AI uses for records
Time-dependent business answers dataset: one request, three right replies
By SourceX Editorial · Updated
10 sections
Short answer
A time-dependent business answers dataset stores, with every historical request, the date it was asked, the customer's entitlement that day and the rule version in force, because the correct reply changes with all three. In the endoflife.date dataset on October 27, 2026, 122 release cycles across 46 products had a support answer that depended on which contract the customer held. Grade each old reply against its own day.
License terms
Time dependent business
Key takeaways
- The correct reply to many customer requests is a function of three inputs: when the request was made, what the customer was entitled to that day, and which rule version applied.
- Across 7,960 dated release cycles in 469 products, 2,063 pass through three or more distinct answers to the same support question during their life.
- On October 27, 2026, the answer for 122 release cycles in 46 products depended on whether the customer held an extended-support entitlement.
- Exports often report today's entitlement status, not the status on the request date, so the snapshot has to come from dated entitlement history.
- Compute every expected answer from stored inputs and re-grade old replies against their own date; a reply that was right then stays right.
Why can one customer request have several correct answers?
Because the correct reply depends on more than the words of the request. It depends on when the customer asked, what the customer was entitled to on that day, and which version of the rule was in force. A dataset that keeps only the request and the reply has thrown away two of the three inputs, so it cannot tell a right answer from a wrong one.
Take one question a help desk receives every day: will my Windows 10 PC still get security updates? Asked in September 2025 about a PC on version 22H2, the answer was yes for everyone. Asked in November 2026, it was no for a home user who had not enrolled in Microsoft's consumer Extended Security Updates program, yes for one who had, and yes for an organization paying for commercial coverage. Same words, three correct replies.
This guide measures how common that pattern is in public support data, shows where entitlement and rule changes come from, and sets out the answer key, data card and change checklist a time-dependent business answers dataset needs before its replies can be used for training or grading.
How often does a support answer change? A count across 486 products
Often enough that a single stored answer is the exception. SourceX counted the support lifecycle dates in endoflife.date, a community-maintained dataset that records release, end-of-active-support, end-of-security-support and extended-support dates for software and cloud products. At commit a3ab802, read on October 27, 2026, it covered 469 products and 7,960 release cycles with both a release date and an end-of-security-support date (486 products have releases).
For each cycle we tracked the answer to one request, will this release get a fix, through its life. Every cycle has at least two answers, yes and then no. A quarter of them, 2,063, have three or more: full fixes, then security fixes only, then fixes only for customers who pay for an extended tier, then none.
The intermediate states last long enough to matter. Where bug-fix and security support end on different dates, the gap had a median of 300.5 days; where an extended tier exists, it ran a median of 760 days past the end of standard security support. Python's developer guide describes the same sequence for its own releases: bug fixes and security fixes after release, security fixes only after two years (18 months for versions before 3.13), and no support five years after release, with dates that release managers can adjust.
The count, with each measure's definition:
| Measure | Release cycles | Products | What it means for a stored answer |
|---|---|---|---|
| Cycles with release and end-of-security dates | 7,960 | 469 | The population counted |
| Three or more distinct answers over the cycle's life | 2,063 (25.9%) | 188 | Yes and no are not enough labels |
| Bug-fix and security answers end on different dates | 1,720 (21.6%) | 150 | The request type changes the answer |
| Extended tier after end of security support | 418 (5.3%) | 56 | The customer's contract changes the answer |
| Inside the extended window on October 27, 2026 | 122 | 46 | Today's answer depends on entitlement |
Same day, different contract: how entitlement splits the reply
On the same date, customers holding different entitlements are owed different answers, and the entitlement is rarely written in the request. Windows 10 shows the structure clearly. Microsoft's support page says Windows 10 support ended on October 14, 2025. Its consumer Extended Security Updates page says enrollment runs until the program ends on October 12, 2027, for devices used personally on version 22H2, excludes devices managed by an organization, and covers only critical and important security updates. The endoflife.date record carries commercial extended support for 22H2 to October 10, 2028.
Editions split answers the same way. The dataset lists Windows 11 version 25H2 with support to October 10, 2028 for the Enterprise, Education and IoT Enterprise editions but to October 12, 2027 for Home, Pro and related editions, so for a full year the same version gets updates for one customer and not for another. Ubuntu 20.04 has three tiers: standard support, which ended in May 2025, Ubuntu Pro coverage to May 2030, and a Legacy add-on beyond that.
For a dataset, the lesson is that a reply cannot be graded without the customer's entitlement at the time of the request. A model trained on replies without that input sees yes and no given to identical questions on the same day, and learns noise, or worse, learns to guess from incidental wording.
When the vendor moves the date after the answer was given
Rules change after replies are written, and a reply that was correct under the rule published at the time stays correct for its date. Extended coverage is defined in yearly steps: Microsoft's lifecycle FAQ lists Extended Security Updates Year 1 as ending October 13, 2026 and Year 2 as ending October 12, 2027, and the consumer page gives the Year 2 date as the end of the personal-use program. Which end date a reply should have quoted depends on which years were on offer when it was written.
Canonical's pages show the same drift. Its blog post on the end of Ubuntu 20.04 standard support says the Legacy add-on extends security maintenance to May 2032, while its current Expanded Security Maintenance page lists Legacy add-on coverage to May 2035. A support agent who quoted 2032 was right on the day, and a grader reading today's page would mark the agent wrong.
So forward-looking answers, the until-when kind, need a fourth input: the date each rule version was published. Present-tense answers, such as whether a fix is available today, depend only on the version in force. Recording both dates per version keeps the two kinds of question from being graded by the same, wrong, standard.
The answer key: compute each expected reply from date, entitlement and rule
Store the inputs and compute the expected answer by a written rule, instead of storing a correct answer typed by someone after the fact. A computed key can be regenerated when a rule changes, checked by a second reviewer and explained to a buyer; a hand-typed key silently freezes one reviewer's reading on one day.
Entitlement snapshots are the input most often lost. Salesforce's documentation says an entitlement's status is determined by the org's current system date and the entitlement's start and end dates, so an export taken today reports today's status, not the status on the day a case was opened. Entitlements there are units of support such as phone or web support, and service contracts cover warranties, subscriptions and service level agreements. The snapshot has to be rebuilt from the dated records, not read from a status column.
Each graded request carries these inputs:
- Request type: what is being asked, from a closed list such as security fix, bug fix, refund, replacement or priority response.
- As-of timestamp: when the request was made, with its time zone.
- Entitlement snapshot: the plan, edition, contract or add-on the customer held at that timestamp, with the record and dates it came from.
- Rule version: the policy, terms or lifecycle table in force, with its effective dates.
- Published date: when that rule version became public, used for until-when answers.
- Expected answer: computed from the inputs above, never typed, with the interval over which it stays valid.
Time-aware data card for entitlement-dependent requests
Add these fields to the dataset's documentation and to each record so a buyer can recompute any expected answer. Values marked illustrative continue the Windows 10 example and describe no real customer.
| Field | What it records | Example | Check before use |
|---|---|---|---|
| Request type | The closed category of question | Will version 22H2 get a security fix? | Every request maps to exactly one type |
| As-of timestamp | When the customer asked | November 2, 2026, 09:14 UTC (illustrative) | Present on every request, time zone stated |
| Entitlement snapshot | Plan, edition or add-on held at that moment, and its source record | Consumer ESU on a personal-use device (illustrative) | Built from dated entitlement history, not a current status field |
| Entitlement validity | Start and end dates of each entitlement record | Enrolled through October 12, 2027 | No gaps or overlaps for one customer and product |
| Rule version | Policy or lifecycle table in force, with effective dates | Consumer ESU terms as published on Microsoft's page | Version table covers the whole request window |
| Published date | When the rule version became public | Date of that page version | Until-when replies graded only against versions published by the as-of time |
| Expected answer and interval | Computed answer and the dates over which it holds | Yes, through October 12, 2027 | Recomputed whenever a rule or entitlement changes |
| Answer-change log | Each event in the period that flipped an expected answer | October 14, 2025: end of standard support | Count of affected requests stored with each event |
Change-impact checklist when a rule or entitlement changes
Run this checklist whenever a policy, a lifecycle date, an entitlement offering or a customer's own entitlement changes, before the next training or evaluation run. Its job is to turn the change into a list of requests whose expected answer moved, without touching any reply already given.
Checklist · 7 items
- Classify the change: a rule (a date or clause), an offering (a new tier or new eligibility) or one customer's entitlement (an upgrade, renewal or lapse).
- Record its effective date and its publication date separately.
- Recompute expected answers for every request whose as-of date falls in the affected interval; never edit stored replies.
- Count requests whose expected answer flipped, by request type and entitlement tier, and log the count against the change.
- Re-grade historical replies only against the expected answer for their own date and entitlement.
- Review evaluation tasks: a task whose answer now depends on an entitlement it does not state is ambiguous and must be rewritten or retired.
- Flag requests that cannot be recomputed because the source system kept only current status, and keep them out of graded sets.
0 of 7 ticked. Tick each item as you go.
Could recent records alone teach today's answers?
Only for tasks that never reason about dates or contracts, and support tasks usually do. Recent records describe current rules, which is useful, but they underrepresent states that appear late in a lifecycle. In the lifecycle count, extended tiers begin only after standard security support ends and then run a median of 760 days, so a one-year sample of recent requests about a product family can contain no example of the extended phase at all, while customers on older releases keep asking.
Newer records can also change the task. If a vendor folds a paid tier into the base plan, answers that used to depend on entitlement become uniform, and scores from before and after are not comparable. Split evaluation series at such changes and report them separately.
A fixed comparison settles the choice for a given task. Train one arm on recent requests only and one on all requests with the five inputs attached, at the same budget, and score both on later requests whose expected answers were computed from the key. Decide by a threshold written down before the run.
Requesting historical requests with their entitlement context
Ask for the entitlement history and the rule versions along with the requests, because a dataset delivered without them cannot be graded or regraded. Name the request types, the systems that hold entitlements, the lifecycle or policy documents with their effective and publication dates, and an acceptance test, for example that every sampled request resolves to exactly one entitlement snapshot and one rule version.
At SourceX the buyer's written description comes first: datasets are found at U.S. businesses to match it, not drawn from inventory, and the form records scope, volume, language, format, timing and licensing terms. Entitlement and rule requirements belong in that description so they shape extraction. The supplying company reviews and approves the scope and terms, and signs, before any records change hands.
Methods and limits of the lifecycle count
Data: products/*.md in github.com/endoflife-date/endoflife.date at commit a3ab802af9b4be24f1d32c4119d99d2b03cbde4a, read on October 27, 2026 from a bare clone. Per release cycle we used releaseDate, eoas (end of active support), eol (end of security support) and eoes (end of extended support), counting only fields that hold dates; boolean or missing values were treated as unknown. A state counts when its interval has positive length; the 1,720 cycles whose bug-fix and security answers end on different dates also need a bug-fix phase of positive length (releaseDate before eoas, eoas before eol). Entitlement-dependent on a date means the date falls on or after eol and before eoes.
endoflife.date is maintained by volunteers who copy vendor pages, so its dates are only as current as its last edit. We spot-checked two products. Windows 10's end of support matches Microsoft's page. Ubuntu 20.04's end of standard support matches Canonical's blog, but the dataset's April 2030 date for extended maintenance differs from the May 2030 on Canonical's maintenance page. The other cycles were not checked, and the dataset holds the current table, not past versions of it.
Not measured: how often real customers ask lifecycle questions, how many historical replies a grader using today's table would mark wrong, and any business support archive. SourceX has published no results on licensed request data; the count shows how common time-and-entitlement dependence is in one public domain.
Frequently asked questions
What is a time-dependent answer in a business dataset?
It is a reply whose correctness depends on when the request was made, what the customer was entitled to then, and which rule version applied. The same question can have different correct replies for different customers on the same day, and for the same customer on different days. A dataset has to store those inputs for the reply to be gradable.
How do you grade an old support reply after the policy changed?
Compute the expected answer from the request's own as-of date, the customer's entitlement on that date and the rule version in force, and grade the reply against that. For until-when answers, use only rule versions published by the as-of date. A reply that was right then stays right, even if today's rule says something else.
Why is a CRM export not enough to reconstruct entitlements?
Many systems report entitlement status relative to today. Salesforce, for example, determines status from the org's current system date and the entitlement's start and end dates, so an export shows whether coverage is active now, not on the day a case was opened. Rebuild the snapshot from dated entitlement records instead.
How common are entitlement-dependent support answers?
In public software lifecycle data they are routine. Across 486 products in endoflife.date, 418 release cycles in 56 products have an extended tier after standard security support ends, and on October 27, 2026, 122 cycles in 46 products were inside that window, where the answer depends on the customer's contract.
Should replies given under an old rule stay in a training set?
Yes, if the date, entitlement and rule version travel with them, because they teach how the answer depends on those inputs. Without the inputs, they teach an outdated rule as if it were current. Records that cannot be resolved to a rule version and entitlement should stay out of graded sets.
Sources
- The endoflife.date dataset of release, support, security and extended-support dates for 486 products, from which SourceX counted answer states per release cycle. Source
- Python's developer guide defines bug-fix, security-only and end-of-life phases for each release, with dates release managers can adjust. Source
- Microsoft states that Windows 10 support ended on October 14, 2025. Source
- Microsoft's consumer ESU page: the program ends October 12, 2027, applies to personal-use devices on 22H2 and covers critical and important security updates. Source
- Microsoft's lifecycle FAQ lists ESU Year 1 ending October 13, 2026 and Year 2 ending October 12, 2027. Source
- Canonical's blog: Ubuntu 20.04 standard support ended in May 2025, and the Legacy add-on extends security maintenance to May 2032. Source
- Canonical's Expanded Security Maintenance page lists Ubuntu Pro support for 20.04 to May 2030 and Legacy add-on coverage to May 2035. Source
- Salesforce determines entitlement status from the org's current system date and the entitlement's start and end dates. Source
- Salesforce entitlements are units of support such as phone or web support; service contracts cover warranties, subscriptions and SLAs. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.