Proof

Coverage shrinks at every step, and each step is a stricter count than the one before it. The headline is the robust core, not the breadth. 947 names carry a credibility read; the distribution and tier table count the 943 with a settled point-in-time series, so the gap of 4 is names scored but not yet settled.

Coverage is a funnel, not a badge. Most names never reach the robust core — showing the drop-off is the point.

Status: Credibility runs on the analytics host and is currently PAUSED at a Gemini extraction spend cap; the universe sweep (~9.6k transcripts pending) is incomplete. Figures reflect the settled subset only.

Tier by resolved-promise count. A high tier in a low column is a thin-sample read, shown here on purpose — a tier is only as strong as the count beside it.

Concall Credibility measures a management's stated guidance against the results that subsequently landed — a measurement of delivery-versus-guidance, not a judgment of management integrity. A company's credibility level becomes meaningful only once its guidance has resolved. Until then the level is held at a fixed ceiling and reflects only the quantification rate — how specific and falsifiable the guidance was — which is a disclosure metric, not evidence of delivery. Roughly one-third of scored names currently have no resolved promises, and many high bands rest on fewer than three. We therefore define a robust core of names with at least ten graded promises, and report credibility tiers for that core only. Momentum is undefined for the first period and stabilises after about four. Every figure carries its as-of call date and its resolved-promise count; verify each against the original transcript.

Every dataset this site reads, where it comes from, how often it lands, and what its timestamp means. 36 data classes declared, of which 2 are MODELED — we estimated when the market could have known them. Nothing reads from a source that is not on this list.

Archive span: 2004-04-17 → 2026-10-01. On a past date, a company is in the universe when the date falls between its first and last trading day.

• Survivorship is correct only from the archive floor 2004-04-17; delistings BEFORE the archive start are not captured (left-censoring).
• Asymmetric: the price/survivorship spine keeps delisted names, but the fundamentals universe is the ~listed-today set — the two universes differ.
• Universe = NSE mainboard EQ/BE/BZ series; SME (SM/ST) excluded.
• Renames auto-confirmed via shared-ISIN clean handover; ISIN is sparse in the historical bhav feed, so rename resolution is biased toward currently-listed names until the authoritative NSE symbol-change feed is loaded.

Our security universe is built from the raw daily bhav-copy archive, not a current-constituents list. Delisted, suspended and surveillance-series names are kept, so historical analysis can see companies that later disappeared rather than only today's survivors. A company is followed across a symbol rename as one continuous security where an ISIN handover confirms it; demergers, mergers and schemes of arrangement are recorded as continuity-break events so a structural gap is never mistaken for a price move. One asymmetry we state plainly: the price/survivorship spine retains delisted names, while the fundamental history covers the names listed today — the two universes are not identical.

Delisted and inactive names are KEPT, not dropped. A universe rebuilt from today's listed companies would quietly delete every failure, and every number measured over it would be flattered by exactly the companies that did worst.

The rule: a financial period counts as known on its period end +90 days (annual results) or +50 days (quarterly), the same lag for every company. It can be wrong two ways: a late filer is treated as known before it was public — a leak — and an early filer is greyed out after it was already public.

Fundamental history is assigned an availability date using a uniform modelled lag — 90 days after period-end for annual results, 50 days for quarterly — applied identically to every company. This is a modelled approximation of when results became public, not an observed filing date. A company that filed late will appear knowable earlier than it was; one that filed early will appear unknown when it was already public. Accordingly these surfaces are labelled “modeled-availability” and must not be read as point-in-time. A true first-seen date is captured only going forward (it does not exist for historical periods).

The learned lag (+119 days annual, +62 days quarterly) is a conservative lag fitted to real filing dates. 84.2% of archived periods now carry a real exchange filing date, which is used in place of any model. Across every period, the expected leak is 0.95%.

Error is the real filing date minus the modelled date. A positive error is a leak — the model said the market knew something before it could have, which is what would flatter a backtest.

Effective look-ahead leak 0.95% (vs 6.6% on the naive +90/+50d model — a 6.9× cut), over 53,672 matched periods.

WHAT WE MODEL: a fundamental period's report_date is a SYNTHETIC +90d (annual) / +50d (quarterly) lag off period-end — applied identically to every company.

HOW WE DE-MODEL: where a real exchange (BSE) filing date exists we PREFER it (leak 0 by construction); 84.2% of archived periods now carry a real date, and a conservative p95 lag covers the rest. The blended expected leak is 0.95%.

THE PROOF: the effective PIT read uses the real date, so a backtest can never see a datum before it was public. The receipts below show the leak the naive model WOULD have injected — and that the real dates fix it.

Error is the real filing date minus the modelled date. Positive: the model was too early, and a backtest on it would have seen the figure before it was public. Negative: the model was conservative and greyed out a figure that was already public.

Ask about a company a dataset does not cover and the page says so, with the sample size — never a made-up value. Three states are kept apart: not covered (outside that dataset), zero (a real, measured zero) and not in source (the upstream feed has no such field).

1. Survivorship addressed first. Universe built from the raw archive with delisted names retained; a survivorship-correct universe-as-of-date exists.
2. Point-in-time honesty is graduated. Fundamentals are labelled modeled-availability with the exact lag disclosed; Credibility is genuinely PIT by construction. We never call modelled data point-in-time.
3. Sample size travels with every score. No tier or credibility read appears without its resolved-promise count and as-of date; a robust core is defined and tiers reported only within it.
4. No performance claim without a backtest. The product speaks in percentile rank-gaps; the lead-time study is openly marked not-built.
5. Grain is never overstated. Market-level data (participant OI, FII/DII flow) is labelled market-level; we do not claim per-name participant flow.
6. Absence is a first-class value. Coverage gaps return coverage + n, never a fabricated number; “not covered”, “zero” and “not in source” are distinct states.
7. Every figure is sourced and reproducible. Each count maps to a named table/column; each company-level claim links to its origin.
8. Characterisation is banned. Copy states stated-vs-actual facts, never judgments of people.
9. Data provenance is declared. Scraped-today vs licensed-later is stated as a known item, not buried.
10. The as-of clock is global and visible. The whole page renders from one stamped moment; freshness is shown per class.

Battery last ran 2026-10-03 06:35:04 · 32 checks · 6 at warn/critical.

The nightly battery checks the tape end-to-end. A handful of these gate pages inline as banners; this table is the complete record — including the checks that never fire a banner anywhere else.

• No alpha claim. No strategy here is a fundable net-of-cost edge versus the index. The one participation-fundable corner is stated with its capacity, and everything else is descriptive.
• No prediction. Every number describes past tape from primary exchange sources. Nothing forecasts.
• No paywall on evidence. Every number's source, method and failure record is free, and so is the reference that makes a number readable — the percentile, the typical value, the trend. Breadth and convenience may cost; the record never does.
• Failures are published. Approaches we killed stay on the site with their numbers — that is what the validation record is for.

Every number on this page names the table it came from — its source, its grain and the timestamp basis are printed beside it.

The same coverage boundary, provenance registry, universe construction and modelled-vs-filed audit, laid out as one dated document you can print or save as a PDF and attach to a diligence pack.

It is the same live read, not a transcription — the memo is assembled when you open it, and it says so on its face.