AI uses for records
Decision ownership collaboration dataset: 359 of 618 records lack one owner
By SourceX Editorial · Updated
10 sections
Short answer
A decision ownership collaboration dataset should label who was accountable for each recorded decision, keep shared, team, self-approved and missing owners as distinct states, and train models to escalate rather than guess. In 618 OpenShift enhancement proposals, 359 (58.1 percent) do not name exactly one approver who is not also an author, although the template has asked for a single approver since April 2023.
License terms
Decision ownership
Key takeaways
- Only 259 of 618 OpenShift enhancement records name one approver who is not an author; 282 list two or more.
- A merge-blocking linter removed placeholder approvers from new records; an advisory single-approver note cut multi-approver records from 55.0 to 29.2 percent.
- Approval and ownership are separate unknowns: of 245 records marked implementable or implemented, 33 still list a placeholder approver.
- Grade ownership answers against a map of record states, with the escalation route as the correct answer when the record settles nothing.
- Request approval events and the approval rule with the documents; a metadata block alone rarely says whether several approvers meant all or any.
Who owns a recorded decision, and how often does the record say?
A decision ownership collaboration dataset has to record who was accountable for each decision as the record itself states it, and label every case where the record does not settle the question. That label is the training signal: a model that meets an ambiguous record should say who could resolve it, not pick a likely name from the surrounding text.
The guess is tempting because collaboration records are full of names. A design document lists its authors, its reviewers, the people who commented and the person who merged it. Any of them can look like the owner to a retrieval system, and an assistant asked who approved a change will usually find someone to name. Whether that answer is grounded depends on a field most teams fill in quickly and rarely revisit.
Public design records let us count how often the owner field supports a confident answer. In 618 OpenShift enhancement proposals read at a pinned commit, 359, or 58.1 percent, do not name exactly one approver who is not also an author. The sections below show where that ambiguity comes from, how two process changes moved it, and what a dataset and a buyer's acceptance test should do with it.
Why OpenShift's enhancement proposals make a fair test
OpenShift's enhancements repository is a public decision log with an explicit owner field and a written rule for filling it, which makes it a strict test: ambiguity there happens despite a process designed to prevent it. The project's template describes the proposals as a record of previous conversations and the decisions that were made, kept so later engineers and support staff can understand the history of a design.
Every proposal opens with a metadata block naming authors, reviewers, approvers and, for interface changes, API approvers. The README says each enhancement should have at least one approver, and that the approver is responsible for recognizing when reviewers have reached consensus so a proposal can be approved or formally rejected. The template's comment on the approvers line is narrower: this should be a single approver, because having multiple approvers makes it difficult to determine who is responsible for the actual approval. The developer guide repeats the point, preferring a single approver to prevent confusion over who is ultimately responsible.
Two machine rules sit beside that guidance. Since April 2022 a metadata linter rejects TBD or empty values in the people fields and requires at least one valid GitHub ID, but its script checks only enhancement files that are new in a pull request, so older records were never re-checked. And the status field that says whether a proposal is provisional, implementable or implemented was removed from the template in November 2021 and restored in December 2025.
What 618 enhancement records name as their approver
Fewer than half of the records give a retrieval system one accountable person to name. Reading the metadata block of every Markdown file under the enhancements folder, 618 records qualify; six files without a metadata block were skipped. The approver field falls into six states, and only the first supports a confident one-name answer:
| Approver field state | Records | Share of 618 | What a model can say from the record alone |
|---|---|---|---|
| One named approver who is not an author | 259 | 41.9% | That person was designated to approve; cite the field |
| One named approver who is also an author | 5 | 0.8% | Designated approver, but self-approval; qualify the answer |
| Two or more named approvers | 282 | 45.6% | Several people were designated; no single owner is recorded |
| A team alias, alone or with names | 10 | 1.6% | A group was designated; the individual is unrecorded |
| Placeholder or empty, such as TBD | 60 | 9.7% | No approver was ever recorded |
| No approvers field at all | 2 | 0.3% | The record predates or ignores the field |
Did the single-approver rule change what gets recorded?
Yes, and the enforced rule worked better than the advisory one. Placeholder, empty or missing approvers appear in 60 of the 320 dated records created before the linter arrived on 11 April 2022, 18.8 percent, and in none of the 285 created since. A check that blocks a merge removed the problem entirely for new records.
The single-approver note, added to the template on 7 April 2023, changed behavior without ending the old habit. Before it, 105 of 389 dated records named one approver who was not an author, 27.0 percent, and 214, or 55.0 percent, named two or more. Among the 216 records created on or after that date, 148 have a clear single owner, 68.5 percent, while 63, or 29.2 percent, still list two or more. In 2026 so far, 21 of 54 new records name several approvers. By creation year, the clear share was 19.5, 21.9 and 21.6 percent from 2019 to 2021, 37.8 percent in 2022, and 68.1, 75.0, 71.4 and 59.3 percent from 2023 to 2026.
For a dataset builder the lesson is about provenance as much as ownership: the meaning of an empty or crowded owner field depends on which rules applied when the record was written. A record from 2020 with TBD in the approver field reflects a process that allowed it; the same value in a 2024 record would point to a broken check. Store the process version with each record, or the label cannot be interpreted.
Approved for implementation, owner still unclear
Even the records that say a decision was taken often leave the owner open. Of the 245 records whose status reads implementable or implemented, 58 name one approver who is not an author, 23.7 percent. The other 187 split into 145 with two or more approvers, 33 that still carry a placeholder or empty value, 6 with a team alias, 2 with no approvers field and 1 self-approval. An approved design whose approver field says TBD is the clearest case where a model should not invent an owner.
The decision state itself is often missing. Mostly because the template dropped the status field for four years, 269 records carry no status value; for records written in that gap, the state lives in the tracking ticket the template asked for instead, outside the record. Another 11 records kept the template's unedited list of options, provisional through informational, as their status. A question such as who approved this design and is it final can have two unknowns, and the dataset should keep them as separate labels.
The alternative explanation deserves a fair hearing. Relevant text elsewhere may name the owner: a pull request approval, a meeting note, a later comment. That evidence can be real, but it may be outdated, unauthorized or insufficient, and a model that reaches for it has to show it and say which record it relied on. The census measures only what the decision record states, which is the floor a retrieval system can rely on.
The knowledge authority map: from ownership state to the right answer
Map every ownership state to the behavior a model should show, and grade answers against the map rather than against a single expected name. The map is a labeling scheme: each record gets one state from the table, the gold answer follows from the state, and ambiguous states carry an escalation target taken from the process documents, not from guesswork about who seems senior.
In OpenShift's case the escalation target is written down. The README sends authors who are unsure about approval to the staff engineers, so a correct answer to an unowned decision names that route. In a company archive the equivalent might be a change advisory board, a product council or the owner of the policy that governs the decision. The map, applied to the six states and to missing decision status:
| Record state | Evidence that settles it | Correct model answer | Escalation target |
|---|---|---|---|
| One named approver, not an author | Approver field at the version in force on the question date | Name the approver and cite the field | None needed |
| Approver is also an author | Approver and author fields, plus any separate approval event | Name the person and flag self-approval | Process owner, if the policy forbids self-approval |
| Two or more approvers | An approval rule stating all, any or a quorum | List them; say no single owner is recorded unless a rule exists | The body the process names for arbitration |
| Team alias | Team membership or queue owner on the decision date | Name the team; say the individual is unrecorded | The team's lead or intake queue |
| Placeholder, empty or missing | Nothing in the record; only outside evidence | Say no owner was recorded; do not infer one | The route the process documents name |
| Decision state missing or unedited | Status field, tracking ticket or approval event | Say the state is unknown from the record | The tracking ticket owner |
The strongest objection: several approvers are deliberate
Many decisions genuinely need sign-off from several groups, and a list of approvers can reflect that. The objection is fair, but it changes the label rather than removing the ambiguity: the record still has to say whether all listed approvers had to agree or whether any one of them could approve. OpenShift's own template treats multi-team input as the job of reviewers and API approvers and keeps the approvers line for one accountable person, which is why it calls multiple approvers confusing.
Business workflow tools store the missing rule as data. In Jira Service Management's team-managed approval step, an administrator sets how many of the approvers must approve a request, and a decline by any approver declines it. A request record from such a system can say whether three listed names meant three required approvals or one; a document's metadata block usually cannot. Datasets built from documents should therefore carry an approval-rule field with an explicit unknown value.
Self-approval is the narrower version of the same problem. In 19 records at least one named approver also appears among the authors, five of them as the only approver. Whether that is acceptable depends on the policy, and the policy is exactly the context a model needs before calling the decision approved.
The evidence-grounding test for an ownership answer
An answer about who owns or approved a decision passes only when every check below holds. A buyer can run these checks on a sample of model answers, and a supplier can run them on its own records before sharing them:
- It cites the field or approval event that names the owner, at the version in force on the date the question asks about.
- It never promotes an author, reviewer, commenter or the person who merged the change to owner.
- With several approvers, it states the approval rule when one is recorded, and otherwise says that joint or alternative approval is unknown.
- With a team alias, it names the team and says the accountable individual is not recorded.
- With a placeholder or no field, it says no owner was recorded and names the escalation route the process documents give.
- It checks which process rules applied when the record was created, such as whether placeholders could still be merged.
- It keeps decision state and ownership as separate claims, so an unknown status does not hide behind a named approver.
- The gold label preserves ambiguity: no grader resolves an unowned record from outside knowledge before scoring.
What a buyer should request in decision-ownership records
Ask for decision records together with the system events that show approval, because the document alone usually cannot settle ownership. The most useful bundle pairs each decision with its approval events from the system of record, such as workflow transitions, review approvals or change approvals, the approval rule that applied, the history of the owner field, and the process documents in force on the decision date, with people replaced by stable role identifiers.
Keep the census measures as acceptance criteria. A buyer can ask a supplier to report the share of records in each ownership state, the share with a recorded approval rule, and how many approved decisions lack an owner, using the table above as the format, and can require that ambiguous records stay in the sample with their labels rather than being cleaned out. Those records are what teach a model to escalate.
SourceX licenses business data directly from companies and does not hold datasets in stock; a buyer describes the records it needs and SourceX follows up about potential sources. A request for decision-ownership data can state the fields above, the de-identification required for names and roles, and the grounding test in this article as the acceptance test.
Methods note: commit, parsing rules and what the census leaves out
Data: every .md file under enhancements/ in github.com/openshift/enhancements at commit 2a301e9ef4b44e0f39d1c7d8d0a6241c8953ffcd (27 October 2026), cloned on 28 October 2026 and read by os_approver_census.py, a standard-library Python script with no randomness that runs no repository code. A record is a file whose first line is --- and whose metadata block, up to the next --- line, has a title key; 6 of 624 files were skipped. Fields are read line by line: a top-level key, an inline value after the colon with any # comment removed, and following lines starting with optional spaces and a dash, up to the next top-level key. On the 583 files a YAML parser could read, this matched the parser's counts for approvers, authors, status and creation-date exactly; the reader also handles the 35 files that YAML rejects.
Rules as applied: quotes, including curly quotes, are stripped from each entry. Placeholders are, case-insensitively, empty, TBD, none, null, ~, n/a, -, @..., ... and todo. A team alias is an entry containing @org/team. Every other non-placeholder entry counts as one named approver. An identity is the first @handle in an entry, or else its first word, lowercased; an approver is also an author when identities match. Status is the first status entry, lowercased, and a missing or empty status counts as no status value; implementable and implemented count as decided, and any status containing a vertical bar is the template's unedited list. Creation dates are the first valid calendar date in creation-date (13 missing or invalid). The boundaries are the commit dates of f012c1c (metadata-lint, 11 April 2022) and e257db5 (the single-approver note, 7 April 2023). Percentages are exact fractions rounded once, half up, to one decimal.
Limits: the field records who was designated, not who granted approval; the approval itself happens on the pull request, which the census did not read. Creation dates are self-reported. Identity matching is textual, so a person listed under two spellings would be missed. The authority map, the grounding test and the request list are editorial frameworks drawn up for this page on 28 October 2026. The saved snapshot keeps file paths, status, creation year, categories and counts, with no names or handles.
Frequently asked questions
What is a decision ownership collaboration dataset?
A set of decision records labeled with who was accountable for each decision as the record states it, the state of that ownership field, the approval rule and the process version in force. Its purpose is to teach and test models to name an owner only when the record supports it and to escalate otherwise.
How should an AI assistant answer when a decision record names several approvers?
It should list them, say whether a recorded rule required all or any of them, and otherwise state that the record names no single accountable owner. If the process names an arbitration route, the answer should point to it rather than choosing one approver.
Is the author of a design proposal its decision owner?
Not by default. Authors drive a proposal; approvers decide when it is ready. In the OpenShift census, 19 records list an author among the approvers, a self-approval that a correct answer should flag rather than treat as independent sign-off.
Why not infer the owner from who merged the change?
Merging is a mechanical act that can be done by a bot or by anyone with permission, and it says nothing about who was accountable for the decision. Treat it as weak outside evidence, show it explicitly, and keep the record's own ownership label.
Where do business decision-ownership records live?
In the systems that run approvals: ticket workflows with approval steps, change management tools, pull request reviews and policy registers, alongside the documents that describe the decision. Licensing both together lets a dataset tie a written decision to the event that approved it.
Sources
- OpenShift's enhancements folder at commit 2a301e9 (27 October 2026): 618 enhancement records with metadata blocks, whose approver, author, status and creation-date fields were counted for this article. Source
- The OpenShift enhancement template calls the proposals a record of previous conversations and decisions, and its approvers comment says there should be a single approver because multiple approvers make it difficult to determine who is responsible for the approval. Source
- The repository README says each enhancement should have at least one approver, describes the approver as responsible for recognizing consensus, and sends authors unsure about an approver to the staff engineers. Source
- OpenShift's developer guide says a single approver is generally preferred to prevent confusion over who is ultimately responsible for approval. Source
- The repository's metadata linter reports an error when a people field holds TBD or an empty string, and requires at least one valid GitHub ID. Source
- The metadata lint script checks only enhancement files that are new in the current pull request. Source
- Commit e257db5 (7 April 2023) added the single-approver comment to the approvers line of the enhancement template. Source
- Commit 7e90d19 (8 November 2021) removed the status field from the enhancement template and replaced it with a tracking link field. Source
- Commit 3b141ff (22 December 2025) re-introduced the status field in the enhancement template. Source
- In Jira Service Management's team-managed approval step, an administrator chooses how many of the approvers must approve a request, and a decline by any approver declines it. Source
- SourceX's data request page says it licenses business data directly from companies, follows up about potential sources, and sources datasets on request rather than holding them in stock. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.