Methodology

Where commitments come from

We read press releases that U.S.-listed companies file with the Securities and Exchange Commission: U.S. companies file them as exhibits to Form 8-K, and non-U.S. companies listed in the U.S. (foreign private issuers) file them on Form 6-K. These are public documents the company itself chose to publish. Each entry links to the original document, to the SEC filing index, and where possible to an archived snapshot at the Internet Archive so the wording remains available if the page changes.

What we record

We record a commitment only when all of these are true:

• It is the company’s own future deliverable that an outsider can later verify, such as a product launch, a facility opening, a project completion, the closing of an announced transaction, or a regulatory filing.
• It has an explicit calendar time frame, such as a quarter, half-year, month, year or date.
• It is stated as a plan or expectation, not as something already done.

We do not record financial guidance (revenue, earnings, margins, cash flow, dividends or buybacks), long-term aspirations without a date, or targets for other parties.

The scope screen

Every published entry is screened once against these rules by a separate check, and again whenever the rules change. An entry that is in fact a financial forecast, too vague for an outsider to verify, a target for another party, an internal accounting step, or a repeat of another entry is withdrawn from publication. Withdrawn entries are not deleted: each keeps its record and the reason, and appears in the public changes log. This is why the number of commitments on record can fall on a day when the record became more accurate.

How deadlines are set

We use the last day of the period the company named. “In the second quarter of 2026” becomes 30 June 2026; “by the end of 2026” becomes 31 December 2026. The company’s own wording is shown next to the date on every entry. If a company states a fiscal period without saying when its fiscal year ends, we leave that commitment out rather than guess.

When we check

Nothing is assessed before its deadline. Checks begin 30 days after the deadline, to give time for announcements to appear. If the first check finds nothing clear, we check again every 30 days, up to 4 times.

How outcomes are decided

Met

Public evidence shows the deliverable happened on or before the deadline. Anything within the stated period counts.

Late

Public evidence shows it happened, but after the deadline. We show the date and how many days late.

Not met

Positive public evidence shows it had not happened by the deadline, for example the company announcing a delay or a revised later date. (A company that cancels a commitment outright is recorded as “Withdrawn”, below.) We re-check these entries later and change the result to “Late” if the deliverable is completed.

No record found

After repeated checks we found no clear public evidence either way. These entries are not counted as met or not met.

Absence of evidence is never treated as failure. A commitment is only marked “Not met” when there is a public source showing it was not delivered on time.

We prefer primary sources, such as the company’s own press releases, SEC filings and regulator records, and then reputable news reporting. Every result lists the sources it relies on.

Permanent identifiers

Every commitment receives a permanent identifier of the form KI-C-2026-NVDA-000417: the year it was announced, the company’s ticker, and a number that is never reused. The identifier is assigned once and never recalculated, so correcting a typo, updating an outcome or recording a revised deadline can never change it.

Each identifier has a permanent address at keptindex.org/c/ followed by the identifier. Anyone citing a commitment, in an article, a report or a filing, can rely on that address continuing to point at the same record.

Outcomes

Once a deadline has passed and been checked, a commitment receives one of these outcomes:

• Met. Delivered by the deadline the company set.
• Late. Delivered, but after the deadline.
• Partly met. Only part of a stated target was delivered by the deadline, for example 8 of 12 stores.
• Not met. Positive evidence that it had not been delivered by the deadline.
• Withdrawn. The company itself cancelled, terminated or abandoned the commitment.
• Superseded. The company replaced it with a genuinely different commitment, which we then track in its place. A later date for the same commitment is not a replacement; it is recorded as a new version.
• Condition not met. The commitment depended on a stated condition outside the company’s control, such as a regulator’s approval, and that condition failed.
• No record found. We could not find public evidence either way. This is never treated as failure.

An entry may also be marked disputed while we examine a challenge to it. Its outcome stays visible, labelled as disputed, and it is excluded from any score until the challenge is resolved.

Partly met, withdrawn, superseded and condition not met are each factual claims about what the company did or announced, so none is published unless it rests on a grade A or B source; otherwise a person reviews it first.

How outcomes count in the Kept Score. Met counts as delivered on time. Late, partly met, not met and withdrawn count against the company: the score measures whether the date held, and counting withdrawal as a miss means a company cannot avoid one by cancelling just before a deadline. Superseded, condition not met and no record found are left out, because none says anything fair about whether the company delivered what it said.

Evidence and grades

An AI summary is never evidence; the underlying source is. Every outcome lists the sources behind it, and each source is graded by what kind of record it is:

• Grade A. Regulatory filings, such as those filed with the U.S. Securities and Exchange Commission, and government or regulator records.
• Grade B. Stock-exchange disclosures, and the company’s own website, results and press releases.
• Grade C. Established independent news organisations and specialist trade publications.
• Grade D. Any other source.

Grades are assigned by fixed rules from the source’s address, so the same source always receives the same grade. Where a source quotes the passage that supports our finding, we fetch the page and confirm the passage is really there, and record the date it was retrieved and a fingerprint of its content so any later change to the page can be detected.

Two rules follow. We do not publish that a company failed to deliver unless at least one source is grade A or B; news reports alone send the finding to a person first. And if the passages we were given cannot be found on the pages cited, the finding is held for review rather than published.

Verification: three checks and a person

AI models read filings faster than any team could, and they make confident mistakes. So no single model decides alone.

1. First check. An AI model reads the company's later 10-Q, 10-K and 8-K filings and records the outcome with the exact passage it relied on. The passage must be found on the page cited, or the outcome is held for a person.
2. Second check, blind. A model from a different provider receives the same evidence documents, including the passage the first check relied on (re-read from the document itself, with its surrounding text), but not the recorded outcome, and decides independently. Using two providers means one model's blind spots are not repeated by the other.
3. Tie-break. Where the two disagree, a stronger model reads the evidence and both answers, shown in random order, without being told which is on record.
4. Third check on negative findings. “Not met” and “Withdrawn” are the findings that could harm a company if they were wrong, so two agreeing checks are not enough. A third, stronger model re-reads the evidence fully blind, with no earlier answer at all. Only three independent agreeing checks on grade A or B evidence confirm a negative finding.
5. A person decides only what the machines cannot settle: disagreements the tie-break cannot resolve, negative findings where the third check disagrees, outcomes resting on weak evidence, and a random sample of agreements that proves the method keeps working. The sample starts at 5% and falls to 2% once people have upheld at least 98% of 200 sampled outcomes; it rises to 10% if accuracy drops below 95%.

An outcome is confirmed when two independent checks agree on a positive outcome on grade A or B evidence; when the tie-break agrees with a positive recorded outcome on grade A or B evidence (two of three checks); when three independent checks agree on a negative finding on grade A or B evidence; or when a person confirms it. Only confirmed outcomes count in a Kept Score. Every entry shows the checks it went through, under “How this entry was checked”.

We publish the agreement rate between independent checks above and track it weekly. A fall in agreement means a model, a prompt or a source changed, and publication of new outcomes is reviewed before it continues.

Advance notice to companies

Before an outcome that counts against a company is published, or when older findings are first recorded, the company's investor-relations contact (taken from its own SEC-filed press releases) receives a notice listing each finding, the evidence, and how to respond: a statement under the right of reply or a correction request. When notices are sent automatically, new negative findings are published only after a reply window of five business days.

How evidence is kept

• Primary first. Commitments come only from documents filed with the SEC, which keeps them permanently and publicly at a fixed address. Outcomes are settled from later SEC filings wherever possible.
• Exact words. The company's sentence is stored verbatim and matched against the filing. Outcome passages are checked against the page they cite.
• Fingerprints and dates. Each source records when it was retrieved and, where fetched, a fingerprint of its content, so a later edit to the page can be detected.
• Archived copies. Pages that are not SEC filings are sent to the Internet Archive so that what was said survives if the page is changed or removed.
• Nothing overwritten. Restatements become new versions; corrections keep the earlier value and the reason; withdrawn entries keep their history. The full log of changes is public and downloadable.
• Permanent identifiers. Every commitment has a Kept ID and address that never change, so a citation made today still resolves in ten years.

Lineage and drift

Companies often restate a commitment before its deadline: a later date, a smaller target, or softer wording. We never overwrite the original. Each restatement is recorded as a new version, with its own source, and the outcome is always measured against the original commitment.

From the versions we measure three kinds of change. Deadline drift is the number of days between the original deadline and the latest one. Scope drift is the percentage change in a stated target, compared only when both versions state it in the same unit. Language drift records whether the company’s wording weakened, from firm language such as “will” to expectation such as “expects” to aspiration such as “aims”.

Drift is recorded as it happens, before any outcome is known. It describes what the company itself said at different times and is not a judgement of its conduct.

Event type and indicative materiality

Each commitment is labelled with the kind of milestone it names, such as a Phase 3 data readout, a regulatory submission, a deal closing or the start of production, using fixed word rules applied to the commitment's wording. From the event type we give an indicative materiality: transformational (separations and mergers), high (late-stage trial results, marketing applications, deal closings, production starts, facility openings), medium (earlier trial milestones, launches, construction and deliveries) or low (roll-outs and other items). These labels help readers filter; they do not take company size into account and they are never used in the Kept Score. Rules version 0.1.

The Kept Score

Proposed next version: KeptScore Methodology 2.0, public preview, open for comment until 15 December 2026.

Not yet published. The definition below is fixed and published in advance, but no scores appear on the site until the verified historical record is deep enough for a score to be meaningful. Publishing a score from a handful of entries would be worse than publishing none.

The Kept Score is the share of a company’s commitments that were delivered by the deadline it set itself, expressed as a number out of 100.

It is calculated as follows.

• What counts. Only commitments with a clear outcome. Met counts as delivered on time. Late, Partly met, Not met and Withdrawn count as not delivered on time. Superseded, Condition not met, No record found and any entry under dispute are excluded entirely, so silence never counts for or against a company.
• The window. Only commitments whose deadline fell within the last 3 years, so a company is judged on its recent record rather than its distant past.
• The minimum. No score is shown until at least 15 commitments in that window have a clear outcome, so a single entry cannot define a company.
• The calculation. The number of commitments marked Met, divided by all commitments in the window that count (Met, Late, Partly met, Not met and Withdrawn), multiplied by 100 and rounded to the nearest whole number.
• Confirmed outcomes only. Each outcome is established from SEC filings and other public sources and shown on its entry at once. A second, independent AI model then re-reads the evidence without being told the recorded outcome. An outcome counts in a score only when it is confirmed, as defined under verification: agreeing independent checks on grade A or B evidence (three for a negative finding), or a person. The two levels are labelled separately on every entry and in the dataset (confirmation_method). Disagreements the tie-break cannot settle, negative findings the third check disputes, outcomes resting on weaker evidence, and an adaptive random sample of agreements go to a person.
• Late is not partial credit. A commitment delivered after its deadline, or only partly delivered, counts the same as one not delivered at all. The score measures whether the date held for the full commitment, not whether the work happened eventually.

Scores are described in bands for readability: 90 and above Very strong, 75 to 89 Strong, 60 to 74 Mixed, 40 to 59 Weak, below 40 Very weak. The band is a label for the number, not a separate judgement.

A Kept Score is a factual summary of a published record, not a rating, a forecast or an assessment of a company’s quality. A low score may reflect ambitious public targets, regulatory delay or factors outside the company’s control. Every commitment behind a score is listed on the company’s page with its sources, so any reader can check the arithmetic.

What the Kept Score is not. It is not a credit rating, and it says nothing about solvency, creditworthiness, the integrity of management or the value of any security. It is not an investment recommendation and it does not predict future performance. It measures one thing: how often a company's own dated public commitments were delivered by the date it set.

Independence. No company, subscriber or data customer can pay to add, change, suppress or rank an entry, an outcome or a score. Companies can respond to any entry through the corrections process or publish a statement under the right of reply, both of which are open to everyone on the same terms.

If this definition changes, the change and its date are recorded publicly so that a score quoted today can be understood later. When an outcome changes, the earlier outcome and the reason are kept in the entry's correction history.

Methodology version 1.3, 18 October 2026. Adds the tie-break: where the first two checks disagree, a third, stronger model decides blind; it can confirm only a positive recorded outcome on grade A or B evidence. Negative findings (Not met, Withdrawn) now need three independent agreeing checks, or a person; findings confirmed earlier by two checks are re-audited. Adds an adaptive spot-check, advance notice to companies, the scope screen, the changes log and the right of reply. No score had been published under any earlier version.

Methodology version 1.2, 24 September 2026. Scores count only confirmed outcomes (two independent checks agreeing on primary evidence, or a person); changed outcomes keep their history. No score had been published under any earlier version.

Version 1.1, 24 September 2026. This section now lists every outcome and how it counts, matching the outcome rules above; version 1.0 named only Met, Late and Not met. No score had been published under either version.

Who the data serves

Who	The question it answers	How they use it
Investors and analysts	When this company sets a date, does it keep it?	Weigh the next announcement against the company's own record; follow warning signals before a deadline.
Lenders and insurers	Will the project we are financing or covering be delivered on time?	A sourced delivery history as an input to credit and underwriting.
Law firms and courts	What exactly did the company say, when, and what happened?	Citable timelines of statements and outcomes with permanent IDs.
Companies themselves	How realistic are our own dates, and how do we compare?	Claim their page, reply to entries, and manage commitments with Kept Desk.
Governments and regulators	Which departments and contractors deliver, and which slip?	Delivery dashboards through the Public Works Ledger.
Universities and students	Do organisations do what they say? Why do some slip?	Free, citable data for research and teaching: the Open Research Dataset.
Journalists and the public	Has this promise been made before, and was it kept?	Check a claim against the record before repeating it.

Why the record is free to read, and free for research

A measure becomes a standard only when everyone can see how it is made and test it. So the record, every source and this methodology are free to read; the settled outcomes are free to download for non-commercial research and teaching; and anyone can challenge an entry. Universities get a clean, labelled dataset on a question no public dataset answered before, students get real research material and paid verification work, and every published study tests the method in public. Paid products add speed, coverage and tools (alerts, the forward calendar, exports, reports and licensed data), never different facts.

What the record does and does not say

Each entry records what a company said and what the public record later showed. It does not assess why a deadline was missed. Delays often have causes outside a company’s control, such as regulatory reviews, supply problems or market conditions, and a missed date is not in itself a sign of bad faith. Nothing on this site is investment advice.

Automation and review

Commitments are identified and outcomes are checked with the help of AI models and automated rules. To limit errors, extraction must meet a confidence threshold, deadlines that do not match the company’s stated wording are held for human review, and “Not met” results below a high confidence threshold are reviewed by a person before they are published. Automated systems can still make mistakes, which is why every entry shows its sources and anyone can request a correction.

Independence

KeptIndex is not affiliated with any company it covers. We do not accept payment to add, change, remove or rank entries.