the screen means, and what the system has not proven.
- The screener picks a monthly book of roughly 50 to 60 stocks. Two paper books follow it from 6 November 2026: the Code book (the screener's picks alone) and the AI book (the same picks, with a slot held as cash when the AI analyst's reading of the filings finds a material event). Both are simulated; no real money is involved.
- The AI analyst reads the filings of the names on the screener's list. Its analysis is on each stock page, with a value only where one can be computed. Many names have no value, and many values sit far from the market price, so read the analysis before the number. The AI analyst has not been graded on returns. See The AI analyst.
- The earlier verdict analyst, with its gate and its AI paper book, was stopped in October 2026 after it was ruled invalid. Its record stays on the Track page, labelled as the stopped analyst's. See What is live, stopped and next.
- A daily follow-up of held names, reading new filings and news, is being rebuilt. It is not running yet. Real-money mirroring is halted; this site never trades.
- Nothing here is proven yet. The comparison that matters, whether the AI book beats the Code book, has not been measured. See How it is validated.
If a word has a dotted underline, it is a technical term — click it to see what it means without leaving the page. A complete, searchable index of every term lives in the glossary section.
The chain, every day — from the economy to a graded record
Behind the page is a chain of six steps. The order is enforced and each step hands the next only a data file, so stale or partial data cannot quietly corrupt the list. The cloud part re-runs on a schedule via . In plain words:
- Macro — the weather. A separate engine reads US economic data () and publishes the economic backdrop and the cost-of-capital number the analyst discounts with. It is context, not a call on the market. See The macro engine.
- Screener — what the analyst reads. Filings, prices and analyst forecasts come in. Safety filters remove stocks nobody could sensibly buy, three — compounder, value gap and trend leader — score what is left, the best cases make the list, and bands sort it. Its job is to decide where the analyst's limited time goes. See The quant screen and Bands, vetoes & warnings.
- Analyst — reads each name. The AI analyst, a cloud AI model (Gemini, with Claude as the fallback), reads one listed company's filings at a time and writes claims by topic; code checks each claim against the filing. Where the filings allow it, code computes a value from the AI analyst's choices. See The AI analyst.
- Gate — the earlier analyst's check (stopped). Code re-judged each verdict of the earlier analyst and marked it blocked or not, with the reasons. It stopped with that analyst; its section is kept as the record. See The gate.
- Follow-up — watches daily. A daily check of held names against new filings and news is being rebuilt. It only ever asks for a fresh look; it never sells. Not running yet. See The daily follow-up.
- Grading — scores the record. Two monthly paper books, Code and AI, are marked every day against benchmarks, beside a control book and your own holdings. The earlier analyst's verdicts were graded against later prices; the AI analyst has not been graded on returns. See Track record.
- 1
Macro
Publishes a for each economic season, alarms and the cost-of-capital anchor. Context only: the sector tilt is off.
hands on: probabilities · shock alarms · cost of equity
- 2
Screener
Safety filters, then three — compounder, value gap, trend leader — each . The better door competes, a gets a small bonus, and keeps the list steady. Bands sort the result.
hands on: the list · bands · ranks
- 3
Analyst
● live · analysis on stock pages
The AI analyst, a cloud AI, reads each company's filings and writes claims by topic; code checks each claim against the filing and, where the filings allow, computes a value from the AI analyst's choices. The AI analyst has not been graded on returns.
hands on: analysis · value where computable
- 4
Gate
Marked each verdict of the earlier analyst or not, with the . verdicts stay visible as a record and are never a recommendation. Stopped with that analyst.
hands on: actionable · reasons
- 5
Follow-up
● being rebuilt · not running
A daily check of held names against new filings and news is being rebuilt. It only asks for a fresh look, and it never sells.
hands on: watch items · buy paused · re-analysis queue
- 6
Grading
Two monthly paper books, Code and AI, are marked every day against QQQ, and , beside the and yours, with real prices and assumed costs. The earlier analyst's verdicts were graded at 30, 91, 182 and 365 days.
hands on: scoreboard · paper records
Each step hands the next only a data file. The cloud part re-runs on a schedule via GitHub Actions; the AI analyst runs separately.
Everything is : each logged signal uses only information that existed at that moment. That discipline is what makes a track record worth reading.
Reading the desk
The desk is the main page. From top to bottom:
- The freshness strip — when prices, the book and the latest verdict were last updated, a health dot, and the macro read on the right. See System health.
- A notice, when something needs saying.
- The status line — one sentence: how many list names the AI analyst has read, how many have a value, and the date of the newest reading.
- The funnel — five counts: US stocks → pass safety → make the list → AI read → valued. Clicking a step only selects it: the summary below changes, and “show how” opens the detail. It does not filter the table. Where a step removes names, its detail panel says why.
- One table for the list, in the sections below, with filters for door, sector and search. A filter that hides rows always says “showing k of n”.
The sections of the table
- Research now — the screener's top names, roughly the first 50 to 60, in screen order. It is a list of names worth your research time, not a buy list.
- Watchlist — the next names on the screener's list, in screen order. This is the screen's own band name; it is not a verdict.
- Disqualified — on the list but removed by a safety filter. The reason is shown beside it.
The AI analyst's reading sits beside each name. Verdicts from the no longer place names; one quiet line counts them, and each stays on its stock page under verdict history.
Reading a row
Each row shows the company, today's price and the screener's rank. The AI analyst's reading sits beside it: when it was read, the value and its range where one could be computed (otherwise the reason), the gap between that value and the price, and how many of the AI analyst's claims were checked against the filings. A large gap is a reason to read the analysis, not a signal: many values sit far from the price. The screener side shows the door that listed the name, and five small bars for the pillars Q M R V G — quality, momentum, revisions, value and expectations gap. A bar grows up when the stock is above its sector peers and down when below. A dashed empty bar means the pillar is missing, not zero.
Opening a name leads to its stock page.
The AI analyst's value and range
Where code can compute a value from the AI analyst's choices, the stock page shows a bear, base and bull value against today's price. A value far from the price is labelled, because the inputs may be wrong: read the analysis first. Some companies cannot be valued yet, and the reason is shown instead of a number. The price shown is the one the screener scored, and it is dated by the trading session it is for.
The stock page
The page for one name opens with the AI analyst's analysis: when it was read, its claims by topic with how many were confirmed in the filings, and its value where there is one. A verdict from the earlier analyst, if the name has one, stays under verdict history. The right column shows how the stock got on the list, what the price expects, a 400-day price chart, key financials and which paper books hold it.
Under ⋯ → On-demand is the request form and the list of the AI analyst's readings on request, each linking to its stock page.
Colour key — one meaning per colour
| Blue | The list: how far a stock has come through the screen. Also marks what you have selected, and links. |
|---|---|
| Violet | Quality: a high-quality business, and the scores behind it. |
| Teal | Value: cheap against the growth it has delivered, and the scores behind it. |
| Pink | Trend: a strong, steady price trend. |
| Green | Good: undervalued, a gain, a verdict that passes the gate. |
| Neutral ink | Fair: the price sits inside the value band. |
| Coral | Bad: overvalued, a loss. |
| Amber | A warning, look closer: a stale stamp, a notice, buying paused, half or quarter size. |
| Grey | Does not count: a blocked verdict, a vetoed stock, no data. |
Every coloured state also carries a word or a glyph, so colour is never the only signal.
The macro engine — the economic backdrop
A separate engine reads the state of the US economy and publishes its read on the Macro page. Its numbers come only from ; news is used as context for the story, never as a number.
What the Macro page shows
- The regime and its probability, as a title in the form “Regime, p % — strength label”. Underneath is a probability for each of five economic seasons — goldilocks, reflation, tightening, stagflation, recession (see ). The single label is only a summary and drives no number.
- The strength label. It comes from the engine's confidence: below 0.10 is “weak signal, context only”; 0.10 to 0.30 is “moderate signal”; above 0.30 is “clear signal”.
- Why — which of the seven dimensions (growth, inflation, policy, credit and liquidity, the yield curve, monetary liquidity, housing) supported or opposed the regime.
- Shocks — seven alarms for fear, credit, interest rates, oil, the dollar, jobs and inflation (), with the active ones marked, and any data-health warnings in full.
- Cost of capital — the market's implied equity risk premium (with where it sits in its own history) and the implied , with a flag if it is degraded. This is the discount rate the analyst anchors to.
What is switched on, and what is off
- Off: sector tilt. The macro read could give extra list places to sectors that suit the economy, but its sector picking failed a pre-registered test, so the read has not shown it can pick sectors. Every sector gets the same .
- Information only: the turbulence flag. It turns on when the fear index (VIX) is 30 or higher. It is shown for information; no real-money book or position size reads it.
- Information only: the regime as dated facts. The regime is shown as a dated fact, not as a forecast. One paper-only book, the Equal book at the risk dial's month decision, follows the dial; no real-money book does.
The quant screen — three doors
The screen does not blend everything into one score. A company can deserve attention by being excellent, by being mispriced, or by being in a strong, steady uptrend — and averaging those into one number describes none of them well. So there are three . A stock needs to clear only one. The screen's job is to decide what the analyst reads, so the analyst's time is never spent on a company that was never worth it.
Door 1 — Quality (compounder)
Companies that earn real money, keep growing and are rising in price. It blends and , with quality counting most.
Door 2 — Value (value gap)
Companies that look cheap against their own demonstrated growth: the price asks for less growth than the company has actually delivered (the ), and the stock is cheap on its . A falling-knife floor keeps out names in steep decline, because cheap and still falling is not a bargain yet.
Door 3 — Trend (trend leaders)
Strong, steady uptrends that the other two doors would miss — for example a sector-wide boom. Up to 20 extra places, with at most 5 per industry group. A candidate must be profitable and reasonably large, must not have falling analyst forecasts, and must have climbed steadily rather than in one lucky month.
How the doors compete
Each of the first two doors turns its score into a , and the better of the two is what a stock competes on. A — a stock in the top 10% on both doors — gets a small bonus of +2, because being excellent and cheap at once is rare. A falling knife cannot be a champion. On the desk the door that listed a name is shown as a coloured square and its name.
Staying power
A name already on the list stays until it falls clearly out. This is called : it stays in Research now while its rank is 60 or better, and on the list while its rank is 150 or better. Without that buffer, stocks sitting on the cut line would hop in and out with every small price move.
Fair comparisons
- Most grades are — a bank competes with banks, not with software companies — and no sector may take more than 18 % of the list.
- is the exception, in a useful way: it is also read across the whole market, so a boom that lifts an entire sector stays visible instead of looking ordinary.
- For cyclical businesses — oil, gas, mining, shipping and similar — whose cash flow swings with the cycle, the screen averages cash flow over several years (8 years for oil, gas and mining; 3 years for other cyclicals) instead of using one good or bad year.
Sector tilt is off
Every sector gets the same base of list places. The macro engine could give extra places to sectors that suit the economy, but its sector picking has not proved itself, so the tilt is switched off until it does.
Bands, vetoes & warnings
The result of the screen is four . The list is the first two together:
- RESEARCH NOW — the top of the list, roughly the first 50 to 60 names. Worth your research time today.
- WATCHLIST — the rest of the list, which holds about 150 names in all. This is the screen's own band name; it is not a verdict.
- PASS — scored, but not on the list.
- VETOED — removed by a safety filter. No band, and the reason is written beside it; on the desk it sits under Disqualified.
What removes a stock
A is a hard disqualifier, applied before the doors score anything. The reasons are: the stock cannot be traded, it is too small or too thinly traded, it has no usable filed fundamentals, it is a shell company, it has chronic operating losses together with heavy debt, or it shows two red flags at once. The last two apply only to companies worth under $10 billion.
What only warns
On large companies the forensic tests are warnings: (likely earnings manipulation), (profits not backed by cash), Altman Z (a distress score) and heavy from issuing shares. A warning never removes a stock worth $10 billion or more. On smaller companies a stock is removed only when two red flags agree. This keeps fast-growing leaders from being thrown out for looking unusual, while still showing you the warning so you can look closer.
Notes are not warnings
Data notes — for example that a stock's momentum was worked out from monthly prices, or that its latest annual report is old — describe how a number was built. They say nothing bad about the company and never remove a stock.
The AI analyst — today, and the earlier verdict analyst
A name on the list was valued in depth by the earlier AI analyst — like getting a careful second opinion. The work was divided so that each part did what it is good at:
- The AI does the research and chooses every valuation input, reading the company's own filings and cited web sources. It runs on a local model (), not a cloud service.
- Python does every calculation. An AI is a good analyst but an unreliable calculator, so no number is left to its arithmetic.
- Code checks the answer against outside anchors: the professional analysts' forecasts and their price-target range (the ).
The verdict is a direction against a band
Each stock is valued in independent runs, and each usable run ends in an estimate of what the business is worth. How many runs a verdict rests on is printed on every row (“n of m usable”) and is never assumed. The verdict is where today's price sits against the whole :
- Undervalued — the price is below the whole band, with at least a 10 % margin of safety: every run says it is worth more.
- Overvalued — the price is above the whole band, with at least a 10 % margin: every run says do not buy.
- FAIR — the price sits inside the band, or the margin is too thin. The uncertainty contains the price: no edge. That is a decision, not a refusal.
- — there is no complete valuation. That is a malfunction and is never shown as a verdict.
A single run is shown as a point, not a range, and its size is capped. A wider band means the runs disagreed, which lowers the size hint. The size hint is the most conservative of four buckets — run disagreement, margin of safety, the analyst's own conviction and the macro turbulence flag — and the stock page prints an upper bound that is a ceiling, not a suggestion. See .
Timing, and what would prove it wrong
Cheap is not the same as now. is a separate call — buy now, wait for momentum or avoid — with the condition that would flip it. And before any position exists, each verdict writes down what would prove it wrong, as checkable rules (see ).
The crux — what the price gets wrong
A buy or sell call has to name the one input that today's price gets wrong. That is the . It is what turns “the analyst says undervalued” into a claim you can check.
Two numbers for one input
Take one input of the valuation — growth in years 3–5, the cost of equity, how long growth fades, the return earned on new capital. is the value that input would need to have for the model to land exactly on today's price. The analyst's value is what it believes the input to be. The gap between the two is the disagreement with the market.
Quotes checked by code
The analyst has to back its number with quotes from the company's filings, and code checks every quote against the named source. A quote that cannot be found word for word does not count.
No valid crux means FAIR
A call that cannot name a valid crux gets no direction: the verdict is held at FAIR, and the stock page says “No valid stated disagreement → verdict held at FAIR”. The gate has its own reason for the same thing, directional_without_valid_crux.
On the stock page
“What the price gets wrong” shows each input on a number line with two markers — what the price implies, and the analyst's value — then the analyst's reason as a quote, and flags that say whether the stated disagreement checked out (“crux valid”) and whether the input was left at its default value. If no value of any other single input would explain the price, a line says so. The crux explains a verdict. It does not predict anything.
The gate — which verdicts count
After a verdict is written, code re-checks it every time the data are rebuilt. A verdict counts only if it passes . Every verdict is marked (yes or no), and when the answer is no, the desk gives the in plain words. A verdict that fails is still published and shown, with its reasons, but it can never be a recommendation and never enters the AI paper book.
Every reason the gate can give, with the wording the desk uses:
- not_usable
- no usable run
- pre_v3.1_gates
- made by the old analyst
- pre_valid_analyst
- made by the old analyst
- single_sample
- only one usable run
- fiduciary_fail
- failed the fiduciary audit
- kelly_on_overvalued
- sizing contradicts the verdict
- high_dispersion
- runs disagree too much
- non_production_row
- test run, not production
- mode_instability
- value jumped since the last verdict
- outside_street_fence
- outside the analysts' price-target range
- mos_beyond_150pct
- implausibly far above the price
- desk_not_used
- calculator not used
- low_effort_rescue
- rushed valuation
- sample_failed
- a run failed
- discovery_failed
- research step failed
- depleting_producer_unsupported
- mine / oil & gas producer (not yet supported)
- depleting_producer_class_unavailable
- producer type unknown
- consensus_data_unavailable
- analyst data could not be fetched
- no_street_fence
- no analyst price-target range
- directional_without_valid_crux
- buy/sell call without a verified stated disagreement
- nci_share_unmeasured
- minority-interest share could not be measured
- per_share_basis_unverified
- per-share basis failed the EPS check
- input_scope_stamps_missing
- input-scope proofs missing
A verdict can have more than one reason. The desk shows these reasons in English in every language, so the words match exactly. The Desk's funnel step “pass the gate” counts how many rows carry each reason.
The daily follow-up — monitoring never sells
Every verdict ends with a watch-list: the things that would matter if they happened. Once the follow-up is live, it checks each held name and each buy candidate once a day ():
- Gather. New SEC filings and company news not seen before, and whether the latest closes sit inside or outside the verdict's value band. No GPU is needed.
- The code floor. Exact signals that always queue a re-analysis, whatever any model says: an earnings release; a filing of the bankruptcy, restatement, delisting or change-of-control kind; a thesis rule that has just become true; a safety veto appearing on a held name; two closes on the far side of the value band ().
- The local model's read. It checks each watch item against the day's material and answers only “re-analyse now” or “nothing new”. It is never shown the verdict or any valuation figure, and every quote it cites must be found word for word in the named source, checked by code.
- Effects. A re-analysis jumps the queue (). New buying can be paused (). And every name is re-analysed at least every 14 days.
Monitoring never sells
The follow-up only asks for a fresh look. A held name is removed only by a new verdict. If a re-analysis of a held name fails, its last good verdict is carried for at most two failures or 21 days, and the desk labels it . Buy paused never means sell. If the follow-up data is older than 36 hours or missing, a name is treated as paused, never as buyable.
On the site this becomes the section of the desk, and “What we're watching” on each stock page: every watch item with its status (reported, not found, unclear), whether it would break or confirm the thesis, the quote, its source and date, and whether the quote was verified.
What is live, stopped and next
Where each part of the system stands on 12 November 2026:
- Live. The screener and its daily scoring; the AI analyst's analysis on the stock pages; and the Code book and the AI book, marked every day on paper from 6 November 2026.
- Stopped. The earlier verdict analyst and its gate. Its AI paper book holds only cash. Its record is kept on the Track page, labelled as the stopped analyst's.
- Being rebuilt. The daily follow-up of held names. It is not running.
- Off. Real-money mirroring is halted. The site never trades.
- Next. A new monthly book is formed each month; the next is due on 30 November 2026. Whether the AI book beats the Code book has not been measured, and it cannot be read until the books have several months of marks.
The earlier AI record
The earlier analyst's record was not reset, because that analyst did not go live again. It stays on the Track page, labelled as the stopped analyst's, and it allows no conclusion ().
Track record — the honest meter
Instead of a flattering , the system its monthly picks every day with real prices and assumed . If the system is wrong, this page will say so. Every number on it is paper (simulated) and covers a short period.
The earlier analyst's verdict scoreboard
It comes first, before any returns, and it is the stopped analyst's record, not the AI analyst's: nothing on this page grades the AI analyst. It shows three counts — verdicts graded from the valid analyst, verdicts graded from the old analyst (labelled legacy and inconclusive), and verdicts still pending () — then four horizon tiles (30, 91, 182 and 365 days) with how many are gradeable at each. The method is printed under it: entry is the first close on or after the verdict date, the benchmark uses the same dates, a horizon is graded only after it has fully elapsed, and repeat verdicts on one name are correlated. The data file's own caveats are shown word for word. While the count from the valid analyst is zero, no average-excess or percent-beat headline is shown.
The banner
A banner at the top says that every book on the page is paper, that the Code book and the AI book began on 6 November 2026 and have run far too short a time to show anything, that the AI analyst has not been graded on returns, and that the scoreboard and the old AI book belong to the stopped analyst.
The paper books
- AI book — the Code book's picks, with a slot held as cash when the AI analyst's reading of the filings finds a material event. Formed monthly, equal weight per slot.
- Code book — the screener's monthly picks alone, the same names before the AI analyst's check.
- Old AI book (stopped) — followed the earlier analyst's verdicts that passed the gate (). It holds only cash.
- Code baseline (engine v2.1) — a comparison book from October, paper only. Its picks stopped updating in mid-October. The Claude book (paper) holds the BUY names of the Claude valuation run (an outside model reads pinned filings three times; code does the arithmetic) at equal weight, paper only, and is not yet graded on returns.
- Control book — every Research now name, with no AI (), since June 2026.
- My book — your own saved holdings, tracked like a fund () so that adding money never fakes performance. Signed out, it asks you to log in.
Each book shows its return, how it did against QQQ, its maximum drawdown, win rate, open positions and average hold. Under the return and the benchmark comparison is the number of behind it. and are hidden while there are too few observations — the page says “not meaningful at n observations” — because annualised ratios on a few weeks of data are noise.
Growth of 100, and the ledger
Growth of 100 plots the books and the benchmarks (, , and a few more) from the same start, with a window selector and a what-if cost box that re-costs every trade. The ledger lists the AI book's trades with a plain reason (for example “entered at the monthly formation”) and shows days with no trade as “no changes”.
Honest rules
- A trade happens at the first close after the signal, because a signal can only be acted on after it exists.
- Costs are included on every trade.
- Benchmarks use the same dates as the trades.
How to read it
- Control vs QQQ — the stock selection works only if the picks beat the , QQQ.
- AI book vs Code book — the AI analyst's filing check earns its keep only if its book beats the Code book. This has not been measured: the books began on 6 November 2026.
- My book vs the others — your own deviations show up as the .
My portfolio — checking your own holdings
The Portfolio page is for your actual holdings. Add a ticker and a dollar value, or paste a broker export. They are saved only in this browser unless you log in. Each holding is checked against the machine, and the result is shown next to it as the machine's stance.
The stance words, and what they mean:
- NO COVERAGE — the stock is not in the screened universe.
- ✕ VETOED · reason — a safety filter removed it.
- NO ANALYSIS YET — the earlier analyst never valued it.
- BLOCKED · reason — a verdict exists but fails the gate. It is never a signal.
- REDUCE (overvalued), FAIR and BUY · size — the verdict passes the gate and says overvalued, fair or undervalued.
- NO PLAUSIBLE RUN — no run was usable.
Stances come from the earlier analyst's verdicts, and that analyst is stopped; the AI analyst's reading is on each stock page and is not part of the stance. BUY and REDUCE are the machine's labels on a research list, never advice to you. Where they exist, the thesis and follow-up status sit beside each holding. The footer gives your total, your cash share, a warning if one sector is over the 25 % rule, and what the old AI book holds today. Your holdings also feed My book on Track record.
The archive — retired pages kept for reference
Under ⋯ → Archive are pages from older versions of the site, each marked retired: the legacy lenses (the 100-bagger and Reverse screens), the AI archive and the YouTube strategy. They are kept for reference. They are not the current system, and nothing on the desk is built from them. The desk itself has one view; the old Quant and Compare switches are gone.
How the system is validated
Four habits keep the system honest:
- Rules before results. Pass/fail rules are written down before the results are seen, so they cannot be tuned afterwards to look good (that would be ).
- Grading verdicts. A grader checks each verdict against what the price did over 30, 91, 182 and 365 days, compared with , and QQQ. A horizon is graded only once it has fully elapsed, the benchmarks use the same dates as the verdict, and repeat verdicts on one name are treated as correlated, not independent.
- A control. The sharpest question is whether the AI analyst's reading adds anything on top of the screen. The Code book and the AI book exist to answer it: if the AI book cannot beat the Code book, the AI analyst's filing check adds nothing to the picks. That comparison has not been measured.
- Paper books. The track record books trade daily with real prices and assumed costs.
Negative results are published too. The macro sector tilt failed its test and was switched off, and the macro page says so.
Where the data comes from, and system health
- — 10 years of as-filed fundamentals via (the ground truth for quality, forensic warnings and demonstrated growth), kept .
- — prices, analyst , and coverage.
- — US economic series behind the macro engine.
- News — narrative context only. It is never turned into a number.
Two clocks
The scoring chain re-runs in the cloud several times each weekday, with a refresh after the close. Cloud schedulers can run late, so every file carries its own timestamp. The AI analyst runs separately and publishes its analysis on its own clock.
System health
The strip under the masthead on every page answers “is anything stale or broken?”:
- Three stamps — prices as of (the trading session the prices are for), book scored and AI analysis (its newest publish). A stamp turns amber with a ● when it is past its expected age: prices when the post-close refresh is more than 3 days old, the book after 36 hours and the AI analysis after 14 days.
- The health dot — green when all is well. Amber for a stale stamp, a discount rate that did not come from the macro anchor, or a degraded macro read. Red when the chain's own integrity check failed or an error alert was raised in the last 7 days.
- The macro chip on the right — the regime, its probability, the strength label and the cost of equity, linking to Macro. Macro stamps older than 45 days turn amber.
Click the dot to open the health drawer. It is read-only: Freshness (each stamp with its expected cadence), Analyst (the state, the baseline progress, and that the real-money mirror is halted), Alerts (from the paper ledgers) and Checks (the chain's integrity checks, the discount-rate source, macro warnings and the follow-up state). A file that failed to load is shown as not loaded, not as a failure. There are no controls in it.
Frequently asked questions
Is this financial advice?
No. It is a research system with the evidence laid out. Nothing here buys or sells anything, and nothing here is financial advice.
Does the site execute trades?
Never. The site places no trades and has no order button. The paper books are simulated — with real prices and assumed transaction costs. A separate real-money mirror exists in the system’s code; it is halted, and this site does not control it.
What is under Research now?
The screener's top names by rank, roughly the first 50 to 60, with the AI analyst's reading beside each. It is a list of names worth your research time, not a buy list. The earlier verdict analyst, whose verdicts used to place names here, is stopped. See What is live, stopped and next.
Why are verdicts marked blocked?
This is about the earlier verdict analyst, which is stopped. One of its verdicts counted only if it passed the gate. Each blocked verdict carries its reasons in plain words, for example “only one usable run” or “runs disagree too much”. Blocked does not mean wrong; it means the system will not stand behind it. A blocked verdict stays visible as a record. See The gate.
What does “No edge today” mean? Is it a sell?
This is about the earlier verdict analyst, which is stopped. No. It means the verdict passed the gate and found no gap: the price sits inside the , or the margin is too thin to call. FAIR is a real answer, not a refusal and not a sell signal. An overvalued verdict also sits there, and it means “do not buy”, not “sell”. See .
Why are some sectors not tilted?
The macro engine could give extra list places to sectors that suit the economy, but its sector picking failed a pre-registered test. So the tilt is off, and every sector gets the same , until the engine earns the right.
Why does a stock stay on the list after its rank slips?
That is . A name already on the list stays until it falls clearly out — rank 60 for the Research now band, rank 150 for the list as a whole — so stocks on the cut line do not flicker in and out.
What does “vetoed” mean?
A : a safety filter removed the stock, whatever else looks good about it. On the desk it sits under Disqualified with the reason written beside it, and the analyst cannot override it.
What if the screen and the analyst disagree?
They are two separate readings. One is pure math over financial statements and prices; the other is the AI analyst, a cloud AI model that reads the filings. The analysis it writes is on each stock page and does not place names on the list. When the two strongly disagree, one of them may be wrong — those are the interesting names. A safety veto is never overridden.
Where are the old analyst’s verdicts?
The old analyst was ruled invalid. Its verdicts are not shown on the desk; one quiet line counts them. Each one stays on its own stock page under verdict history, labelled as old, so nothing is erased.
How can I trust the track record?
You should not treat it as proof yet. It is built forward, day by day (), with included and benchmarks measured on the same dates, but it is a short paper record. The Code and AI books began on 6 November 2026, so there is almost nothing to read yet. Read it as an early scoreboard; see .
The glossary — every term in this handbook
A searchable index of all 117 defined terms. Click any term chip to open its definition — or click a term inline in any section above.
Disclaimer
StockPeak publishes the output of a research system. It is not investment advice, not a recommendation to buy or sell any security, and not personalised to your circumstances. The AI analyst's analyses and values, like the earlier analyst's verdicts, are model outputs built on assumptions; they can be wrong. Track records shown are paper (simulated) portfolios, include assumed costs, cover a short period, and have not been shown to predict future returns. Historical macro data are revised data, not point-in-time. The site is not a trading bot and not a crystal ball: the screen works on average over years, not on every stock every month, and the AI analyst's analysis has not been graded on returns. Do your own research; you can lose money.