joris

Glossary

Plain-language meanings of the terms used across the analyses. This page shows the file framework/04-glossary.md, where the terms are defined. Each analysis explains only the words specific to its own drug and disease.

04 — Glossary (generic)

This glossary explains, in plain language, the general terms used across the analyses and the framework files. It has three sections: clinical and regulatory terms, financial and market terms, and the framework’s own process words. Each analysis keeps its own short glossary for the terms specific to it: its drug, its trials and its disease vocabulary. For every other term it links here (01-rules.md, rule 22). Add a new generic term here, not in an analysis.

Other framework files are named in backticks, such as 02-connectors.md. The website has a page for this glossary and for each analysis, but none for the other framework files, so the glossary never links to them.

Clinical and regulatory

TermMeaning
IndicationThe disease or condition a drug is being developed to treat. One drug can be developed for several indications. In this repo a program is one drug in one indication, so the indication is part of what identifies a program. Example: “foralumab in na-SPMS” names the drug and then the indication.
CatalystA scheduled event that can move a company’s share price. Examples: a trial announcing its results (a readout), the FDA deciding whether to approve a drug (a PDUFA date), or an advisory committee meeting. BPIQ flags a program that has one with has_catalyst: true, and only such a program gets its own analysis document. A prediction is identified by the ticker, the program and the catalyst together, because one program can have more than one event.
ReadoutThe moment a trial’s result is made public. It is the catalyst most predictions in this repo are about. The first announcement is usually the topline (next row). The benchmark study in 07-benchmarks.md dates a readout to the earliest public report that addresses the trial’s primary endpoints.
ToplineThe first high-level headline results of a trial. The full data come later, often at a medical congress or in a paper.
Mechanism of action (MoA)How a drug works in the body.
TargetThe molecule in the body, usually a protein or a gene, that a drug acts on.
Target validationEvidence that the target matters in the disease, independent of the drug itself. Human genetics is the strongest kind: people who carry a variant of the target’s gene get the disease more often, or less often. The Open Targets connector supplies this evidence in the program-tier sweep (02-connectors.md), and the program document records it in its evidence tables.
Line of therapyWhich round of treatment a patient is on. First line is the first treatment. Second line is the next one, after the first has failed.
Standard of careThe treatment patients normally receive today for that disease and stage.
SponsorThe company or institution that runs and pays for a trial. It is usually the company that owns the drug. The program document says whose drug each trial tests (01-rules.md, rule 18).
InvestigatorA doctor or researcher who runs a trial at a hospital or clinic. The principal investigator leads it; sub-investigators work under them. An investigator is paid by the trial, so they have a relationship with its sponsor by definition. The program document lists each one with a name, a date, a source and a conflict disclosure (01-rules.md, rule 40). The sources are CT.gov search_investigators and get_trial_details (02-connectors.md § KOL sources).
Key opinion leader (KOL)A senior clinician whose views shape practice in a disease area. A program document records them in two separate groups, because the two groups are not the same kind of witness. Investigators are paid by the trial, so they have a relationship with its sponsor by definition. Independent voices have no such relationship. Each person is recorded with a name, a date, a source and their own conflict disclosure. A view that lacks any of the four is not recorded at all (01-rules.md, rule 40). See Independent voice and Conflicts checked under Framework and process, below.
Phase 1, Phase 2, Phase 3The three stages of testing a drug in people. Phase 1 is a small trial, usually the first in humans, mainly about safety and dose. Phase 2 is a larger trial that looks for the first signs that the drug works and settles the dose. Phase 3 is a large trial designed to prove the effect well enough to support approval. Some sources write the numbers as I, II and III. The benchmarks in 07-benchmarks.md are split by phase, because shares react differently at each: on average a positive Phase 2 readout lifts a small drug developer’s shares by 12% and a miss cuts them by 16%.
Pivotal trialThe trial whose result an approval decision rests on. It is usually a Phase 3 trial, and sometimes a Phase 2 trial in a rare disease. The program-tier sweep is run on this trial’s NCT number, and the design benchmark is read for this trial’s design (07-benchmarks.md § Trial design). In the program document it is the A.5 (A.2 before 6.2.0) trials-table row the readout settles on.
Single-arm trialA trial with no comparison group. Every patient gets the drug. Results are judged against historical results in similar patients.
Randomized controlled trial (RCT)A trial where patients are assigned by chance to the drug or to a comparator (a placebo or an existing treatment).
Placebo, placebo-controlled, placebo-adjustedA placebo is a dummy treatment with no active drug in it. A placebo-controlled trial assigns patients by chance to the drug or to the placebo, so the drug’s effect can be measured as the difference between the two groups. A placebo-adjusted figure is that difference. Example: 40% of patients respond on the drug and 15% on placebo, so the placebo-adjusted response rate is 25 percentage points.
Head-to-head trialA trial that assigns patients by chance to the new drug or to an existing active treatment, rather than to a placebo. It produces the strongest evidence, and the largest share-price moves. A negative Phase 3 head-to-head readout averages −73%, against −22% for Phase 3 trials of all designs (07-benchmarks.md § Trial design).
Open-label / blindedOpen-label means everyone knows who gets what. Blinded means they do not.
Interim analysisA planned look at a trial’s data before the trial is finished. Depending on what it shows, the trial can be stopped early for benefit (the drug has clearly worked), for futility (next row) or for safety. The program document’s risk table treats an adaptive design with an interim analysis as a middling design, between a pivotal Phase 3 with an SPA and a single-arm trial with no control.
Futility stop (early stop)Stopping a trial early because an interim analysis shows the drug cannot reach its goal. Early stops are not symmetric in the market’s eyes. A trial stopped early for benefit is not rewarded above an ordinary positive result. A trial stopped for futility or safety falls slightly less than an ordinary miss: −14% against −17% (07-benchmarks.md § Surprise).
First / last patient in (FPI / LPI)When the first and last patients were enrolled.
Last patient out (LPO)When the last patient finishes follow-up.
Database lock (DBL)The point where trial data are finalised before analysis.
Primary completion dateThe date the last patient in a trial is examined for its primary endpoint. It is a field on the trial’s ClinicalTrials.gov record, primary_completion_date. It is known to the month, and it moves when the trial moves. This repo reads it as the ctgov source of a readout window, and cross-checks every catalyst date against it (01-rules.md, rules 24 and 32). The modelled readout estimate starts from it: primary completion plus two to four months, by default (rule 34).
ClinicalTrials.gov (CT.gov) and NCT numberClinicalTrials.gov is the public registry of clinical trials run by the US National Institutes of Health. Every registered trial has an NCT number, such as NCT05109975, which is its permanent id. An analysis links the registry record wherever it cites a trial (01-rules.md, rule 20), and the CT.gov connector is called on the program’s pivotal trial in every program-tier sweep (02-connectors.md).
Primary / secondary endpointThe primary endpoint is the main measure a trial is designed to prove. Secondary endpoints are the supporting measures. A trial succeeds when it meets its primary endpoint. Every endpoint is spelled out in an analysis: full name, what it measures, its scale, which direction is better, and the minimal clinically important difference if one exists (01-rules.md, rule 16).
Surrogate endpointA measure that stands in for the outcome patients care about, because it can be read sooner. Examples: tumour shrinkage instead of survival; a blood marker instead of symptoms. A surrogate is validated when it has been shown to predict the real outcome. Accelerated approval is granted on a surrogate. An unvalidated surrogate is a risk the program document scores.
Counting window (of an endpoint)The period over which an endpoint’s events are counted, and the visit at which the count is taken. The two can differ. A trial may count events “within 28 days” at a Day 28 visit that the protocol allows from Day 25 to Day 31, so the company can report two counts. A positive definition locked from framework 6.3.0 names which count settles (05-prediction-protocol.md § Pre-registered settlement criteria).
p-valueThe chance of seeing a difference at least as large as the one observed between drug and comparator, if the drug did nothing. The smaller it is, the less likely the difference is luck. A result is called statistically significant when its p-value is below a threshold set in advance, usually 0.05.
One-sided / two-sided testTwo ways of reading a p-value. A two-sided test asks whether the drug differs from the comparator in either direction. A one-sided test asks only whether it is better. For a symmetric test, the two-sided p-value is double the one-sided one, so the side can decide whether a result clears 0.05. A p-value published with no stated side is read as two-sided, which is the convention. A positive definition locked from framework 6.3.0 names the side it requires (05-prediction-protocol.md § Pre-registered settlement criteria).
Successor product (next-generation version)A newer version of the same drug from the same company, meant to replace the one in the trial: the same active ingredient in a new formulation or delivery system, or a closely related molecule. Example: ARCT-2601 is ARCT-810’s messenger RNA in a new lipid particle. A company can move its plan to the successor without saying the original failed. A positive definition locked from framework 6.3.0 says how such a switch scores (05-prediction-protocol.md § Pre-registered settlement criteria).
Objective response rate (ORR)The share of patients whose tumours shrink by a set amount, measured by a standard rulebook (RECIST v1.1 in solid tumours).
Duration of response (DOR)How long a response lasts.
Progression-free survival (PFS)Time until the disease worsens or the patient dies.
Overall survival (OS)Time until death from any cause.
Clinical benefit rate (CBR)Share of patients with a response or prolonged stable disease.
Minimal clinically important difference (MCID)The smallest change patients actually notice or benefit from.
BiomarkerA measurable feature (often a protein or gene) used to select patients or track disease.
Companion diagnostic (CDx)A test that identifies which patients are likely to respond to a drug.
Antibody-drug conjugate (ADC)An antibody that carries a chemotherapy payload directly to tumour cells.
MyelosuppressionReduced production of blood cells by the bone marrow, causing low white cells, red cells or platelets.
SepsisA life-threatening reaction to infection. A risk when white-cell counts are low.
Synthetic lethalityTwo conditions each survivable alone but fatal in combination.
Target product profile (TPP)A written statement of what the drug needs to achieve to be worth developing: the efficacy, safety, dosing and other properties it is aiming for, each with its supporting evidence. The program document records it in section A.1 (A.3b and A.4 before 6.2.0). The expectation benchmark for the stock call is often read from it.
Label expansionApproval of an already-approved drug for more uses: another disease, another stage of the same disease, or an earlier line of therapy. The label is the FDA-approved description of what a drug may be prescribed for. The chance of later expansions is one reason oncology shares fall less on a miss than shares in other therapy areas (07-benchmarks.md § Therapy area).
Food and Drug Administration (FDA)The US government agency that decides whether a drug may be sold in the United States. It reviews the approval application, sets the decision date and grants the designations listed below. Most catalysts in this repo are either a trial readout or an FDA decision.
NDA / BLAThe application a company files with the FDA to ask for approval. An NDA (New Drug Application) is for a chemical drug. A BLA (Biologics License Application) is for a biological drug, such as an antibody. The FDA’s target date to decide on it is the PDUFA date. On the published transition rates, 90.6% of applications filed are approved (07-benchmarks.md § Probability of approval).
Probability of approvalThe chance that a drug will eventually be approved. This repo uses one specific form of it: the probability of approval given a positive readout, recorded as poa_if_positive. It is the product of the transition rates still ahead of the drug: 15.1% after a positive Phase 1 readout, 52.4% after a positive Phase 2, 90.6% after a positive Phase 3 (07-benchmarks.md § Probability of approval). It feeds the commercial value ceiling (01-rules.md, rule 43). It is a conservative estimate and never a scenario anchor. See Transition rate.
Transition rateThe share of drug programs that moved from one phase to the next, out of all the programs that either moved on or were suspended. Example: if 100 programs finish Phase 2 and 29 go on to Phase 3, the Phase 2 to Phase 3 transition rate is 29%. A program dropped for a business reason counts as a failure, so the rate sits below the odds of a program whose data are good. The published rates come from the BIO/Informa/QLS 2021 report (07-benchmarks.md § Probability of approval, and by area and drug type in § Probability of success — base rates).
Accelerated approval (AA)An FDA route that allows earlier approval on a surrogate endpoint. A confirmatory trial is required afterwards.
Fast TrackAn FDA designation that speeds development and review for a serious disease with unmet need.
Breakthrough TherapyA stronger FDA designation for drugs showing substantial early improvement over available therapy.
Orphan DrugAn FDA designation for rare-disease drugs, carrying incentives and extra exclusivity.
PDUFA dateThe FDA’s target decision date for a drug application. The name comes from the Prescription Drug User Fee Act, the law that sets the deadline.
Advisory Committee (AdCom)An external expert panel that advises the FDA. Its vote is not binding.
Special Protocol Assessment (SPA)A written FDA agreement that a trial’s design is adequate to support approval.
Clinical holdAn FDA order pausing a trial, in full or in part, usually over safety.
Health technology assessment (HTA)A national body’s review of whether a drug is worth funding (for example NICE in the UK, CADTH in Canada).
DMPKDrug metabolism and pharmacokinetics: how the body absorbs, distributes and clears a drug.
CMCChemistry, manufacturing and controls: how the drug is made and quality-checked.
ModalityThe kind of thing a drug is made of: small molecule (a chemical drug, usually a pill), antibody, bispecific antibody, antibody-drug conjugate (an antibody that carries a toxin to a target cell), cell therapy, gene therapy, RNA drug, radioligand (a drug that carries radiation to its target), peptide, vaccine, or other. It decides how the drug is made, how it is tested and how often it succeeds. The program document records it in A.7 and the manufacturing section M. The allowed values are fixed by the schema.
PICOPopulation, intervention, comparator, outcomes: the four parts of a question a health technology assessment asks of a trial. Example: adults with moderate psoriasis (population), the new drug (intervention), the current standard drug (comparator), and clear skin at week 16 (outcome). The EU joint clinical assessment sets the PICOs a company must answer.
Quality of life (HRQoL) and EQ-5DHRQoL is health-related quality of life: how a disease and its treatment affect daily life, as the patient reports it. EQ-5D is the most common questionnaire for it. It scores five areas (mobility, self-care, usual activities, pain, anxiety) and turns the answers into one number that payers use to work out QALYs.
Joint clinical assessment (JCA)The European Union’s single review of how well a new drug works, run once for all member states under the EU health technology assessment regulation. It covers cancer drugs and advanced-therapy medicinal products from January 2025, orphan drugs from January 2028 and all new medicines from 2030 (07-benchmarks.md § Rules of thumb, [UNVERIFIED — illustrative]). Each country still decides price and reimbursement on its own.
G-BA, IQWiG and AMNOGGerman drug pricing. AMNOG is the 2011 law that makes a new drug’s price depend on its added benefit over the standard treatment. IQWiG is the institute that writes the assessment. G-BA, the Federal Joint Committee, decides the added-benefit category. A drug with no added benefit generally gets the comparator’s price.
NICEThe National Institute for Health and Care Excellence: the body that decides whether the English health service pays for a drug. It weighs cost per QALY against a threshold (see QALY and ICER, under Financial and market).
ChuikyoJapan’s Central Social Insurance Medical Council. It sets the national reimbursement price of a drug in Japan.
Complete response letter (CRL)The FDA’s letter saying it will not approve an application in its present form. It lists what must be fixed. A manufacturing problem is a common reason. The FDA publishes many of them.
Form 483 and warning letterRecords of an FDA inspection. A Form 483 lists what the inspector observed that may break the rules. A warning letter is the FDA’s formal next step when the problems are serious or not fixed. Either at a drug maker’s plant can delay an approval.
REMSRisk evaluation and mitigation strategy: safety steps the FDA requires of a drug beyond the label, such as restricted prescribers or special monitoring. A REMS adds cost and slows the launch.
Orange Book and Purple BookFDA lists. The Orange Book lists approved chemical drugs, with their patents and exclusivity periods. The Purple Book lists approved biological drugs and their biosimilars (copies of a biological drug). Both are public and are the first place to check the protection around a drug.
Contract development and manufacturing organisation (CDMO)An outside company that develops and makes a drug for the sponsor. Most small drug developers have no plant of their own, so the CDMO’s record and spare capacity are part of the risk (section M.4 of the program document).
CQA, PPQ, HPAPI, DAR, ATMPManufacturing terms. A critical quality attribute (CQA) is a property of the drug that must stay in range for it to be safe and effective. Process performance qualification (PPQ) is the set of full-scale batches that proves the process works. HPAPI is a highly potent active ingredient, which needs special containment. DAR is the drug-to-antibody ratio of an antibody-drug conjugate: how many toxin molecules each antibody carries. ATMP is an advanced-therapy medicinal product, the EU term for a gene, cell or tissue therapy.
QTcA measure on an electrocardiogram of how long the heart takes to reset between beats. A drug that lengthens it can cause a dangerous rhythm. Regulators expect most new drugs to be checked for it.
Drug–drug interactionOne drug changing how another is absorbed or cleared, so that the level of either in the body rises or falls. Example: a drug that blocks the enzyme which clears a second drug makes the second drug’s level climb.
Exposure–response and population PKExposure–response is the link between how much drug is in the body and the effect or side effect. Population PK is the study of how that drug level varies across patients (by weight, age, kidney function and so on), using data pooled from many people. Together they support the dose chosen.
Organ impairment studyA small study of how a drug behaves in patients whose kidneys or liver work poorly. The label’s dosing advice for those patients rests on it.
Statistical power and assurancePower is the chance that a trial finds an effect of a stated size if the effect is really there. A trial with 80% power misses a real effect one time in five. Assurance is the chance that the trial succeeds once the uncertainty about the true effect is allowed for: it averages power over the effects that might be true. Assurance is the lower and more honest figure when the early evidence is thin.

Financial and market

TermMeaning
Share, stock, tickerA share is one unit of ownership in a company. A company’s stock is all of its shares together. The ticker is the short code a stock exchange uses for the company, such as TLSA. In this repo the ticker identifies the company: its files live under analyses/<TICKER>/, and a batch dispatches one subagent per ticker.
Market capitalisation (market cap)Share price × number of shares: the market’s price tag on the whole company. BPIQ’s market_cap field is unreliable, so the company document recomputes it from the filed share count and the price actually wanted (02-connectors.md § Known data bugs).
Emerging biopharma (EBP)IQVIA’s term for a smaller drug developer. The benchmark study in 07-benchmarks.md covers 1,502 trials sponsored by publicly listed EBPs, 70% of them under $1B market capitalisation. That is the same group of companies this repo screens.
Pre-commercialA company with no product on sale yet. Its value rests entirely on drugs still in development, so its shares react more strongly to trial results. Among pre-commercial EBPs, companies under $1B market capitalisation react 13.3 times more strongly than larger ones at Phase 1, 9.5 times at Phase 2 and 6.2 times at Phase 3 (07-benchmarks.md § Company size).
Spot (current share price)The current share price: what one share trades at right now. Spot is the price everything else is measured against. In this repo spot is the price cache’s most recent close, recorded with the day it was read. It is stored on the prediction at lock time rather than looked up later.
Price cacheThe daily price history kept in the repository itself, one file per ticker at data/prices/<TICKER>.json. Prices are not looked up fresh each time they are needed. Every figure derived from price is read from the cache: the 52-week range, a run-up entry or exit price, a realised peak. So the same figure recomputes to the same number from the same commit, however long afterwards, and a past date can be priced as honestly as today’s. scripts/fetch-prices.mjs is its only writer and lib/prices.mjs its only reader. The cache is never read into a session’s context. The orchestrator refreshes it before any run that reads a price (02-connectors.md § Pre-run price fetch).
Price barOne trading day’s prices in the price cache: at minimum the date and the closing price, and usually also the open, high, low and volume. The cache holds five years of daily bars per ticker. A halted or untraded day keeps its bar with null values; nothing is filled in. Settlement counts bars to measure trading days.
Trading dayA day the stock exchange is open. Weekends and market holidays do not count. Run-up exit rules count trading days (see Exit rule). A date set by a rule can fall on a non-trading day; the price used is then the last close on or before that date.
Intraday moveThe change in the share price within one trading day. The company document’s table of past catalyst reactions (C.7) records intraday moves. The benchmarks in 07-benchmarks.md and a settled actual_move_pct use a different window: from the average of the two closes before the event to the average of the two closes around it. Compare the conventions before comparing the numbers.
52-week positionWhere the last price sits between the 52-week low and the 52-week high, as a fraction: 0 at the low, 1 at the high. It is derived from the price cache for whatever date is being asked about. It is not captured once at the moment of writing (03a-company-spec.md C.4). It is deliberately not clamped: a price outside its own 52-week range shows as below 0 or above 1 instead of being pulled quietly back to the edge.
Enterprise value (EV)Market capitalisation minus cash. It is closer to the cost of the business itself, because a buyer would also get the cash. In this glossary EV always means enterprise value and never expected value; expected value is written out in full. BPIQ’s enterprise_value field cannot always be reconciled with its own cash field, so the company document recomputes it (02-connectors.md § Known data bugs).
Cash runway (runway)How long the company’s cash lasts at its current spending rate, stated in months or as an end date. The company document records the runway as the company states it, and a second figure recomputed as cash ÷ burn (03a-company-spec.md C.3). A runway that ends before the catalyst means a raise is likely first, which changes the trade. It is one of the two inputs to the run-up’s financing and clustering risk driver.
BurnHow much cash the company spends per month or quarter.
Cash per shareThe company’s cash divided by its economic share count (below), not by the reported share count. The cash floor (below) is the form of it that a scenario price range uses as an anchor (01-rules.md, rule 30). It often sets a floor under the miss scenario: when cash alone is worth close to the share price, the shares cannot fall far.
DilutionRaising money by issuing new shares. Each existing holder then owns a smaller share of the company. A dilution mechanism that fires on the event, such as warrants, an at-the-market facility or a convertible trigger, is an admissible anchor for a scenario price range, counted on the economic share count (01-rules.md, rule 30).
Private placement (PIPE)Selling new shares directly to selected investors to raise cash.
WarrantThe right, but not the obligation, to buy a share at a set price (the exercise price) until an expiry date. Companies often issue warrants alongside shares in a financing. Warrants that are exercised add new shares, so they are a dilution mechanism.
Pre-funded warrantA warrant whose price has been paid almost in full up front, leaving a token exercise price such as $0.001. It is economically the same as a share, so it is counted in the economic share count (03a-company-spec.md C.4).
In-the-moneyA warrant or option is in the money when its exercise price is below the current share price, so exercising it is worth doing. In-the-money warrants and options are counted in the economic share count, because they are likely to become shares. Out-of-the-money ones are not.
Convertible note (and its conversion trigger)A loan to the company that can be turned into shares instead of being repaid in cash. The terms say when and at what price the conversion happens. A conversion trigger is the condition that sets it off, such as a share price being reached or a date passing. A trigger that fires on the event is a dilution mechanism and an admissible anchor (01-rules.md, rule 30).
At-the-market (ATM) facilityAn arrangement that lets a company sell new shares gradually into the open market at the going price, through a bank, whenever it chooses. It is a standing source of dilution. The company’s remaining ATM capacity is read from its SEC filings (02-connectors.md), and an ATM that fires on the event is an admissible anchor (01-rules.md, rule 30).
Shelf registration (S-3, 424B5)A Form S-3 registers a block of securities with the SEC in advance, so the company can sell them later at short notice (they sit “on the shelf”). A 424B5 is the prospectus supplement filed when some of them are actually sold. Together they show how much the company can raise before a catalyst. Both are read from the EDGAR filing feed (02-connectors.md).
SEC filings (Form 3, Form 4, 8-K, 6-K)Documents a listed company must file with the US Securities and Exchange Commission (SEC). The ones this repo reads: Form 3 and Form 4 report an insider’s holdings and trades; an 8-K reports a material event, such as a trial result or a financing, within days; a 6-K is the equivalent of an 8-K for a foreign company listed in the US. EDGAR is the SEC’s public database of filings, and the EDGAR submissions connector reads it (02-connectors.md). See also Shelf registration and 13F.
13FA quarterly filing in which large investment funds disclose their holdings. A 13F is filed up to 45 days after the quarter it reports (02-connectors.md § Known data bugs).
Insider, insider flowAn insider is an officer, a director or a holder of more than 10% of the company’s shares. Insider flow is their buying and selling, reported on Form 4. The company document separates open-market buying, which shows conviction, from routine share grants and tax sales (03a-company-spec.md C.5). BPIQ’s insider connector can return no rows while Form 4 filings exist, so the EDGAR filing feed is the cross-check (02-connectors.md § Known data bugs).
FloatThe shares available to trade publicly, leaving out those held by insiders and other affiliates. A small float means a thin market, in which a surprise moves the price further. It is one of the three inputs to the run-up’s squeeze mechanics driver (lib/runup.mjs). Short interest is usually quoted as a percentage of the float.
Short interestThe share of a company’s stock that investors have bet against, by borrowing shares and selling them in the hope of buying them back cheaper. High short interest signals scepticism, and can fuel sharp rallies on good news. BPIQ’s short_float units are unconfirmed. The FINRA short count divided by the EDGAR-filed share count gives a percentage whose units are known (02-connectors.md).
Days to coverShort interest divided by the average daily trading volume: how many days of normal trading it would take for every short seller to buy back. A high figure means a rally would be hard to escape. FINRA’s semi-monthly short-interest data supplies it (02-connectors.md).
Short squeezeA sharp rise in a heavily shorted stock. Good news forces short sellers to buy back shares to limit their losses, and that buying pushes the price higher still. A high short interest can turn good news into a squeeze. That is why the triage screen reads it and the run-up’s squeeze mechanics driver scores it.
Reverse splitMerging several existing shares into one, which multiplies the price per share by the same factor without changing the company’s value. Example: a 1-for-10 reverse split turns ten $1 shares into one $10 share. A reverse split between a prediction’s lock and its event would make a settlement read as a loss that never happened, so the settlement carries a spot_scale factor and a note saying where it came from (05-prediction-protocol.md § Locking and settlement).
Options, options chainAn option is a contract to buy (a call) or sell (a put) a stock at a set price (the strike) by a set date (the expiry). When traders expect a big move, options get more expensive. The options chain is the full list of options traded on a stock, by strike and expiry. The company document reads the chain for the expiry nearest the next catalyst (03a-company-spec.md C.6), and says plainly when the chain is too thin to read.
StraddleA matched call and put at the same strike. Its combined price shows the size of move the market expects.
Implied moveHow large a price swing the options market is pricing around an event. Unreliable when the options barely trade.
Implied volatilityThe volatility level implied by options prices. A floor or ceiling value usually means missing data, not real volatility.
Open interestHow many option contracts are outstanding. High open interest with no volume still means the price is unreliable.
Sell-side analystAn analyst at a bank or brokerage who publishes research and price targets on stocks for the bank’s clients. Their targets on the small companies this repo analyses sit far above spot as a rule, so a target is a directional bias, not evidence (01-rules.md, rule 30). A claim about what analysts think needs a name, a date, a figure and an evidence tag, like any other claim (rule 40).
ConsensusThe view most investors and analysts share about an event or a figure: the result they expect, or the average of their forecasts. The expectation benchmark can be read from published consensus. A gap between our probability and the market’s needs evidence of something consensus has not absorbed (01-rules.md, rule 42).
Price targetAn analyst’s estimate of where a stock should trade in about twelve months. Wide disagreement between analysts is itself information. It is never an anchor for a scenario price range, because targets on small companies sit far above spot as a rule (01-rules.md, rule 30).
Priced inAlready reflected in the share price. When the market expects a result, the price moves before the event, and the event itself then moves it little. Example: a positive Phase 3 readout lifts shares by 11% on average, less than a positive Phase 2, because the market treats it as confirmation of something it had already partly priced in (07-benchmarks.md § Phase). See Priced-in-ness and Sell-the-news.
Sell-the-newsWhen a stock falls on good news because the good news was already priced in.
Re-ratingA lasting change in the price the market puts on a company, because investors have re-set their view of it. After a readout they re-set two things together: the probability of success and the peak sales. The commercial value ceiling is an upper bound on a positive re-rating, not a forecast of it (01-rules.md, rule 43).
UpliftHow much a price rises above a reference point, as an amount or a percentage. The commercial value ceiling’s uplift is the ceiling price minus spot. The positive scenario’s high edge may not exceed spot plus 1.5 times that uplift, or 3 times with a named exception (01-rules.md, rule 43). Value-uplift potential is the run-up driver that reads the ceiling’s ratio to market capitalisation.
Upside and downsideUpside is how far the price could rise; downside is how far it could fall. The benchmarks report each as an average move and divide the two: a negative-to-positive ratio of 2.0 means a miss moved the shares twice as far down as a success moved them up (07-benchmarks.md § Phase).
Event studyA statistical method that measures how share prices move around many events of the same kind, on average. The benchmarks in 07-benchmarks.md come from one: IQVIA’s 2024 study of 1,502 trial readouts from 2017 to 2023 (07-benchmarks.md § Source and method).
Prior (class average)A starting assumption about a whole class of events, made before looking at the specific case. Example: the average move on a Phase 2 readout. A prior may shape a range or a band. It is never evidence about one program and never an anchor, and a program-specific fact wins wherever the two disagree (07-benchmarks.md § Standing).
Event-move benchmarkA published average of how shares move on a class of clinical readouts: by trial phase, therapy area, trial design or company size. It is a prior over a class, not evidence about one program. It may shape a scenario range or a run-up band through the written basis beside it. It is never an anchor and never counts toward the two anchors a range requires (01-rules.md, rule 30). The tables, their source and their limits are in 07-benchmarks.md.
Impact amplitudeThe spread between the average positive reaction and the average negative reaction to a readout, in percentage points. A larger amplitude means the result moved the shares more, in whichever direction. Single-arm trials have an amplitude of 7%, placebo-controlled trials 29%, head-to-head trials 72% (07-benchmarks.md § Trial design).
Hype cycleA period in which a therapy area is fashionable and shares move far more than the class average on its readouts. Past examples: obesity, rare disease, mRNA during the pandemic, antibody-drug conjugates, GLP-1 drugs. No benchmark number exists for hype. It is a named, argued reason to widen a positive range or a peak band beyond the class average (07-benchmarks.md § Surprise). Recorded as the hype exception, it raises the commercial ceiling’s allowance from 1.5 times to 3 times the uplift (01-rules.md, rule 43).
Peak salesThe highest annual sales a drug is expected to reach, some years after launch.
Expected net present value (eNPV)A drug’s future value in today’s money, adjusted for the probability of success.
RoyaltyA percentage of sales owed to whoever licensed the drug out.
Sublicensing incomeMoney received from partnering a drug to another company. A licence may take a share of it.
Milestone paymentA one-off payment triggered by a development or sales achievement.
Sales multipleThe number of years of peak annual sales an approved drug is worth. It turns a sales forecast into a value. This repo uses a band of 2 to 5 (low 2, mid 3.5, high 5), tagged [UNVERIFIED — modelled convention] because no published source is attached (07-benchmarks.md § Probability of approval).
Scenario price rangeThe price range the shares are expected to trade in if one side of a binary event lands. Two are written for each event: one for a positive result and one for a miss. It is a range and not a single price, because nobody can know where inside it the stock settles. Each range is built from at least two anchors (next row). The positive range runs from the price if the result just meets the positive definition to the price if it beats the expectation benchmark (01-rules.md, rule 30).
AnchorA named, checkable figure a scenario price range is built from, such as the cash floor the company holds. Each anchor is tagged with its own evidence and carries a kind from the short list rule 30 in 01-rules.md admits: cash_floor, dilution, own_past_moves, peer_precedent and options_implied. Each range needs at least two. A range with no anchors is a guess, so the framework does not allow one. An analyst price target, a class benchmark and the commercial value ceiling are never anchors. The ceiling is a cap on the positive range (rule 43).
MidpointThe value halfway between a range’s low and its high. Example: the midpoint of $16 to $24 is $20. The expected value uses the midpoint of each scenario range, and the break-even probability is where spot sits between the two midpoints.
MedianThe middle value of a list sorted from smallest to largest. One extreme value changes it far less than it changes the average. The anchor built from a ticker’s own past catalyst moves uses the median of the comparable moves, never the largest (01-rules.md, rule 30).
Percentage point (pp)The unit for the difference between two percentages. Going from 45% to 60% is a rise of 15 percentage points, not of 15%. Distances between probabilities are measured in percentage points: rule 42 in 01-rules.md records, in percentage points, how far our probability band clears the break-even range.
Signed percentageA percentage written with its plus or minus sign, so the direction is part of the number. +17.5% means above the reference, −12% below it. The expected value against spot, realised moves and returns are all written this way.
Bracket, bandA low-to-high range given instead of a single number, such as a probability of 55% to 75% or an expected move of 20% to 35%. The two words mean the same thing. “Band” is used for probabilities and run-up moves; “bracket” for an expected move when the options chain cannot price one. Every modelled figure in this framework is given as a range (01-rules.md, rule 10). This is not the square-bracket punctuation around the templates’ italic instructions.
Expected valueThe two scenario midpoints weighted by the probability of each. Example: a stock is $10, the positive case is $20, the miss case is $5, and the positive case is 40% likely. The expected value is 0.4 × $20 + 0.6 × $5 = $11. It is arithmetic. It is not advice and it is not a price target. It is written out in full here because EV in this glossary already means enterprise value.
Break-even probabilityThe chance of a positive result at which today’s share price is fair, if our own two scenario prices are right. It is where spot sits between the two scenario midpoints: 0% at the miss midpoint, 100% at the positive midpoint. Only spot comes from the market. Both scenario prices are ours, so the figure describes our scenarios and is not the market’s view. The market’s own probability is future work (build plan item 2.5). Before 6.2.0 the repo called this figure the market-implied probability. breakEven in lib/edge.mjs computes it, together with its range, and it is never hand-computed (01-rules.md, rule 42). A value at or below 0 or at or above 100 means spot sits outside the two midpoints, and the ranges are rebuilt.
Probability gapThe distance between our probability and the break-even probability, in percentage points. It is recorded on every stock call but it no longer decides the call. The call depends on whether our whole probability band clears the whole break-even range (01-rules.md, rule 42). Before 6.2.0 a gap wider than 10 points needed evidence and a gap wider than 25 points was never locked. Those lines are gone.
EdgeA reason to believe the market has mispriced the event. In this repo a stock call has an edge when our whole probability band clears the whole break-even range and at least one item of evidence says why the market is wrong. A call with no edge is written no-edge. See up / down / no-edge under Framework and process, below.
Commercial value ceilingThe most a positive readout can add to a company. It is the drug’s value at approval (peak sales × sales multiple × economic share) times the probability of approval given a positive readout, spread over the economic share count and added to spot. It assumes none of that value is priced in yet, which makes it an upper bound, not a forecast of the re-rating (01-rules.md, rule 43). The positive scenario’s high edge may not exceed spot plus 1.5 times the ceiling’s uplift, or 3 times with a named hype or read_through exception. A program with no peak-sales estimate, or that keeps none of the drug’s economics, has no ceiling and takes a no-edge call. commercialCeiling in lib/edge.mjs computes it; it is never hand-computed.
Economic shareThe fraction of a drug’s economics the company keeps after partnering, royalties and territorial splits. A fraction above 0 and up to 1, never a count of shares. A company that keeps none of the drug’s economics has no commercial value ceiling for it, declares the hole and takes a no-edge call (01-rules.md, rule 43).
Economic share countThe number of shares a per-share figure is divided by: common shares plus pre-funded warrants plus in-the-money warrants and options (03a-company-spec.md C.4). It is not the economic share, which is a fraction of a drug’s economics and not a count of shares.
Read-throughWhen one result changes the value of another asset. Example: a positive result for a drug in one disease raises the value of the same drug in a second disease, or of another drug with the same mechanism. Recorded as the read_through exception and named as a pipeline row in company.md C.2, it raises the commercial ceiling’s allowance from 1.5 times to 3 times the uplift (01-rules.md, rule 43).
Long position, long return, position returnA long position is owning the shares; it makes money when the price rises. A short position is the opposite bet. The long return is what owning the shares from the lock spot to the post-event price would have earned, as a signed percentage (long_return_pct). The position return is what the call implied (position_return_pct): the long return for an up call, its negative for a down call, and zero for no-edge. A stock call is scored on its position return, never on whether the direction word was right (05-prediction-protocol.md § Calibration).
Run-upThe tendency of a stock to drift upward in the weeks before a major expected event. It is also the name of the third call a prediction scores, beside outcome-direction and stock-direction: a position taken before the event and closed before the result lands. The call is pre-registered when the prediction is locked and never written afterwards (01-rules.md, rule 37). It carries an entry, an exit rule, a move band, a predicted peak and a priority score. It is scored on its own, not folded into the other two (05-prediction-protocol.md). A program whose readout date is not known well enough to time a trade carries no run-up call at all (rule 39).
Run-up entryThe start of a run-up call: the date and price a position would have been taken at, and the reason for that date rather than another. In practice it is the prediction’s own lock date and the spot price on it.
Exit ruleWhen a run-up position closes. It is written as a rule relative to the earliest edge of the readout window, not as a fixed calendar date, so the same rule still gives the correct date if the window moves. The three allowed rules are T-5, T-2 and T-1 trading days before readout.window.earliest (EXIT_RULES in lib/runup.mjs). Trading days, never months: see “T-5 / T-2 / T-1 (two different meanings)” under Framework and process, below. The actual exit date and price the rule later resolves to is a separate fact, recorded separately as exit.
Move bandThe predicted share-price move from entry to exit, written as a low-to-high percentage range rather than a single figure. Every modelled number in this framework is banded (01-rules.md, rule 10). At settlement the realised move is compared against it, and whether the band contained it is recorded.
Predicted peakWhere and roughly when a run-up is expected to reach its highest point: a low-to-high percentage move from the entry price, with an estimated date. It need not sit inside the move band and need not fall on the exit date. A stock can reach its highest price and then fall back before the position closes.
Priority scoreOne number, higher first, that ranks programs against each other as run-up trades. It combines seven scored drivers: unmet-need relevance, value-uplift potential, the probability of a positive outcome, date confidence, squeeze mechanics, priced-in-ness, and financing and clustering risk. Each driver carries its own score out of 100 and the reasoning behind it, because a score with nothing behind it is not evidence. Six drivers combine as a weighted average; date confidence then multiplies the result (see Date confidence). The number is stored together with the formula_version that produced it and is never recomputed when it is displayed. A changed formula is therefore a version bump a reader can compare, never a silent drift (01-rules.md, rule 38). lib/runup.mjs is the one place the formula is implemented.
Priced-in-nessHow much of the expected move the market has already made before the event. It is one of the seven priority-score drivers, and it is scored as the room still left to run: a high score means the move has not been priced in yet, which pushes the program up the ranking. That reads backwards from the name, so say which way it reads wherever it appears. Its computed part is the 52-week position. Drift since the last catalyst, ownership crowding and the spread between analyst targets are read alongside it.
Exit-timing errorHow far a run-up’s planned exit missed the peak that actually happened, counted in trading days and signed. Positive means the peak came after the exit: sold early, and part of the move was missed. Negative means the peak came before the exit: held too long, past the top. Zero means the exit landed on the peak’s own trading day.
BacktestA test of a formula or a trading rule against past prices, to see how it would have done. scripts/runup-backtest.mjs scores the run-up priority formula against the committed price cache. It only reports; it writes nothing. The driver weights in lib/runup.mjs are a documented first guess, and a backtest or a settlement is what can argue for changing them.
Rho (rank correlation)A number between −1 and 1 that says how well two rankings agree. 1 means the same order, 0 means no relation, −1 means the opposite order. The run-up backtest reports Spearman’s rho between the priority score’s ranking of programs and their realised run-up moves (scripts/runup-backtest.mjs, lib/runup.mjs).
Quality-adjusted life year (QALY)One year of life in full health. A year in worse health counts for less: a year at 70% of full health is 0.7 QALY. Payers use cost per QALY to judge whether a drug is worth funding.
Incremental cost-effectiveness ratio (ICER)The extra cost of a new treatment divided by the extra QALYs it gives over the old one. Example: it costs $60,000 more and adds 0.5 QALY, so the ICER is $120,000 per QALY. Also the name of a US body, the Institute for Clinical and Economic Review, that publishes a price benchmark. The payer figures are in 07-benchmarks.md § Rules of thumb, and are [UNVERIFIED — illustrative].
Inflation Reduction Act (IRA) and Medicare drug price negotiationA 2022 US law. It lets Medicare negotiate the price of some high-spending drugs. A small-molecule drug becomes eligible 7 years after approval and a biologic 11 years after, so the date a drug could be pulled in limits its late-life sales (07-benchmarks.md § Rules of thumb, [UNVERIFIED — illustrative]).
Risk-adjusted net present value (rNPV)A drug’s future cash flows in today’s money, each weighted by the chance that the drug reaches that stage. It is the same idea as eNPV and the two names are often used for one thing. In the program document, B.4b gives the value in this form.
Base rateThe published share of similar programs that moved on from the same phase, in the same disease area and drug type. It is the starting point for our probability, before any fact about this program is used. It counts moving to the next phase, not meeting this trial’s endpoint. Rule 31 requires the outcome call to name it. It comes from baseRate in lib/success-rates.mjs, which reads data/benchmarks/success-rates.json (07-benchmarks.md § Probability of success — base rates).
Cash floorThe company’s net cash at the expected event date, divided by its economic share count. Net cash is the latest reported cash, less debt, less the cash spent each quarter until the event. It is the starting point of the miss scenario (01-rules.md, rule 30). cashFloor in lib/edge.mjs computes it.
Market Access ScoreThe mean of the five payer criteria in section B.3 of the program document. It rates how likely payers are to reimburse the drug and at what price, on a scale of 1 to 10. scorecard in lib/scorecard.mjs computes it (01-rules.md, rule 45).

Framework and process

TermMeaning
Framework version and stale_beforeframework/version.json declares two versions. version is the framework release the repo is on, such as 6.0.0. stale_before is the oldest per-analysis version stamp that still counts as current. Every company and program record carries a framework_version stamp saying which release it was written to. stale_before is raised only when a release changes what a finished analysis contains: a change to 01-rules.md, 03a-company-spec.md, 03b-program-spec.md, 05-prediction-protocol.md or a schema. A change to how a run is dispatched, how a figure is produced or how a fact is rendered never raises it. Some rules bind only records stamped at or above a fixed version, such as EDGE_REQUIRED_SINCE (6.0.0) in lib/paths.mjs; those anchors never move.
StaleA company or program record whose framework_version stamp is below stale_before. It was written to an older standard than the current one. scripts/stale-check.mjs lists stale records, and § Refresh stale in 06-prompts.md brings them up to date.
Trigger promptA named set of instructions in 06-prompts.md that starts a run. There are ten: Triage, Next shortlisted, Run shortlisted, Refresh stale, Run due, Program analysis, Program refresh, Program re-lock, Company refresh and Settlement. A run starts by citing one with its arguments, never by pasting it. CLAUDE.md § Trigger prompts holds the pointer table.
OrchestratorThe session you type into. It resolves the work, runs the pre-run price fetch, takes the lane-check snapshot, dispatches subagents, runs the two guards when they return, merges the shared ledgers (run-order 6b) and reports. It analyses nothing itself. It writes nothing to the repo between the snapshot and the verify.
SubagentA separate session that the orchestrator starts to do one piece of work: one ticker’s analyses, run one program after another, inside one declared lane. It stops at run-order 6a and returns its records. A subagent never dispatches another subagent and never writes a shared ledger.
DispatchStarting a subagent. Every dispatch names the model it requests. The effort is inherited from the orchestrating session, because a dispatch cannot set one. The two guards, the lane check and the model check, bracket every dispatch: a snapshot before, a verify after.
Lane, lane checkA subagent’s lane is the set of files it may write. For a ticker dispatch it is analyses/<TICKER>/. scripts/lane-check.mjs enforces it: snapshot records the whole working tree before the dispatch, and verify compares the tree afterwards against the declared lane. A subagent that wrote outside its lane fails the check, and its records are held out of the merge. The check reads paths and bytes, never intent, so the orchestrator must not edit anything between snapshot and verify (CLAUDE.md § Scope).
Model, model id, effortThe model is the AI model a session runs on. The model id is its full name, such as claude-opus-5-5; a short alias such as “Opus” is never passed to a script. Effort is a setting for how much reasoning the model spends on each step, such as medium, high or max. CLAUDE.md § Model and effort says which model and effort each task takes. A dispatched subagent inherits the orchestrating session’s effort.
Model checkscripts/model-check.mjs. It reads a subagent’s transcript and reports the model and effort that actually ran, because the runtime can substitute a model mid-run without saying so. The orchestrator passes --expect-model with the full model id. A mismatch is recorded in the prediction’s provenance block and the records are still merged. Unlike a lane violation, it holds nothing back. The orchestrator never reads the transcript itself.
Fallback (model substitution)The runtime delivering a different model from the one a dispatch requested. One substitution is expected: a Fable dispatch delivered on Opus. A dispatch that requests claude-fable-5-1 declares it with '--expect-fallback=claude-fable-5-1->claude-opus-5', written in single quotes and with =. The declared substitution passes the model check and is still recorded. An undeclared substitution, or one to a third model, fails the check. An Opus or Sonnet dispatch declares no fallback (CLAUDE.md § Model and effort).
ProvenanceThe block in a prediction record that says which model and effort produced it: model_requested (what the dispatch asked for), model_delivered (what the model check found) and the effort that ran. The orchestrator fills it from the model check. A subagent cannot observe its own model, so it leaves the block empty.
Shared ledgersThe four files every analysis adds to: data/predictions.json (every locked prediction), data/coverage.json (which connectors were called for each ticker and program), analyses/index.json (the generated index of analyses) and data/triage/queue.json (the programs waiting to be analysed). SHARED_LEDGERS in lib/paths.mjs names them. Only the orchestrator writes them, once, by script, in run-order 6b. They are never read into a session’s context.
Run-order 6a and 6bThe two halves of the last step of the run order in CLAUDE.md § Run order. In 6a the subagent writes the program’s own files: analyses/<TICKER>/<TICKER>.json, company.md, the program README, program.json and the diagrams. Each belongs to one company or one program, so only one run ever writes it. In 6b the orchestrator writes the four shared ledgers through scripts/ledger-merge.mjs, regenerates analyses/index.json with npm run gen-index, and runs npm run validate. 6b happens once per batch, after every subagent is back.
BPIQThe biotech catalyst data feed this repo uses, reached through the BPIQ connector. It lists companies, their drugs and their catalysts. Every program starts from one of its drug records: the record’s integer id is the program’s bpiq_drug_id. Its company-tier calls supply the price and market cap, the drug rows, past catalyst reactions, the options chain, insider transactions and press releases. Several of its fields are wrong in known ways. 02-connectors.md § Known data bugs lists them, and rule 11 in 01-rules.md says to recompute those fields rather than quote them.
ConnectorA tool that fetches data from one outside source. This repo’s connectors are BPIQ, ClinicalTrials.gov, PubMed, Open Targets, ChEMBL, EDGAR (SEC filings), FINRA, the Yahoo Finance options chain and price history, openFDA, the Federal Register, Europe PMC, CTIS, NIH RePORTER and web search. 02-connectors.md says which are mandatory, at which tier, and what each answers. Every call ends in one of the three tool states.
Coverage sweep, coverage gateThe coverage sweep is the full set of connector calls an analysis must make, issued first and all at once, before any interpretation. The coverage gate is the rule it enforces: no program document ships and no prediction locks while any mandatory tool reads NOT CALLED (02-connectors.md § The gate). The sweep’s outcome is recorded in C.0 of the company document, section 0 of the program document and data/coverage.json.
Company tier, program tierThe two tiers of the coverage sweep. The company tier is called once per company and answers company questions: price, cash, the pipeline, past reactions, options, insiders, filings. It is reused by every program under the ticker while the company document is current, meaning swept within 30 days with no earnings release or financing since (02-connectors.md § Company-sweep currency). The program tier is called once per program and answers program questions: the trial record, the literature, target validation, the molecule, commercial estimates. A third tier, the regulatory tier, is called only when the catalyst is a regulatory decision.
CALLED / BLOCKED / NOT CALLEDThe three states a tool in the coverage sweep can end up in. In short: the tool was called and answered, the tool was called and refused, or the tool was never called. 02-connectors.md § Three states is their only definition, so this entry does not repeat the details.
CLEARED / PROVISIONALThe two coverage verdicts a document carries, in the heading of C.0 of the company document and section 0 of the program document. CLEARED means no mandatory tool at that tier reads NOT CALLED; a BLOCKED tool does not prevent it. PROVISIONAL means at least one mandatory tool reads NOT CALLED. A PROVISIONAL program cannot be signed off: its prediction is not locked, the orchestrator does not merge it, and any older prediction stays live (02-connectors.md § The gate; 05-prediction-protocol.md).
Evidence tagsThe label every factual claim carries (01-rules.md, rule 1). [VERIFIED — source]: confirmed by a live tool call, a primary filing or label, a peer-reviewed paper or a trial record, with the source named. [UNVERIFIED]: not confirmable this session, an extrapolation, or a field of unestablished reliability. [WEB ESTIMATE — source, date]: a third-party figure, or a labelled model with stated assumptions. Some tags carry a qualifier after the dash, such as [UNVERIFIED — modelled] for a probability or [WEB ESTIMATE — IQVIA 2024] for a class benchmark. An unverified claim is never written in the same voice as a verified one (rule 2).
ProgramOne drug in one indication: a single row of a company’s drug list, with its own stage and its own expected event. A company can run several at once. Each program is keyed by BPIQ’s integer id, recorded as bpiq_drug_id, and carries a slug (next row). Only a program with has_catalyst: true gets its own document.
SlugA program’s short readable name: lowercase words joined by hyphens, such as arct-810-otc-deficiency. It is unique within its ticker and names the program’s folder, analyses/<TICKER>/<slug>/. The prediction record carries it as program_slug beside bpiq_drug_id, which is the authoritative key.
PipelineAll of a company’s programs together, including the ones that have been stopped. The pipeline table in company.md (section C.2) lists every one. A company that has shut programs down has a track record, and its surviving programs carry more of the value.
MaterialityHow much a single program matters to the share price: dominant, meaningful or immaterial. Good news on an immaterial program moves the stock very little, however good the science is. The stock-direction call must cite the materiality recorded in the pipeline table (01-rules.md, rule 27).
Placeholder dateBPIQ’s catalyst_date when the company has only named a period. BPIQ then stores the last day of that period: “Late 2026” becomes 2026-12-31, “Late Q3 2026” becomes 2026-09-30, “Mid-2027” becomes 2027-08-15, and even “TBA” becomes 2026-06-30. A placeholder is not a date. It is never used for timing, because it always says “hold longer” (01-rules.md, rule 23). An analysis reports catalyst_date_text instead and reads every timing decision from the readout window. The site labels a program’s timing as placeholder or readout so a reader can tell which it is showing (lib/timing.mjs).
Readout windowThe framework’s own judgement of when a program’s result is expected. Every timing decision reads it instead of BPIQ’s catalyst date. It is three dates. Earliest is the first date the readout could plausibly land. Likeliest is the single best guess, which need not be the midpoint of the other two. Latest is the last date it could plausibly land. All three are judged from the sources gathered for the program, never copied from any one of them. Everything is anchored on the earliest date: the re-analysis cadence, the stock-call window, clustering, and a run-up’s entry and exit. The reason is that a position has to exist before the window opens, so the conservative edge is the one that matters. A window is absent when no source discloses a date at all, and nothing falls back to BPIQ’s placeholder in its place (01-rules.md, rule 23).
Readout precisionHow precisely a readout window is known. DAY: a source names a day. MONTH: a source names a month or a season. PERIOD: only a half-year, a quarter or a year is named. UNKNOWN: nothing is disclosed at all. UNKNOWN is a valid answer, written out plainly, never an empty cell. Being open that there is no date is the requirement; inventing one is the failure this framework exists to prevent (01-rules.md, rule 33). The shared list is PRECISION in lib/paths.mjs.
Readout confidenceHow much the window judgement trusts the sources it was built from: HIGH, MEDIUM, LOW or UNKNOWN. It is a separate question from precision. A source you have reason to doubt can name a day, and a half-year can be as solid as the company’s guidance ever gets. UNKNOWN is valid here for the same reason it is valid for precision. The shared list is READOUT_CONFIDENCE in lib/paths.mjs.
Readout disagreementWhether the sources behind a window agree with each other. NONE: there is only one source to judge. CONSISTENT: several sources agree. UNRESOLVED: they conflict. A disagreement is reported, saying which source points which way. It is never averaged away and never settled by picking one (01-rules.md, rule 24). The shared list is READOUT_DISAGREEMENT in lib/paths.mjs.
Readout lag benchmarkThe gap between a trial’s registry primary completion date and the day its sponsor actually announced topline results, collected for comparable completed trials in data/benchmarks/readout-lag.json. A modelled readout estimate is measured against it. Until a matching entry exists, the estimate uses the stated default of primary completion plus two to four months and says so in its tag, [UNVERIFIED — modelled, default lag]. Once one exists, the tag names how many comparable trials the figure rests on, [UNVERIFIED — modelled, benchmarked n=N], so a reader can tell a stated default from a benchmarked estimate at a glance (01-rules.md, rule 34). An observation is added only after a trial has actually read out and both its dates are confirmed from a primary source. Nothing is estimated, interpolated or backfilled. The file ships empty by design.
Date slippageHow often a company has pushed back the expected date of a readout. It is read from BPIQ’s note field, which logs every change to a catalyst date, and from the company’s own releases, oldest statement first. It is recorded as part of the readout judgement. Repeated pushes signal execution risk. A date that never moves is evidence of the opposite. Zero slips is a finding too, and is written down as zero rather than left out.
Date confidenceHow well a readout date is known, scored from 0 to 100 from readout precision combined with readout confidence. It is one of the seven priority-score drivers, and the only one that works structurally: the other six are multiplied by it. So a program whose date is not known well enough to time an entry or an exit cannot out-rank one whose date is, however attractive everything else about it looks. It is exactly 0 when precision is UNKNOWN, and such a program carries no run-up call at all (01-rules.md, rule 39). lib/runup.mjs implements it.
Cadence gateA checkpoint at which a program is re-analysed as its readout approaches: 5 months, 2 months and 1 month before the earliest edge of its readout window (01-rules.md, rule 35). A gate is open once that distance is reached. It is covered once a locked prediction sits inside the gate’s own span. An open gate that no prediction covers is what makes a program due. The gates are derived fresh on every run from what is already on disk; they are never stored as a schedule. A program with no readout window is reported UNSCHEDULABLE, never scheduled off BPIQ’s placeholder date. A month here is a fixed 30 days, so a boundary does not shift with which months are being counted through. lib/cadence.mjs computes the gates and scripts/cadence-check.mjs reports them.
T-5 / T-2 / T-1 (two different meanings)The same three labels name two unrelated things in this framework. Only the unit separates them. A cadence gate’s T-5, T-2 and T-1 are counted in months before the readout window’s earliest date. They say when a program is re-analysed (01-rules.md, rule 35; GATES in lib/cadence.mjs). A run-up exit rule’s T-5, T-2 and T-1 are counted in trading days before that same date. They say when a position closes (EXIT_RULES in lib/runup.mjs). A cadence T-5 falls five months before the window opens. A run-up T-5 falls about a week before it. Reading one as the other would move a trade by months. The numbers are reused because 5/2/1 is memorable, not because the two things measure anything alike. Write the unit every time: “T-5 months” or “T-5 trading days”, never a bare “T-5”.
Catalyst clusteringThe check for whether two catalysts on the same ticker land so close together that a price move around either one cannot be attributed to it alone. Two rows conflict when their date ranges sit closer than CATALYST_CLUSTER_MIN_MONTHS apart: six months, set in lib/paths.mjs as a judgement call rather than a figure derived from data. The check covers every row in the company’s pipeline table that carries a catalyst, analysed or not, because a program with no document of its own can still land inside another program’s window. An imprecise row is widened rather than narrowed, so it collides with more than a narrow one would. A false conflict costs a caveat sentence a reader can discount. A missed one costs a stock call presented as clean when it was not. lib/clustering.mjs computes it (02-connectors.md § Catalyst clustering).
AttributionWhat the clustering check concluded about one particular program: whether a price move around its own catalyst can be read as belonging to it alone, together with every conflicting row the check found. It is computed and transcribed, never judged by eye. A placeholder half-year and a real day can sit at opposite ends of the six-month bar in ways that are easy to misread. A program whose window is contaminated may not present its stock-direction call as though the move were attributable to it alone (01-rules.md, rule 36). It carries exactly one of the four statuses below.
CLEANNo other catalyst on the ticker lands close enough to conflict with this one. The list of conflicts is empty.
CONTAMINATEDAt least one conflict rests on a date somebody actually disclosed, on at least one of its two sides: a readout window of DAY or MONTH precision, or a pipeline row whose catalyst date is exact. That is a real collision, not two placeholders touching. It requires a note explaining it, and the stock-direction call must not be presented as attributable to this program alone.
INDETERMINATEConflicts exist, but every one of them is a guess on both sides: two PERIOD- or UNKNOWN-precision ranges merely touching. That is evidence of not knowing, not a finding, and it is reported as its own answer instead of being folded into CONTAMINATED. No note is required on each conflict. The only thing such a note could say is that the dates on both sides are placeholders, and repeating that on every conflict would teach a reader to stop reading the notes. The stock-direction call must still say plainly that attribution could not be determined for this program.
WAIVEDAn analyst has read a CONTAMINATED or INDETERMINATE finding and judged that it does not contaminate this particular call. Only a person can decide it; the computation never produces it on its own. It requires a note saying why.
Independent voiceA named commentator on a program’s endpoint who is not an investigator on its own trial and has no disclosed relationship with its sponsor. “Independent” names the role being looked for. It does not certify an absence of bias. What makes that absence checkable is the record of where the search for conflicts was made (see Conflicts checked). A disclosed relationship is a contradiction here, not an ordinary finding: it moves the person into the investigator group instead. A panel of independent voices is never collapsed into a single sentiment score. One number would hide the very conflict structure that makes the individual views worth reading (01-rules.md, rule 41).
Conflicts checkedThe record of where a search for one person’s conflicts of interest was actually made, named source by named source, each with its date. It is required beside every investigator and every independent voice, and above all beside the ones whose conflict list came back empty. An empty list is a finding only once the search behind it is on record. It is never a default assumed from silence (01-rules.md, rule 40). The record says that the search happened. It never certifies what the search did not find: a journal with a thin disclosure policy, or a relationship formed after a paper was submitted, produces a clean-looking record for a person who is not clean.
Probability (modelled)The chance, written as a whole-number percentage, that a trial meets the exact positive definition written down before the result is known. It is worked out from evidence rather than measured, so it is always tagged [UNVERIFIED — modelled] and given with its reasoning and a band (01-rules.md, rule 31). A number can be scored afterwards; a word like “medium” cannot.
Probability bandThe range around a modelled probability, for example 55% to 75% around 65%. It says how firmly the single number is held. A probability without a band claims a precision it does not have.
Outcome-direction callThe first of a prediction’s three calls: will the readout succeed on its primary endpoint? It states a direction (positive, miss or uncertain) and the modelled probability, with its band, that the positive definition is met. The probability must agree with the direction: above 50 for positive, below 50 for miss, a band spanning 50 for uncertain (01-rules.md, rule 31). At settlement it is scored on the probability with the Brier score, not on the direction word.
Positive definition (settlement criteria)The written test for what counts as a positive outcome, settlement_criteria.positive_definition. It is written at lock time, in enough detail that a third party would score it the same way, together with source (the primary source that will settle it) and settles_on (the event that triggers settlement). The modelled probability is the chance this definition is met. Writing it afterwards would fit the test to the result. From framework 6.3.0 it also names the endpoint’s counting window, the side of any p-value test it requires, and how a switch to a successor product scores (05-prediction-protocol.md § Pre-registered settlement criteria).
Stock-direction call (stock call)The second of a prediction’s three calls: which way the shares move on the result, within a stated time window. The window names the event the call is a view on. The call settles on the readout’s announcement day, inside or outside that window (05-prediction-protocol.md § The stock call carries a window). It carries the spot price, two scenario price ranges with their anchors, the expected value, the break-even probability with its range and the clearance of our band, the commercial value ceiling, and the call itself. It cites the program’s materiality (01-rules.md, rule 27). It is scored on the position return, not on the direction word (05-prediction-protocol.md § Calibration).
up / down / no-edgeThe three values a stock call can take. The call follows from the bands; nobody chooses it. up needs the low edge of our probability band above the high edge of the break-even range, with at least one item of evidence. down needs the high edge of our band below the low edge of the range, with evidence. Anything else is no-edge. The clearance in percentage points is recorded. A call carries no confidence rating: unrated for up and down, none for no-edge. callFromBands in lib/edge.mjs applies the rule (01-rules.md, rule 42). no-edge is legitimate whenever the bands overlap or the evidence is thin.
Expectation benchmark, expectation gap, expectation-gap bucketThe expectation benchmark is the result investors are pricing in before a readout, recorded as scenarios.positive.expectation_benchmark. It is read from the target product profile (A.1; A.3b before 6.2.0), a competitor’s data or published consensus. The expectation gap is the distance between the actual result and that benchmark. The benchmark study sorts readouts into four expectation-gap buckets: significantly exceeded, met, missed and significantly missed expectations. The dramatic moves live in the two outer buckets (07-benchmarks.md § Surprise). The positive scenario range runs from the price if the result just meets the positive definition to the price if it beats the benchmark, and says which bucket each edge corresponds to (01-rules.md, rule 30).
Verdict blockA short piece of unscored prose at the end of a program analysis, directly above the locked prediction. It says in plain words what a reader should do with the scored calls: what I would do, why, what would change this, and what to watch. Evidence tagging still applies to it (05-prediction-protocol.md § The Verdict block).
Locked prediction (lock)To lock a prediction is to write it down and fix it before the result is known. The locked prediction is the record appended to data/predictions.json at the end of an analysis. It is identified by the ticker, the bpiq_drug_id and the catalyst, and it carries the three calls, the settlement criteria and the provenance. Once locked it is never edited. A later view is a new record that supersedes it.
Supersede, live record (live prediction)When a program is analysed again, the new prediction supersedes the old one. The new record names the old one in supersedes; the old one gains superseded_by and nothing else. The live record is the one with no superseded_by. It is the call of record, and exactly one record per program is live. Everything else in the chain is history and is scored as history. A settled record is never superseded, and a PROVISIONAL result supersedes nothing (05-prediction-protocol.md § A refresh supersedes rather than revises).
Settlement (settle)Scoring a locked prediction against what actually happened. It happens only when the user explicitly asks, against the pre-registered settlement criteria, from a confirmed primary source. It records actual_outcome, actual_move_pct and settlement_source, plus stock_direction.actual and runup.actual where the call exists. actual_move_pct is measured on a fixed window: the average of the two closes before the announcement day against the average of the announcement-day close and the next trading day’s, read from the price cache. The original prediction text is never revised. scripts/ledger-merge.mjs settle writes the result (05-prediction-protocol.md § Locking and settlement).
Brier scoreThe score an outcome call gets at settlement: the squared difference between the probability, as a fraction, and the result, where the result is 1 for a positive and 0 for a miss. Lower is better. Example: a 65% probability followed by a positive result scores (0.65 − 1)² = 0.1225; the same 65% followed by a miss scores 0.4225. It separates confident wrong calls from hedged right calls, which the direction word alone cannot. Across settled records, our Brier score is compared with the base rate’s, the independent benchmark, and with the break-even probability’s, which is labelled as coming from our own price ranges (05-prediction-protocol.md § Calibration).
CalibrationTwo related uses. First, the long-run check of whether our probabilities match how often things actually happen, measured with the Brier score across settled predictions. It is the reason the ledger exists (05-prediction-protocol.md § Calibration). Second, shaping a range or a band against the class averages in 07-benchmarks.md: how deep a miss range sits, how wide a run-up band is. A re-lock re-derives a prediction “under the current calibration” in this second sense.
Re-lockA superseding prediction produced without a new evidence sweep. The judgement layer is re-derived under the current calibration: the scenario ranges, the expected value, the break-even probability and the call it implies, the commercial ceiling, the run-up entry and bands, and the computed drivers. The probability, the anchors, the readout window and the settlement definition are carried verbatim from the record it replaces. It is allowed only while the underlying sweep is within 30 days and a two-call currency check finds no intervening event. Its written rationale names it a re-lock, so the ledger can tell it from a refresh without a diff (06-prompts.md § Program re-lock; 05-prediction-protocol.md). A re-lock’s record covers a cadence gate exactly as a refresh’s would (01-rules.md, rule 35). The document shows the split: a re-lock stamp sits beside the original header stamp, while as_of and the document’s framework version keep describing the evidence sweep (program.schema.json, relock).
Triage boardThe running list of screened candidates, built from every triage run file in data/triage/ and shown at /triage on the site. Screening is cheap and shallow, so nothing on the board may be cited in an analysis. It exists to decide what is worth analysing next, not to support a claim (02-connectors.md § Triage pass).
Board stateWhich of the four states a screened candidate is in: screened, shortlisted, dismissed or analysed. A candidate is in exactly one at a time. Two states are decisions a person writes down; the other two are derived from what is already on disk (02-connectors.md § The board lifecycle).
ScreenedA candidate has appeared in at least one triage run and no decision has been taken on it. Taking no decision is a valid position and needs no record.
ShortlistedA person has queued the candidate for a full analysis by adding a line to data/triage/queue.json. The next-shortlisted and run-shortlisted prompts read that queue. Run-order 6b deletes the line once the analysis is on disk.
DismissedA person has screened the candidate out by adding a line to data/triage/dismissed.json. A dismissal carries both a dismissal reason and a written note, because a dismissal with neither is a deletion rather than a decision. A dismissal is never deleted, so it can be reviewed later against what actually happened.
Dismissal reasonThe short tag attached to a dismissal, chosen from a fixed list such as financing-first or not-a-binary. The tag says what kind of dismissal it is, so dismissals can be counted. The note beside it says what specifically was wrong with that candidate. The list lives in data/triage/reasons.json.
AnalysedA full program analysis exists for the candidate. It is never set by hand: it is derived from analyses/index.json, so it cannot disagree with the analyses on disk.
Derived stateA state worked out from what is already on disk rather than typed by a person. Analysed and archived are derived, so they cannot drift from the analyses and the settled predictions they are read from. A row is archived once its catalyst date has passed or a prediction on it has settled.
Executive scorecardThe first page of a 6.2.0 program document. It scores the asset on seven domains, from 1 to 10, with a weight for each, and adds them into a weighted total. It carries the asset verdict. The scores inform and never decide a call (01-rules.md, rule 45).
Asset verdictThe scorecard’s judgement of the asset’s long-term value: Invest, Monitor or Pass, with a reason. It never moves the outcome, stock or run-up call.
Mode A and Mode BThe two ways the program template is used. Mode A is a program still in development: it ends in a locked prediction. Mode B is a look back on a drug that is approved, failed or withdrawn, written as an asset review. It ends with no prediction.
Look-backThe section (D) that compares what we expected with what happened. In a Mode A document it holds one fixed sentence until the prediction settles. At settlement it is written in its own file, lookback.md, beside the program document.
Asset reviewA Mode B document: the look back on a whole drug, once it is approved, failed or withdrawn. It lives in reviews/<slug>/ and carries no prediction. It serves as a precedent for later programs.
Stage-gate questionOne of the Go / No-Go questions in section A.3 of the program document. Each asks whether the drug has cleared a hurdle that a decision to continue depends on, with a gate column for each step (to Phase 2, to Phase 3, filing), a rating and a confidence.