Assumptions
Every judgement call in the analysis, with the reasoning and how to change it. Anything not listed here is either read directly from the data or defined in the organisers’ glossary.
About the data
The dataset is a retail prescription audit. Value, volume and price to retailer by pack, brand and company, in the IQVIA-style format used across the Indian industry. It is treated as the market, which it is not: institutional, hospital and government-tender channels are not represented. That matters given the scale of public screening programmes, and it is stated in the limitations rather than adjusted for, because the dataset gives no basis for an adjustment.
Values are in INR crore. The workbook carries no unit label. Crore is inferred from the total of ₹23,244 for a therapy area covering anti-hypertensives, lipid regulators and anti-angina, which is consistent with published estimates of the Indian cardiac market. Every output states the unit explicitly so a misreading is visible.
Quantities are in the pack-dependent units the glossary describes (tablets, millilitres or doses). They are therefore comparable within a space and not across spaces of different dosage forms. Volume growth is used as a within-space rate, never as a cross-space level, which is what makes this safe.
MAT CP for the latest period is stated at the previous year’s prices. This comes from the glossary and determines how real growth is computed: MAT CP Feb’26 against reported MAT Feb’25, not against MAT CP Feb’25. Getting this wrong would misstate the price contribution across the entire analysis.
A trailing asterisk marks a consolidated corporate group. CIPLA*, SUN*, ZYDUS CADILA*. The asterisk is stripped for display and the raw value kept for joins. Consolidated groups are the right unit for competitive analysis because they share a field force.
Zero-value rows are real. Roughly 1,300 packs report zero in the latest MAT. These are discontinued or pre-launch, not missing data, and are retained so player counts and share churn are correct.
About molecules and brands
Salt forms are the same active ingredient. “AMLODIPINE BESILATE” and “AMLODIPINE” roll into one franchise, as do metoprolol succinate and tartrate. The stripped suffix list is deliberately conservative and excludes nitrate esters, because isosorbide mononitrate and dinitrate are genuinely different drugs rather than salt forms of one.
The metoprolol case is a real simplification: succinate is extended-release and tartrate is immediate-release, and they are not clinically interchangeable. For franchise adjacency, which is what the field uses, treating them as one molecule is correct. For a formulation decision it would not be.
A combination pack’s value splits evenly across its ingredients when computing molecule adjacency, so a triple FDC does not count three times over. An even split is a simplification: the therapeutic and commercial contribution of each component is not equal. No allocation basis exists in the data, and an even split is at least neutral.
Anchor-molecule spaces overlap by design. A Cilnidipine + Telmisartan pack counts towards both franchises, so anchor values sum to more than the market. This is the point, and it is why anchor spaces are ranked separately rather than pooled with the hierarchical levels. Every output that shows them says so.
Brand root recovery is heuristic. A hyphen is treated as an unambiguous line-extension marker, so AMLOPRES-AT reduces to AMLOPRES. A trailing token reduces only when it is short or a known modifier, so ROSULIP GOLD reduces while ACE REVELOL does not. The rule is tested against real brand names in the suite. It will occasionally be wrong on an unusual name, and the effect is bounded: a wrong root affects one component of one pillar for one space.
About the framework
Percentile normalisation within a level. Audit data is heavily skewed and a common min-max scale across levels would measure list length rather than market structure. See PRIORITIZATION_FRAMEWORK.md.
Weights are judgement, and are exposed as such. Every weight in config/settings.yaml is a defensible choice rather than a derived optimum, because there is no labelled outcome to optimise against. That is precisely why the sensitivity analysis exists and why the agent is required to run it before presenting a recommendation.
Future potential outweighs current attractiveness, 40 to 34. The case asks which spaces will outperform over three to five years, not which are largest today.
Reported value growth carries no weight in future potential. Price gains do not compound the way prescription volume does, and a material part of the essential cardiovascular basket sits under administered ceiling prices. It appears in market attractiveness, where it belongs.
The right-to-win gate is a multiplier, not a fourth pillar. Adding it into a single weighted score would let a very attractive, completely inaccessible space outrank a moderately attractive winnable one. A multiplier cannot do that, which is the behaviour the case asks for.
The gate floor is 0.35 rather than zero. A space Cipla cannot win today may still be worth watching, and a zero floor would remove it from the output entirely rather than ranking it low.
The gate curve exponent is 1.35. Above 1.0, so weak right to win hurts more than linearly. For a company at 1.68 per cent therapy share, chasing growth into spaces with no structural advantage is the failure mode to protect against.
Size floor of ₹100 crore. Below this a win cannot move a company of Cipla’s size, and small bases produce unstable percentages. It is the single most consequential filter: it removes 213 of 321 constructed spaces.
Minimum of three active players. Fewer than three is treated as structurally closed, most often a single-brand or originator-held space.
About the forecast
The therapy area keeps growing ahead of the wider pharmaceutical market. Grounded in the disease-burden and screening evidence in the corpus, not assumed. It is the assumption that most affects the projections and would be the first thing to revisit against a longer history.
Excess growth over the market decays 28 per cent a year. Nothing sustains 45 per cent for five years, and assuming it does is how forecasts produce indefensible numbers. The rate is a judgement calibrated so a high-growth space lands at a plausible five-year CAGR rather than an absurd one.
Growth is capped at 45 per cent and floored at minus 15 per cent. Bounds on extrapolating a two-year rate forward five years.
Bull and bear shift the starting rate by 4 percentage points. A band rather than a distribution. With two observations there is no basis for a variance estimate, and presenting one would imply precision the data does not support.
Base rate weights: 40 per cent real growth, 30 per cent volume, 20 per cent two-year CAGR, 10 per cent momentum. Demand-led measures dominate, for the same reason they dominate the future-potential pillar.
About the external signals
Fourteen documents, chosen for relevance rather than systematically sampled. This is curation, not a literature review, and the corpus is therefore incomplete by construction. Its influence is bounded to a narrow band for exactly that reason.
Confidence weights: high 1.0, medium 0.6, low 0.3. A judgement about how much a stated magnitude should be trusted.
Within a category, the strongest signal counts fully and the rest at half weight. Two guideline documents describing one clinical shift are one piece of evidence reported twice.
Tilts are centred on the median tilt of their space level. A signal that applies to every space carries no information about which space to prefer, so it must not move the ranking. Without this the multiplier saturates at its ceiling for nearly every space and stops discriminating.
The multiplier is clipped to 0.80 to 1.25. Secondary research can shade the ranking; it can never overturn what the audit shows.
Four signals are derived from the supplied dataset rather than published. They are labelled internal:// and marked as derived in every citation, so a reader can tell them apart from external sources at a glance.
About the agent
The model never calculates. Every figure comes from a tool result, and the numeric guardrail enforces it. This is an architectural constraint, not a prompt instruction.
Baseline tool calls run regardless of what the model decides. The planner schedules evidence by intent before the model is consulted, which is what makes the agent gather the right data even when the model chooses badly, and what makes the no-credentials path work.
One rewrite on a failed grounding check, then fall back. If an answer cannot be grounded in two attempts, the honest outcome is to return the deterministic rendering rather than keep asking until something slips through.
Numeric tolerance of 2 per cent, ignoring values below 3. Rounding in prose is legitimate; fabrication is not. The floor stops list positions and small counts generating false positives.
Scope enforcement is asymmetric. Topics the dataset cannot contain, such as share price or clinical advice, refuse unconditionally. Adjacent therapy areas refuse only when nothing in scope is present, because “how does cardiac compare with respiratory” is a legitimate framing question. A question with no in-scope vocabulary at all is refused, because answering it would produce market context with no bearing on what was asked.