The evidence base is 20 printed pages.
Here is exactly where it ends.

The top three clients sum, to the second decimal, to the share of revenue the prospectus reports on contracts of at least seven years. The document never joins those two tables, and joined they say the durable contract base and the client concentration are the same three counterparties.

build fails: The reconciliation that carries the finding no longer holds.

Clients one, two and three are Hyperscalers in every period. The concentration exhibit calls them that, and the filing is what says so.

build fails: the filing states clients 1, 2 and 3 are Hyperscalers in every period

The same headline capacity figure is defined twice, on two different printed pages. If they ever land on one page the finding is gone, so the guard requires them to differ.

build fails: the two capacity definitions must cite different printed pages

One government order is 69 per cent of the order book. An order larger than the book holding it would push that share past a hundred, which is arithmetic nonsense rather than a finding.

build fails: cr order book it is measured against

The order book must be struck after the order was awarded, or the order is not inside it and the concentration exhibit is measuring nothing.

build fails: does not postdate the award at

The page's claim about the struck off list. The list sits a few notes below the related party disclosures and four of its eleven rows are related parties, so the obvious reading is that this is a related party problem. It is not: the largest balance on it, by two orders of magnitude, is with a counterparty the report itself classifies as Others. If a later edit moved that row the sentence on the page would be wrong rather than merely stale.

build fails: and the page says it is not

The government scheme's deployment ledger needs at least one provider reported ahead and at least one behind. That separation between announcement and delivery is the only reason the ledger is on the site, and without both sides the exhibit shows a list of names rather than a finding.

build fails: the deployment ledger needs a provider on each side

Every Indian operator in the table, not merely the issuer, earned a lower return on capital in the second year than the first. The sector page leads on that, so if one of them rises the page is claiming a decline its own primary source no longer shows.

build fails: every domestic operator's return on capital fell across the first two years

The exhibit claims the published forecasts disagree with each other by a factor of three. If one publisher held both the lowest and the highest number, that would be a single house's own band drawn as a disagreement, which is a different and much weaker claim.

build fails: the lowest and highest forecasts come from one publisher

Three foreign firms have pledged more than the annual size of the entire Indian data centre market they are pledging into. That mismatch is the exhibit's headline. Below that line the pledges are an ordinary capital cycle and the exhibit has nothing left to say.

build fails: no longer exceed the annual market they are pledged into

The capacity half of the pledge exhibit rests on one sentence: a single announced site is most of what the whole country currently operates. Only one of the three pledges named a capacity at all, so if either that figure or national live capacity moves, the sentence goes with it.

build fails: the largest single announced site is under half of national live capacity

The page argues the grid is the binding constraint partly because electricity is the largest single thing a colocation operator buys. Under half of operating cost it is one line among several and the argument is weaker than the page states it.

build fails: power is no longer the majority of operating cost

The finding, asserted rather than described. Annualising the exit monthly recurring revenue must land below the revenue of the year reported beside it: a year described as up seventy three per cent that ends at a run rate a fifth under its own average has its growth behind it. If a later report closes that gap the claim has stopped being true and the build should fail instead of the sentence going stale.

build fails: no longer annualises below the year reported beside it

Capital expenditure is the four lines the investing section prints. Against operating cash flow as filed it is more than covered; against operating cash flow with the borrowings line taken out it is comfortably covered. The exhibit's whole sentence is that one presentation choice moves the answer across one, so if a restatement ever lands both readings on the same side, the build stops instead of the sentence going quietly wrong.

build fails: must change side when that line leaves operating activities

Operating cash is under half the profit reported beside it, and it stays under half on the other reading of the borrowings classification as well. That is the cash conversion pillar's headline for this name, and the whole point of publishing it is that it does not depend on choosing between two presentations of the same statement. If either reading ever clears the threshold the sentence is gone and the build stops.

build fails: must stay under half the profit reported beside it on both readings

Every filed period spends more on investing than it recovers. That is what makes the accrual ratio on this filer move with the size of the build rather than with the quality of the earnings, which is the caveat published beside the ratio. A period that turned positive would make that caveat wrong.

build fails: investing cash is no longer an outflow

The policy takes the lower of three tests, and the profit one binds by a wide margin because profit is small against both net worth and turnover. That is what the exhibit says, and it is the reason the disclosure bar is far lower than a net worth test alone would set. If earnings ever grow into either of the other tests the bar rises several times over without the policy changing a word, and the sentence resting on it is gone.

build fails: the disclosure threshold is no longer set by the profit test

The data centre arm this company is priced on is about one per cent of the revenue the group actually reports. That is the deep dive's central claim, and above a tenth of group revenue the sentence is gone.

build fails: the data centre arm is no longer a small share of group revenue

The data centre subsidiary made a loss in the year read. The page says the capacity everyone quotes sits in an entity that does not yet earn, so a profitable arm would be a different page and should fail the build rather than quietly change the meaning.

build fails: the data centre arm no longer loses money

The group reports one segment and it is real estate development, which is why no data centre revenue, margin or asset line exists anywhere in the consolidated accounts. If the group ever reports the data centre separately, the central claim on that page changes and the exhibit has to be rebuilt rather than left standing.

build fails: the group no longer reports a single segment

The Noida campus carries a headline of 16 MW and a first phase of 500 kW, so the announced against delivered gap this site exists to draw appears inside a single site rather than only across an estate. If a restatement ever makes a stated first phase a material share of its campus, the sentence resting on it is gone and the build stops.

build fails: a stated first phase is no longer a small fraction of the campus

Across every filed period taken together, spending on property, plant and equipment exceeds the cash operations produced. That is the claim the cash flow exhibit makes out loud: the estate was built with money the business did not generate. If it ever inverts the headline has stopped being true.

build fails: no longer exceeds the

Net debt must equal borrowings plus lease liabilities less cash. The formula the document prints for capital employed never says whether a lease is a borrowing, and the return on capital rebuild only lands on the published figures if it is, which is an inference drawn backwards from a result. This identity states the same convention forwards, from a line printed beside the one that needed it, and it holds in every filed period. If it stops holding, the claim that the issuer treats leases as debt has stopped being true.

build fails: The issuer may have stopped counting leases as debt.

The key management remuneration line must cover fewer officers than the note names, or the finding it carries has gone away. A disclosed figure covering every officer named is an ordinary disclosure and the page should stop calling it a shortfall. A guard on the amounts would not catch that, because the amounts would still be there.

build fails: officers named, so it no longer understates anything

The filer targets 250 MW and calls the business its most consequential decision in a generation, and the only line in either set of accounts that records contracted future capital spending is smaller than the tax demands it is disputing. The page draws the two against each other, so a reversal removes the comparison rather than weakening it.

build fails: contracted capital now exceeds the tax it disputes

No segment disclosure appears anywhere in 433 printed pages, so the data centre business cannot be sized apart from the engineering business funding it. The guard fires when that improves, because the page states the absence and would need rewriting.

build fails: the report now separates the data centre business

One clause of the consolidated statutory information note tables a balance with a company struck off the register, and the next clause, sharing its numeral, states the Group has no such transactions. The contradiction is the finding, so the balance has to be present for the page to say so.

build fails: no longer tables a balance beside the sentence denying there is one

Almost everything the standalone reports as current borrowings arrived by moving a liability from trade payables to borrowings with no cash involved. The page states that share as a percentage, so it has to keep holding.

build fails: the transfer is no longer most of what the standalone calls current borrowings

The ageing table heads its columns from the due date of payment. The whole micro and small enterprise balance sits in an overdue column and nothing sits in the column for amounts not yet due, in both filed years. The page states that, so a single rupee moving into the not due column changes what it can say.

build fails: the micro and small enterprise dues no longer sit entirely past their due date

Interest on payment after the appointed day is automatic under the Act, and all five statutory clauses report nothing in both years, one printed page after the table that placed the balance past its due date. The contradiction is the finding, so both halves have to stay as they are for the page to state it.

build fails: so the five that report nothing are no longer five

In every filed period, built capacity is at least commissioned capacity, which is at least what is sold.

build fails: capacity rungs must descend, got built

The capacity rungs are stored widest first. The finding is that the widest one is an engineering maximum while being called built.

build fails: capacity rungs must descend, widest first

Announced, then claimed operational, then actually handed over. That descent from 307 to 28 to 8 megawatts is the entire Anant Raj case.

build fails: the capacity ladder must descend, got

What is commissioned and live sits inside the three campus portfolio, which sits inside the capacity targeted by Fiscal 2030. This is the same descent every other operator on the site is held to, and an inversion means a campus was misread or a target has quietly been restated downwards.

build fails: must descend from the target through the portfolio to what is live

The third statement of the customer concentration, and the only one inside the accounts the auditor examined. The client table and the long contract share are both risk factors, written by the issuer for its own document, so their agreeing proves less than it looks. Note 33 reports the same revenue as an amount and lands on the client table to the paisa in every filed period. If an edit breaks that, the risk factors and the examined accounts have stopped agreeing and the build should say so.

build fails: The risk factor and the examined accounts no longer agree.

The site rows must sum to the totals printed in the filing, within the rounding the filing itself introduces. A wider gap means a row was mistyped.

build fails: sites sum to

The fresh issue and the offer for sale must add up to the total offer, so the exhibit showing how much never reaches the company is built on the issuer's own split.

build fails: the fresh issue and the offer for sale must add up to the total offer

For each object, money already spent plus net proceeds plus borrowings equals the estimated cost. That identity is what lets the funding gap exhibit say the borrowings are the issuer's own figure rather than our subtraction.

build fails: deployed plus net proceeds plus borrowings must equal the total estimated cost

A segment can never be larger than the group it sits inside, in any filed year.

build fails: segment revenue exceeds group revenue in at least one year

Property, plant and equipment is one line inside total assets, so a total smaller than its own component means the two were harvested against different periods or different units. Neither failure announces itself: the accrual ratio would still divide, and still return a number a reader would believe.

build fails: total assets are smaller than the property, plant and equipment inside them

Confirmed, partial, deflected and declined must account for every question pressed, so the refusal rate has a denominator a reader can check.

build fails: the four response qualities must account for every question pressed

The per topic counts must add up to the stated denominator for the company.

build fails: the family counts must add up to the stated denominator

The headline refusal count must match the topic breakdown beneath it.

build fails: the refusal count must match the family breakdown

Every counted refusal must be listed individually, so a reader can read the numerator rather than being told it.

build fails: every counted refusal must be listed, so the numerator can be read

The annual report prints one headline capacity number in its highlights and, elsewhere, the three pieces it is made of, only one of which is operational. Both are stored and the identity is asserted, which is what lets the exhibit say the composition is the company's own rather than our subtraction.

build fails: the headline capacity figure must equal the parts the report says it is made of

The consolidated income statement adds up to the profit before tax printed under it. Every line is typed from one printed page, and a mistyped expense would change the margin the page quotes while changing nothing a reader could see.

build fails: does not reconcile to its own profit before tax

The summary income statement must add up to the profit before tax printed under it. Every line is typed from one page, and a mistyped one would move the ratio the page quotes between operating profit and other income while changing nothing a reader could see. Worded away from the Anant Raj guard on purpose: the ledger requires each documented fragment to identify exactly one guard, so two guards checking the same kind of thing must say it differently.

build fails: does not add up to the profit before tax printed under it

Cash generated from operations, less the income tax paid against it, equals the filed operating cash flow. All three are printed on the same statement and all three are stored, so the identity is asserted rather than assumed.

build fails: less income tax must equal the filed operating cash flow

The movement in current borrowings is presented among the working capital adjustments inside operating activities, and the repayment of borrowings is presented again under finance activities on the next printed page. Both filed years do it. The cash conversion exhibit says the classification is the observation, so the document has to keep doing it: the guard fails if either side stops being an outflow or if the two stop being printed on different pages.

build fails: must be an outflow in both the operating and the finance sections

The data centre subsidiary is printed twice in the same annual report, once in the statement of subsidiaries and once in the consolidated entity table. Its net assets must agree across both, and against its own share capital plus reserves. That agreement is what lets the exhibit call the figure the report's own rather than a subtraction performed here.

build fails: net assets do not agree across the two pages that print them

The same subsidiary's result for the year must agree across the two pages that print it. One is the statement of subsidiaries, the other the consolidated entity table.

build fails: result does not agree across the two pages that print it

The document prints the return on capital formula inside its commissioned industry report and the answers in the business section, and never joins them. Rebuilt from the issuer's own key performance indicators, cash flow statement and balance sheet, every period that can be rebuilt lands on the published figure to the second decimal, and the exhibit says so. The earliest period is deliberately not covered, because an average needs a year the balance sheet does not carry, and that absence is itself the finding.

build fails: the published return on capital no longer reconciles to the balance sheet

The filing prints cash generated from operations, the tax paid against it, and the net figure. All three are stored and the identity is asserted, so a mistyped row is caught rather than quietly changing the size of the funding gap the exhibit draws.

build fails: cash generated from operations less tax paid must equal net cash from operations

Every period the associate exposure is stated for must be a period this file already carries a balance sheet for. The exhibit divides that exposure by net worth at the same date, and a period without one would either throw or, worse, quietly reach for a neighbouring year and print a plausible percentage over the wrong denominator.

build fails: which has no balance sheet in this file

Two audited figures the page sets against each other: the trade receivables and other financial assets the auditor drew attention to as substantially overdue and unprovided, and the capital the company has contracted to spend. The first is several times the second, and the page says so.

build fails: no longer exceed the capital the company has contracted

Financing arranged so that suppliers can be paid later is larger than everything the company has contracted to spend on capital account. Both figures are audited and in the same unit, and the page sets them against each other.

build fails: the supplier finance has fallen below what the company has contracted to build

Twenty cells transcribed by hand from a printed table across two years. A single mistyped bucket would leave every derived share slightly wrong and nothing else would notice, so each year is summed against the total the filing prints.

build fails: against a printed total of

One concept restated in one period is one point. Counting it twice would inflate the rate the page leads on, and the rows are promoted from a raw file where the same period can appear under several concepts.

build fails: counts one restated period of one concept more than once

The three cash conversion inputs are served per period rather than per filing, and often not by the filing recorded against the rest of the row. Each carries its own, and a figure without one could not be cited on a page even though it would render.

build fails: a harvested figure carries no filing to cite it against

A coverage row cannot claim the primary tier without naming a filed document. This is why only one operator in eight carries it.

build fails: claims PRIMARY without naming a filed document

The capacity ladder on this page comes from a research note, and it may never claim the primary tier however often its figures are repeated. The annual report has since been read and sits in its own block with its own printed pages; this guard keeps the two apart, so the note cannot quietly inherit the report's authority.

build fails: no Anant Raj figure is traced to a filing, so it cannot claim PRIMARY

No Netweb filing has been opened either, and the same guard applies.

build fails: no Netweb figure is traced to a filing, so it cannot claim PRIMARY

A coverage row whose own filing has been read and states no power capacity cannot claim a primary source for the megawatts beside it. Reading the document is what makes that checkable, and what it establishes is that the number is not in there, so the figure stays a research estimate however carefully it was read for.

build fails: so the megawatt columns cannot claim a primary source

A capacity forecast is a research house projection, not a figure read out of a filing, and cannot claim the primary tier however widely it is quoted.

build fails: is a forecast and cannot claim PRIMARY

A corporate investment pledge is an announcement, not a figure read out of a filing, and cannot claim the primary tier however widely it is quoted. The same rule the capacity forecasts already carry.

build fails: a pledge is an announcement and cannot claim PRIMARY

A risk row on this name may cite only a printed page the file already records, taken from the annual report block or from one of the source blocks on the audited statements. A page number appearing on a risk and nowhere else would mean a magnitude arrived without the source block that lets a reader check it. The register mixes rows resting on a research note with rows resting on the audited statements, so the rule is enforced per row rather than by banning pages outright.

build fails: a risk row cites a printed page this file does not record

The report is laid out as two page spreads, so one PDF page carries two printed pages and a printed page number can only run to about twice the PDF page count. A citation beyond that is a misread offset rather than a typo, and the offset is the single most likely thing to be wrong when a document is paginated this way.

build fails: a cited printed page falls outside the document it is attributed to

A printed page is what the primary tier means on this site, so the two have to agree. A secondary risk row that attaches a page is borrowing a filing's authority, and a row that cites a page while calling itself secondary understates what it actually rests on. Both directions are rejected.

build fails: a risk citing a printed page must be primary, and one without a page cannot be

A risk row may only cite a printed page this file already cites elsewhere. The register is authored from what was read, so a page number appearing on a risk and nowhere else would mean a figure arrived without the source block that lets a reader check it.

build fails: which this file does not cite anywhere else

On the two pages where no filing has been opened, no risk row may cite a printed page or claim the primary tier. A risk sounds like a fact, so it is the easiest place for a page number to appear unchallenged, and one schema shared by both names means the rule is written once rather than twice.

build fails: no risk row on a page with no filing may cite one

Every company in the disclosure comparison must be measured over the identical set of topics. A rate computed over a different set is a different measurement wearing the same label.

build fails: every company must be measured over the identical family set

Exactly one row in the issuer's chosen peer set is the issuer. Without it the peer exhibits cannot tell which bar to highlight.

build fails: exactly one peer row must be marked self

The revenue mix exhibit draws a quarter against the nine months that contain it and says so out loud. Exactly one row of each, or that sentence describes rows which are not there.

build fails: revenueMix must carry exactly one

Each of the eight operators in the peer benchmarking table carries one reading per fiscal year the table covers, present or explicitly absent. The source prints NA for operators that had not reported, and a short row rather than an explicit absence would shift every value after it into the wrong year.

build fails: does not carry one reading per fiscal year the peer table covers

The four filed columns are not one reporting entity. Two are consolidated and two are standalone, and the document titles all four Restated Consolidated. What makes them comparable anyway is that the associate contributed nothing in the standalone years, so a consolidated statement for those years would have been the same statement. That condition is asserted rather than assumed: a standalone column carrying an associate share would mean every exhibit spanning all four periods is comparing a group against a parent.

build fails: a standalone column carries an associate share

Nobody can have more capacity live than announced. If that inverts, either a row was read wrong or an announcement has quietly been restated downwards.

build fails: MW live against

What has been handed over sits inside what is live. Anant Raj is the whole reason this field exists.

build fails: MW handed over exceeds

Money already spent on an object cannot exceed what the object is estimated to cost.

build fails: money already spent cannot exceed the estimated cost of the object

A single provider's published offer of processors must sit inside the number the scheme has installed nationally. The exhibit says one offer is more than half that national total, and past a hundred per cent the sentence stops being a comparison and starts being a sign that two different quantities have been put in the same ratio.

build fails: offered more processors than the scheme has installed nationally

A 2030 capacity forecast must exceed what is already operational. A projection below today's actual means a row was misread, or a publisher is forecasting a contraction and saying so somewhere this data does not record.

build fails: MW already operational

Where a publisher gives a band rather than a point, its top must sit above its bottom. An inverted band would draw backwards.

build fails: gives a band whose top is not above its bottom

Three firms cannot have pledged more than every investor in the sector has committed in total. The exhibit draws the one inside the other, so past that point the two figures are counting different things and the drawing is meaningless rather than merely wrong.

build fails: pledged more than the sector's whole recorded commitment

The government's estimate of data centre electricity demand must sit above what data centres already draw. A demand curve that does not rise has nowhere to go, and the page's claim that the grid rather than land or capital is the binding constraint rests on it rising steeply.

build fails: does not rise above what data centres draw today

The capital requirement is drawn as a band from the low build cost to the high one. Inverted, it draws backwards and a reader sees a saving where the data says a range.

build fails: the capex per megawatt band inverts

The stabilised colocation margin is quoted inside a published range for the same measure. A stabilising figure sitting outside its own range means two different measurements have been carried in as one, which is exactly the failure this project found in a capacity definition.

build fails: the stabilising margin sits outside the published margin range

Total assets exceed net worth in every filed period, which is an identity rather than a finding. It is asserted because total assets were read out of one column of a four column restated statement, and a column taken one place across would move the accrual ratio without looking wrong on the page.

build fails: total assets do not exceed net worth, so a column has been read across

The board resolution adopting the materiality policy is dated before the offer document that publishes it. That is the ordinary order, and it is asserted because the page states the sequence and a mistyped date would leave the sentence describing an impossibility.

build fails: the materiality policy is dated after the document that publishes it

Consolidation begins and does not stop, so the most recent column is the consolidated one. An inversion would mean the columns have been read in the wrong order, which is the error most likely to go unnoticed on a statement whose four columns run most recent first while the stored rows ascend.

build fails: the most recent period is no longer the consolidated one

A printed page on a two page spread is half a PDF page and has to say which half. On a single page layout there are no halves to name. An anchor that names one where none exists, or omits one where it matters, cannot be checked against the document.

build fails: an anchor names a half on a single page layout, or omits one on a spread

A spread has to say which half of which PDF page carries printed page one, because the two halves are consecutive positions and starting on the wrong one shifts every citation in the document by one page. A single page layout has no half to declare.

build fails: must say which half carries printed page one

The filing gives two payment ranges and the whole point of the arrangement is that one is longer. The slowest comparable payable must still settle sooner than the fastest financed one, or the two ranges have been read into the wrong columns.

build fails: the arrangement no longer buys more time than the payables it replaced

A filled cell has to name a subject that exists and may only be filled once for a pillar. Both mistakes would inflate the count of measured cells the page publishes as its own completeness.

build fails: claims a subject that is unknown or counted twice

The brief names six pillars and the page counts them. A duplicate identifier would leave one pillar unrepresented while the count still read six.

build fails: the six pillars are no longer six distinct ones

A filer cannot restate more annual points than were ever taken from it. The page publishes a rate over that denominator, so a count larger than the harvest would produce a percentage above a hundred and nothing else would notice.

build fails: restates more points than were ever harvested from it

A forward claim that has been graded must carry the evidence that graded it. An ungraded one stays open and says so.

build fails: a graded claim must carry the evidence that graded it

A claim's horizon must fall after the date the claim was made, or it was never a forecast.

build fails: horizon_end must fall after made_on

A refusal recorded as published elsewhere must name where. Refused on the call but published in a filing is ordinary investor relations; refused and published nowhere is a disclosure gap, and only the second is a finding.

build fails: published_elsewhere requires naming where it was published

Every primary claim on this site names a printed page, so the mapping from a printed page to a position in the PDF carries the whole citation chain. Each document stores folios it prints on itself, recovered by pipeline/find_folios.py, and this recomputes every one of them from the declared mapping. Editing the mapping without the evidence fails the build. Before this existed the mapping was a sentence, and two documents wrote the same field name with opposite sign conventions.

build fails: the declared pagination does not reproduce a folio the document prints

Front matter is frequently numbered on its own terms, so a mapping is declared to hold only from a stated index onward. An anchor taken from before that point would appear to confirm a rule it is not governed by, which is the most flattering way to be wrong about pagination.

build fails: an anchor sits before the index the mapping is declared to hold from

One anchor cannot disagree with anything, so a mapping resting on a single folio is an assertion wearing the shape of a measurement. Two distinct positions are the minimum that can contradict each other.

build fails: a pagination needs at least two distinct anchors to be checkable

A known limit marked closed must name what closed it, and an open one must not pretend to. Closing a limit is a claim like any other, and without this a limits list decays into a list of things that simply stopped being mentioned.

build fails: a closed limit must name what closed it, and an open one must not

The worst cell of the matrix, high severity against high likelihood, is the one a reader looks at first, and it cannot hold a row whose magnitude is not derived from data. The standard failure of a risk register is asserting a catastrophe with nothing behind it, and this refuses to render one.

build fails: the worst cell of the matrix cannot hold an unmeasured risk

The same rule inside the measured register: a refusal that points somewhere must say where it points.

build fails: a refusal that points somewhere must say where

The four states are checked against the coverage they claim rather than being labels anyone can type: built covers most of the subjects it could, terminal covers none, and the two in between cover some but not most. Without this the picture and the badge on it could drift apart silently.

build fails: so the label no longer describes the coverage