AI uses for records
Collaboration intent dataset: four states between an idea and a yes
By SourceX Editorial · Updated
9 sections
Short answer
A collaboration intent dataset should label a statement a commitment only when a record shows an authorized state change and an owner, not when the room sounds agreed. Ethereum's upgrade process shows the states: of 58 proposals on the draft Hegotá list at commit d4c5755, 2 are scheduled and 26 declined. In 210 machine-written call summaries, 82 of 493 recorded decisions use deferral wording.
Agent run
Collaboration intent dataset
Key takeaways
- Label a statement as a commitment only when it links to a ratified state change, a decision entry that settles the substance, or an owned work item.
- Ethereum's EIP-7723 separates proposed, considered, scheduled and declined; on the draft Hegotá list 2 of 58 proposals are scheduled and 26 declined.
- Decision fields hold postponements: 82 of 493 decisions in 210 machine-written 2026 call summaries contain deferral wording.
- Owners are thinner than they look: 183 of 508 summary action items name a collective such as all client teams.
- Ask suppliers for the joins between conversations, decisions and work items, plus state definitions and any summarizer prompts.
What separates an idea raised from a commitment made?
A statement becomes a commitment when a record shows that the people with authority moved it into an accepted state and that someone now owns the next step. Agreement in the room supplies neither, so a collaboration intent dataset has to take its labels from the records around a conversation: state changes, decision entries, linked work and owner fields.
The difference matters because language models read tone well and authority badly. A thread full of 'love this' and 'let's do it' looks like a promise; the ticket that was never created, or the status that never changed, says otherwise. Train on the thread alone and a model learns that enthusiasm is a yes. Train on the thread joined to its records and it can learn which statements actually bound anyone.
Few companies write the intermediate states down, which is why this article uses a public process that does. Ethereum's core developers move proposed changes for each network upgrade through named inclusion stages, record their calls in a public repository and, since February 2026, publish structured call summaries with separate fields for highlights, decisions and owned action items. Each of those records can be counted, and each maps onto a record that enterprise tools already keep.
How does Ethereum record the states between a proposal and a yes?
Ethereum separates four states for a proposed protocol change, defined in EIP-7723, Network Upgrade Inclusion Stages: Proposed for Inclusion, Considered for Inclusion, Scheduled for Inclusion and Declined for Inclusion. Only the scheduled state signals that the change is expected to ship. The definitions are precise about what each state does not mean.
Considered for Inclusion is described as 'similar to "concept ACK" in other open source projects, and is not sufficient to result in deployment to mainnet.' Scheduled for Inclusion requires core developers to 'agree upon a strong intent to include the EIP in the next network upgrade' and to agree that specifications and implementations are mature, and the EIP says status changes are ratified on the execution or consensus layer All Core Devs call before the status is assigned. Ownership starts at the first rung: the proposer should serve as the primary point of contact for the upgrade cycle or designate someone who will.
The Hegotá upgrade list shows how far apart the states sit. Its Meta EIP, still a draft at EIPs commit d4c5755 (27 October 2026), lists 58 proposals: 10 proposed, 20 considered, 2 scheduled and 26 declined. On that snapshot, 3.4% of what was raised for the upgrade carries the state that signals intent to ship, and 44.8% carries an explicit no.
The same process also deletes its own record of ideas. EIP-7723 says that when an upgrade's Meta EIP moves to Review, the proposed and declined lists 'SHOULD be removed'. The Glamsterdam Meta EIP, now in Review, lists 25 EIPs, 18 scheduled for inclusion and 7 networking or informational EIPs under a separate heading, and the Final Meta EIPs for Fusaka and Pectra list only the 13 and 12 EIPs they included. Anyone who wants the rejected ideas has to read the file's history, which is exactly where a dataset builder should look.
| State in EIP-7723 | What the record commits to | Hegotá count (draft list) | Label for a dataset |
|---|---|---|---|
| Proposed for Inclusion | Someone asked; reviewers should look; the proposer is the contact | 10 | Idea with an owner |
| Considered for Inclusion | Client teams intend to try it in devnets; not enough to ship | 20 | Interest, not commitment |
| Scheduled for Inclusion | Agreed strong intent to include, plus maturity, ratified on a call | 2 | Commitment |
| Declined for Inclusion | Client teams are against including it in this upgrade | 26 | Recorded refusal |
What did older meeting notes file under Decisions?
Older All Core Devs notes show what happens while the states are still loose: the word Decisions becomes a heading, and the heading collects everything. Of 345 notes files in the repository's execution and consensus layer meeting folders at commit bf338dc, 97 contain a heading that names decisions, 82 one that names actions, and 235 have neither; headings that also contain the word review, such as 'Review previous decisions made and action items', count for neither.
Inside those headings, ideas and commitments sit side by side. The notes for 7 February 2020 list five decisions: one says an EIP 'will not be included', and four say 'Discussion to be continued in EthMagician thread.' The table headed 'Decisions Made / Action items' for 20 August 2021 opens with 'It would be good to get feedback from Geth focus group' and, two rows later, records that a 64 bit integer 'has been chosen to be the nonce limit'. A wish, a request and a binding technical choice share one column.
The notes also show the community inventing intermediate states as it went. In August 2019 a decision line says five EIPs 'are tentatively accepted and will require benchmarks'; in November 2019 the notes record a status called Blessed being renamed Eligible for Inclusion. Organizations add a middle rung because a single 'accepted' label keeps swallowing things nobody has agreed to deliver. A dataset that collapses those rungs back into one label throws away the very distinction it is meant to teach.
What 210 machine-written call summaries counted as decisions and owners
Since February 2026 the repository's meeting bot has stored a structured summary for nearly every recorded call, 210 of the 211 calls in its manifest, with separate lists for highlights, decisions, action items with an owner field, and target dates. The 210 summaries at commit bf338dc, dated 2 February to 27 October 2026, record 493 decisions and 508 action items. They are machine-written: the pipeline code says it uses the Claude API with the transcript, chat and agenda, so they show how one production prompt draws the line rather than what any person decided.
The prompt's own rule is strict: 'Only formal decisions with clear consensus', with 'Proposals or suggestions' excluded. Even so, 82 of the 493 recorded decisions (16.6%) contain deferral wording: the words deferred, defer, postponed, postpone, pending, revisit, TBD or undecided, or the phrases not yet, next call, next ACDC, next ACDE, next ACDT or no decision; on the 70 main All Core Devs calls the share is 47 of 252 (18.7%). A decision to wait is a real decision, but it commits no one to the substance, and a model trained on the field without that distinction will count postponements as yeses.
Owners are thinner than the field suggests. Of the 508 action items, 183 (36.0%) name a collective owner such as 'All client teams', 101 (19.9%) name several parties and 224 (44.1%) name one person or one team. On the main calls, 87 of 176 actions (49.4%) go to a collective. Part of that reflects the prompt, which asks for cross-team coordination items, but the result is the same for a dataset: an action owned by everyone has no accountable owner until a work item says otherwise.
One more detail shows why labels must come from the governing document. The summarizer prompt expands PFI as 'Pending for Inclusion' and DFI as 'Deferred for Inclusion', while EIP-7723 defines them as Proposed and Declined, and it tells the model that during a proposal window 'the act of presenting is the decision.' Twenty-three recorded decisions mention PFI. Read with the prompt's glossary, a declined proposal looks merely deferred.
The commitment evidence ladder for collaboration records
Label every statement by the strongest record that exists for it, not by how it sounds. The ladder below gives each rung the record that proves it, the Ethereum marker counted for this article and the place the same evidence usually sits in business tools. The counts come from the EIPs and pm repositories at the commits named above; the business column describes documented product features, not any company's data, and nothing here measures how often enterprise records reach each rung.
Two rules make the ladder usable. A statement can only climb as high as a record it links to, by ID, URL or a matching title and date. And 'no usable record' is a label in its own right, because a model should learn to say that a discussion never reached a decision rather than guess which way it went.
| Rung | Record that proves it | Ethereum marker (count) | Typical business record |
|---|---|---|---|
| Idea raised | A message, document or agenda item with no state and no owner | Wishes filed under Decisions in older notes (examples above) | Chat thread, brainstorm doc, idea backlog entry |
| Owned proposal | A written proposal with a named contact | Proposed for Inclusion: 10 on the Hegotá list | Idea with an owner field, draft RFC |
| Interest recorded | A state that says a group will try it, explicitly short of delivery | Considered for Inclusion: 20 | Discovery idea marked in progress, concept approval |
| Decision deferred | A decision entry that postpones the substance | 82 of 493 summary decisions with deferral wording | Meeting decision 'revisit next quarter' |
| Commitment | A ratified state change by the body with authority | Scheduled for Inclusion: 2 | Status change to committed, signed minutes, approved change request |
| Owned delivery | A linked work item with a single accountable owner and date | 224 of 508 summary actions name one person or team | Delivery ticket with assignee and due date |
| Recorded refusal | A state or entry that says no, ideally with a reason | Declined for Inclusion: 26 | Closed as won't do, rejected decision |
| No usable record | Discussion with no state change, entry or link | 235 of 345 notes files with no decision or action heading | Meetings with no notes, notes without outcomes |
Where business tools keep the same line
Business tools draw the same line through links and fields rather than words, so the labels have to come from those joins. Jira Product Discovery keeps ideas apart from delivery work: from an idea's Delivery panel a team can create a new Jira work item or link an existing one, according to Atlassian's documentation. The link, not the idea's description, is the evidence that someone turned interest into planned work.
Confluence covers the owner rung. An action item is assigned by mentioning a person, it then appears in that person's task list, and Atlassian's help says a task with a date turns red once it is past due. That gives a dataset three separate facts to keep: that an action was written down, who it was assigned to, and whether it was ticked off.
A reaction or a plus-one reply carries no state at all, so on its own it can never climb above the idea rung, however many people add one. The practical labeling rule is the same in every tool: a statement in a thread becomes a commitment candidate only when it links to a record on the commitment rung or above, and a collective owner such as 'the team' counts as no owner until a work item names one.
How should a buyer test a model on collaboration intent?
Test a model on the hard cases these records expose: deferrals filed as decisions, interest states that read like approval, collective owners and discussions that end without any record. A model evaluated only on clean, cheerful threads that did lead to tickets will learn that agreement means commitment and will look accurate while doing it.
At a glance
How should a buyer test a model on collaboration intent
A test set built from collaboration records can cover those cases with the checks below:
- Rung labeling: give the model the discussion only and ask for each candidate statement's rung; score it against labels derived from the linked records, not from annotators reading tone.
- Grounding: require the model to cite the record behind any commitment label, such as a state change, a decision entry or an assigned work item; an uncited commitment counts as wrong even when the label matches.
- Deferral handling: include decision entries that postpone the question, and score a model that calls them commitments as wrong.
- Owner resolution: ask who owns each action; 'all teams' or an unassigned item should come back as no accountable owner, not as a guessed name.
- Abstention: for discussions with no usable record, the correct answer is that no outcome was recorded, and a confident yes or no counts as an error.
- Glossary drift: when labels come from a summarizer or a team's shorthand, test that the model applies the governing definitions, as with PFI meaning Proposed for Inclusion in EIP-7723.
- Report precision on the commitment label first: a false commitment sends work to someone who never agreed to do it, which costs more than a missed one.
Could a protocol community mislead an enterprise model?
It could, and the direction of the error is predictable: Ethereum's process is more explicit than most companies' meeting culture, so the gaps measured here are likely smaller than in a typical enterprise archive. Public governance also writes for an outside audience, while internal threads assume shared context and leave more unsaid.
The strongest alternative explanation for the numbers is that the summaries, not the meetings, create the noise: deferrals and collective owners may be artifacts of a prompt that asks for cross-team items and errs on the side of including decisions. That is a fair reading, and it is also the lesson for buyers, because enterprise archives increasingly contain AI-written meeting summaries too. Any label drawn from a summary needs the summary's instructions beside it.
The counts have limits worth stating. The heading census measures note structure, not whether decisions were made. The deferral and owner rules are keyword rules, published with the scripts; a decision that commits to one thing and defers another counts as deferral wording, and a team name and a person's name both count as a single owner. The Hegotá list is a draft and will change. None of these figures says anything about how decisions were later revised, which needs its own evidence.
What to ask for when licensing collaboration records
Ask for the joins, not just the conversations: meeting notes with their decision entries, the tickets or work items those entries link to, the owner and due-date fields, and the status history that shows a proposal moving between states or being declined. Without the joins, a supplier can deliver thousands of threads and still not tell you which statements committed anyone to anything.
Ask also for the definitions: the team's own state names, any summarizer prompts used to produce meeting notes, and the date each definition changed. Owner fields should be pseudonymized consistently, since the useful signal is whether a single accountable owner exists and whether the same owner closed the item, not who that person is.
SourceX licenses business data directly from companies and sources each dataset on request, so a buyer can describe the record types, the link keys between them and an acceptance test, such as the share of commitment labels backed by a linked state change, on the request-data form. Nothing changes hands until the company supplying the records has approved the scope and terms and signed the agreement.
Frequently asked questions
What is a collaboration intent dataset?
A collaboration intent dataset labels statements from meetings, threads and documents by what they committed anyone to: an idea, an owned proposal, recorded interest, a deferral, a commitment, an owned delivery or a refusal. Each label is backed by a record such as a state change or a linked work item, so a model learns to tell interest from obligation.
Can a model learn commitment from words like agreed or plus one?
Not reliably. Even fields labeled as decisions hold postponements: 82 of 493 decisions in Ethereum's 2026 call summaries contain deferral wording, and older notes filed 'Discussion to be continued' under Decisions. Wording is a useful feature, but the label should come from the record that shows a state change.
What if our meeting notes have no decision field?
Label from the joins that do exist: tickets created or linked after the meeting, status changes, assignees and due dates. Where nothing shows an outcome, keep the label 'no usable record' instead of inferring one, so the model learns to say that no decision was recorded.
Are AI-written meeting summaries good enough to label from?
They are a starting point, not ground truth. Ethereum's summarizer prompt expands two stage acronyms differently from the governing EIP and treats presenting a proposal as a decision. Keep the prompt and the governing definitions with the data, and check labels against linked records.
How is this different from a decision log?
A decision log records outcomes after the fact. A collaboration intent dataset labels the statements made before and during a decision, including the ideas that never got one, and uses logs, tickets and notes as the evidence for each label.
Sources
- EIP-7723 defines the four network upgrade inclusion stages, says Considered for Inclusion is not sufficient to ship, sets the conditions for Scheduled for Inclusion and says the proposed and declined lists should be removed when a Meta EIP moves to Review. Source
- The draft Hegotá Meta EIP lists 10 proposed, 20 considered, 2 scheduled and 26 declined EIPs at commit d4c5755. Source
- The Glamsterdam Meta EIP, in Review, lists 25 EIPs; the Final Fusaka and Pectra Meta EIPs list only included EIPs. Source
- Human-written All Core Devs notes show decision headings in 97 of 345 files and mix continued discussions, wishes and binding choices under them. Source
- The ACDbot summarizer prompt defines decisions, action items and stage acronyms for machine-written call summaries generated with the Claude API. Source
- The 210 structured call summaries from February to October 2026 hold 493 decisions and 508 owned action items, counted for this article. Source
- Jira Product Discovery lets a team create or link a Jira work item from an idea's Delivery panel. Source
- Confluence assigns action items by mention, lists them under the assignee's tasks and marks dated tasks red when past due. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.