The facilitator's process, from planning to handover

This guide runs a study from the empty draft to the handed-over result. It follows the order the work actually happens in, and it says at each point what is on the screen, what the panel sees at the same moment, and which actions cannot be taken back.

The companion guide, How Delphi Pronoia's two interfaces are built, explains why the facilitator view is organised the way it is. The one-sentence version, because everything below follows from it: your phase is not the study's phase. The panelist asks "what is open for me"; you ask "what is my work right now", and while the response phase is open your work is monitoring — and preparing the next round.

0. The map you are working on

Four kinds of work, always the same four buttons:

Planning · Collection · Analysis · Reporting

Ten things that belong to no round and no phase, in the bottom bar marked Always available:

Settings · Responses · Panelists · Theses · Community · Summary · Team & access · Results · Export · Share an observation · Method library

The names are the screen's own, not the guide's: a bottom-bar item is a door, and a door is named after the place it leads to.

The phase decides where you land and what gets weight. It does not decide where you may go: in a Delphi you return to planning in the middle of collection, because round 2's theses are built while round 1 is still running. The button marked default is where the study says you belong; when you are somewhere else, the view tells you so — "You are looking at {a}. The study is at {b}." — with a Back that returns you. It never hides anything; it only refuses to let you believe you are looking at the current round when you are not.

The navigation is at the top of the screen. The four buttons and the tabs of the chosen kind of work sit in one block that stays in place; a choice brings the work directly beneath it. Collection shows the open round — closed rounds are read in Analysis.

1. Planning: a completeness list, not a wizard

Planning is both a sequence (the first time through) and a place (round 3's theses, built mid-study). Pronoia resolves that by showing the first pass as a list of what is missing rather than as numbered steps — Building the study:

"Five things the first round is made of. The order is for reading; readiness is computed from what the study contains."

The five, each of which is also the way in:

Role play and the anonymous one-shot poll are settings, and each has its own working guide: Role play: a role in a box of the panel matrix and One-shot poll at a seminar: QR code and anonymous answers. Role play needs the panel matrix to be built first.

The list is computed, not ticked off. Readiness comes from what the study contains, and the header states the situation without arithmetic on your part: "Opening is waiting on {n} item(s)", or "Everything that blocks opening is in place." Items that are advice rather than obstacles are marked recommended — a thesis without a scale is open-ended, which is information, not a fault.

Two of the gates are enforced by the server, not by the screen. A study does not leave draft while the AI mode is unconfirmed, and invitations do not exist until the data-controller declaration is complete. The gaps are named as they are: "AI mode not confirmed", "controller name", "controller contact", "legal basis for processing", "consent text still has placeholders". The list reads the server's answer instead of guessing it, so you never meet a gate you were not warned about.

And the gap is closed in the view the list leads to. In Settings, under the AI-mode choice, it says which of the two holds: "The AI mode is confirmed" or "The AI mode is not confirmed — a round cannot be opened until it is". Confirm the AI mode opens the same choice dialog that otherwise appears only after you hit the gate. The confirmation applies to the selected mode: change the mode and it has to be confirmed again, because a change does not carry the old confirmation with it.

Every planning view says what it applies to. "Applies to: the study" or "Applies to: round {n}", and the panel row says the hybrid out loud: "Applies to: the study · participation is per round". When you are preparing round 3 this is the difference between editing round 3's theses and quietly editing round 1's. Round- scoped editors take the round as an explicit parameter — never "whatever is current".

Rounds are not fixed at creation. "The number of rounds is not locked at creation: the length of a Delphi can be a result rather than a precondition." Dialogue is on by default for an added round; a pure feedback round is made by switching it off, deliberately:

"This round has NO dialogue phase — it collects positions but the panel never argues about them."

Building the panel. The panel is one list in which AI participants are marked — SHOW: All · People · AI filters the view and changes nothing. Under it, PANEL COVERAGE reports the Coverage score, the primary and secondary coverage, the cells that are empty, and — when they are — "Panel supplement recommended before the round".

Reserve enough rounds — an unused round costs nothing. The number of rounds is set when the study is created, and it is not a promise you have to keep: the program never requires a reserved round to be run. You may close the study after round two even if you reserved four; the rounds you did not open simply never existed for the panel. The reverse is harder — a round can be added mid-process, but that is the exception (see the round-management card), not the path. So when the number is in doubt, reserve the larger one.

2. Round 0: orientation, run as a sequence

Round 0 is the one place where the work genuinely is a sequence, and the view shows it as one, and the steps are named on the screen: Configure, Open round, Send invitations, Monitor chats, Close & go to R1.

• Configure sets the orientation questions, the opening text shown to the panelist, the AI assistant's deepening task and how many follow-ups it may ask. With no configuration yet: "Create default config".
• Who round 0 is for is a choice of its own, made before opening. Whole panel is the default; Selected key informants narrows the round to the ones you tick. The scope covers AI panelists as well as humans, and the same choice drives the AI run, human access to the conversation and the cost estimate. The scope is an explicit choice, not derived from the ticks: one tick must not quietly shut the rest of the panel out.

From then on the scoped number appears everywhere round 0 is discussed — readiness, the R0 counters and the invitation card — beside the panel size. In the collect view the scope has a line of its own, and Who round 0 is for → leads back to the choice.

• Open Round 0 publishes it. From this moment the panelists' links lead somewhere.
• Invitations are generated per panelist — Generate invite links — and sent from the communication centre or by hand. In a blind study the interface says what it cannot do rather than failing quietly: "Blind mode hides the addresses: the links are created, but invitations cannot be mailed by hand."
• Monitor chats gives the state of each session — Started, In progress, Completed, Not opened — and the average number of turns. You can read any session.
• AI panelists' orientation chats are generated here, not from the AI actions control elsewhere: this screen is a lifecycle you run in order, and the two buttons belong to that order. Regeneration deletes the existing AI chats and cannot be undone; the confirmation says so.
• Close Round 0 → open Round 1 ends orientation and starts the study proper.

A study can skip Round 0 — "Advance straight to Round 1. Sensible e.g. for a pure AI panel."

Round 0 belongs to every variation. It comes before the panel opens, and the variation's own character begins only when Round 1 opens — Real-Time Delphi and Barometer Delphi get the orientation round too, and the only way to skip it is the same in all. Before 13.9.2026 the state machine jumped straight to Round 1 in those two, and the card was hidden.

3. Collection: the round-management card

Collection's tabs are in the order of the work: Situation · Invitations · Facilitation · Community · Responses (with Round 0 first when it is run). The round-management card is in Responses. The settings mentioned below live in their own homes: visibility & interaction and the deadline in Collection's Facilitation, the opening time and adding a round in Planning's Rounds section, and the phase exception and resetting an old round in Reporting's Maintenance.

Everything about running a round happens in Collection, and the card marked Round management is where the round's state and its controls live: the round, the Phase, the accumulation of Responses, and a button whose label is the act — "Move to the dialogue phase", "Open Round 0 (intake)", "Close round 2 (pause)".

Closing a round is a pause, not an ending. Closing even the last round leaves the study open: rounds can be added. Ending the study has its own control, End the study (⏹), and it is a decision — it asks for a closing letter to the panel and marks the handoff final. Adding a round afterwards undoes it. Before 31 Aug 2026 finality was a side effect of a round, and the facilitator had no way to say the study is complete.

There may be several buttons that advance a round; there is exactly one computation of what advancing means, and it happens on the server. That is why the label reads as a destination rather than a verb.

A step says that it was taken. After a successful transition a notice names the destination — "Round 0 opened · 4 key informants". Without it, opening looked like it had failed: the only thing that changed was the button's own label, which became the next destination, and a second click in the same spot carried the study past a round.

And a step that leaves something unrun is asked first. If round 0 is unopened or unfinished, moving to round 1 asks for confirmation that names both the destination and what is being left out: "Opening round 1. Round 0 (intake) will not be run." Round 0 is one-shot — orientation gathered afterwards is not orientation but post-hoc justification. Other transitions are not asked: a confirmation asked every time is learned as something to click through.

Before you open a round, two settings are worth a deliberate look, because both change what the panel experiences and neither can be repaired afterwards:

• Visibility & interaction — per phase. Round 1a — response is blind by default; Round 1b — dialogue is where the panel reads each other. The choices are Blind, After you answer, Open.
• Deadline — not set until you set it, then "{n} days left", +7 days, Clear, and deadline passed when it has. The deadline is your countdown; the panelist learns of it through a message, not through a clock on their screen.

View as panelist. The response-link card has a button that opens the survey in a new tab exactly as a panelist sees it at this stage — "The preview identity is not a panel member and nothing is saved." The previewer is an identity of its own (an ESIK code), the consent card and the Round 0 conversation are skipped, and every attempt to write returns "Not saved (preview)". It does not appear in the response rate or in the reciprocity figures. End preview closes the link and removes the previewer; a new one can be opened at any time. This is where the question what does this look like to a panelist gets its answer before opening, not after.

4. The response phase: your work is monitoring — and preparing

The study's phase is response. Yours is monitoring, and the screen is built for watching rather than doing: accumulation per thesis and per panelist, and one intervention that matters — Remind inactive ({n}), with a confirmation that names the number, and a result that says what happened: "Reminder sent to {n} panelists".

Read the response rate carefully. If the panel is mixed, an AI panelist answers every thesis on every round; a combined rate is therefore partly a property of the panel's construction rather than of its engagement. Pronoia computes the human rate, the AI rate and the combined rate separately and treats the human panel as primary when there are humans — and where there is nobody to divide by, the figure is absent rather than zero, because nobody was asked and nobody answered are different facts.

The AI panel's actions live behind a control. AI actions ▾ holds four: Generate thesis responses, Generate dialogue comments, Generate reactions, Generate revisions. They are behind a button because a block of machine operations rendered into the main column grows exactly the column this design is shortening, and puts machine work in front of the work that is actually yours: watching a human panel answer.

When the actions are not available the card says why rather than disappearing:

• "Actions become available once the first round is open."
• "Round 0 is open — the orientation conversations are started here."
• "The study has ended — no new responses are generated."

The middle one is not only an explanation. While round 0 is open the card carries the command itself — 🤖 AI panellists’ chat, and Regenerate AI chats when there are sessions to redo — next to Open the Round 0 view. It used to say the run was started from another view, and there was no way there from that sentence: a sentence pointing at another view is not a path to it.

Every generation passes a cost gate first. AI run cost estimate states the scope, the estimate in credits, the budget remaining, who pays and which limit binds — wallet balance, monthly cap, personal cap or this study's own cap — before Continue and generate. If the estimate is over: "The estimate exceeds the remaining budget — some responses may be produced rule-based."

Meanwhile, the other half of your work. This is when round N+1 gets built. The rail is free, Applies to: keeps the two rounds apart, and Results lets you glance at an earlier round without leaving what you are doing (§7).

5. The dialogue phase: monitoring and moderation

The panel now reads each other. Your view gains the comment stream, the dialogue threads — "Dialogue threads (comment-on-comment)" — and the AI panelists' activity in the same stream, marked.

Dialogue does not close answering. The phase adds a capability — commenting — and removes none: a panelist who did not get to answer can still answer during the dialogue, and changing an answer in the light of someone else's reasoning is the method's idea, not an exception. You need not open anything for it. A comment is locked against editing only once another panelist has replied to it; a panelist's own refinement of their own comment does not lock it.

Moderation here is mostly restraint. The dialogue phase is the study's learning phase, and a facilitator who answers arguments becomes a panelist with special powers. What the interface supports is the facilitator's own channel — the COMMUNICATION CENTRE — addressed to the Whole panel, a Group, or an Individual, with templates for Reminder, Round open and Results and placeholders ([link], [handle], [study]) resolved per recipient. Group reaches a panelist through both the home cell and any additional membership on either axis — a person who belongs to two groups receives the group message from both.

Two options there are worth understanding:

• Also send by email delivers the message to inboxes as well as the panel.
• Notice only, no text sends a mail that says a message is waiting and links to the panelist's own page, leaving the text inside the panel: "Choose this when the message contains panel content that is not meant for inboxes."

When a message goes to email it goes out as a letter: the study's name at the top, [link] as a button and the contact person at the end — the text version travels along for mail programs that do not display the letter. Two more choices:

• Include »What we ask of you« adds, after the button, a short three-point description of what the panelist is asked to do. The description is the same in every variation. It is on by default.
• Preview the letter shows the letter exactly as it will be sent — the same builder makes both. Placeholders are filled with sample values, and the preview sends nothing.

The delivery report is explicit about who did not get it — recipients without an address, withdrawn panelists skipped, and whether mail was sent at all.

6. The revision phase

Panelists revise their answers in the light of the dialogue, and say what moved them. Your view is the accumulation of revisions and the movement between rounds: Revisions, Revised, Revision rate, under the subtitle "Panelist changes to R1 responses after dialogue".

Revisions is a count, not a share: there is no maximum number of revisions, and the raw figure tells you something about the panel's diligence that a percentage would flatten. Revision rate, by contrast, is one share, computed in one place: the people who changed their position divided by everyone who responded, humans and AI panelists together. Before 14.9.2026 it was computed in four places in four ways, and the community view showed 1600 %.

7. Glancing at results without leaving your work

Results in the bottom bar is the only item there that is not a tab. It opens Results so far over what you are doing: round pills for the rounds that have results, and a row per thesis with n, median and IQR.

The window contains no editing surface at all, and there is exactly one way onward — Open in Analysis → — which carries the round you were looking at with it: "Viewing only — editing happens in Analysis." When there is nothing to show it says which nothing: "No results yet — the first round has not been opened.", or "Round 0 is the orientation; its results appear in the Round 0 view."

Glancing is by definition something that does not interrupt. That is why it is a window and not a destination.

8. Between rounds: your richest screen too

The round closes. For the panel this is the interim result; for you it is preparing, and it is the only point in the process where the main column sits on the boundary between two kinds of work — analysis produces what planning consumes.

What Analysis gives you here

• Summary — distributions, central tendency and spread per thesis.
• Statistics & change (thesis stats, convergence, key figures, dialogicity, revisions) and ROUND-OVER-ROUND MOVEMENT — Converged, Diverged, Stable, and "Attention — split theses": "Theses where views diverge the most — often the most interesting part of the analysis."
• A readability verdict that is about your data rather than your patience: Not readable yet ("Stability cannot be measured before the round is closed and there is something to compare against"), Needs action ("Coverage is thin. Remind the non-responders before reading the result"), or Readable.

What builds the next round

• PIRE — the inter-round engine — proposes, one thesis at a time, how to continue: convergence, stable dissensus, noisy spread, an emergent theme, a bandwagon flag, and an operation to match. You accept, change the operation, flip the mode, or lock the thesis out of the next round. Its own help states the limit plainly: "PIRE proposes — you confirm. It does not replace deep analysis."
• Build the round. "Build the next round's survey from the previous round's results — copy as a starting point and edit, or open with the same theses (re-rating)." The first of these is the default in a round-based Delphi: the next survey is built from the previous results by editing, removing and adding. Plain re-rating is an exception worth justifying — in a barometer it is, conversely, the whole idea. Proposals that involve interpretation are flagged for you to reformulate before opening; you always edit the theses before the round opens. A thesis can also be copied as the base of a new one and hidden or shown mid-round behind the same gates; hiding does not delete answers.
• Reading the discussion and picking. The pause is the moment to go through the previous round's discussion question by question and branch by branch. Open the whole discussion shows a thread in full; the Timeline shows reasoning and dialogue in one stream, newest first. Any reasoning or comment can be starred (☆) — "Mark as thesis material for the next round" — noting at the same time the reason for the pick. Picks collect in the basket Thesis material for the next round: "A pick does not change the comment and is not visible to the panel. It is your memory between rounds."
• The basket is source material, not a thesis. When you write a new thesis, the basket opens beside the editor under Source material for the new thesis: "Write the thesis yourself. The contributions you select are recorded as its sources." No button turns a pick into a thesis. The selected ones are recorded on the thesis card as Source material, which later tells where the question came from. Basket and source material are different things: the basket is your tool between rounds, the source material is bookkeeping. That is why they behave differently when a contribution is deleted — the basket skips it, the source material keeps the record and the reason but the content is emptied.
• The panel's readiness for the next round is a different question from the panel's composition: who answered last round, who withdrew, where the coverage gaps are. Composition is edited in Planning; readiness is read here.
• If a round was opened past the pause and you need its survey built after all, Return Round {n} to preparation closes it back to the pause and keeps the answers already given.

When you want real interpretation, it happens outside Pronoia: Send round to DAE analysis opens the export. Theming, scenarios, argument and discourse analysis and synthesis are the analysis ecosystem's work, under its own quality gates. Pronoia describes; it does not interpret, and this button is where that boundary is crossed on purpose.

9. Reporting and handover

Generate results report produces a descriptive report — summary, per-thesis result cards, the consensus–dissensus map, round comparison R1→R2, and highlights (greatest disagreement, most discussed, most revised). Print / Save as PDF and Download Word (.docx) take it out of the browser.

The report says what it is, in its own footer, and the sentence is worth quoting to whoever receives it:

"This report is not an analysis but its source material."

Export produces three files for the analysis pipeline: the CSV (quantitative data), the Handoff Package (responses, arguments, dialogue, revision history) and DAE_STATE.json (pipeline control data). Export sits in two places on purpose: Reporting is a phase — what you do at the end; Export is a function — something you can do at any moment.

Anonymize study is the last step and it is irreversible: panelist names and email addresses are cleared and invite links revoked, leaving anonymous research data — pseudonyms without a key. Do it when the study is over and the material has been exported.

10. Exceptions, and what they cost

Three things sit outside the normal flow. Each is available, each is marked as an exception, and each is recorded.

11. What the panel sees while you work

Read that column downwards and the design is one design: the panelist gets content when it can no longer anchor them; you get it immediately, because your job is to see it.

12. Five things that cannot be undone

1. Regenerating Round 0 AI chats — the existing ones are deleted.
2. Resetting a thesis, a round or a study's responses — responses, revisions and comments for that scope are deleted. A round reset asks you to type the study name.
3. Withdrawing a panelist — name, email and link are permanently deleted.
4. Anonymizing the study — identities are removed and links revoked.
5. Closing a round — a closed round stays closed. Changing it afterwards would silently change a response rate that has already been shown and possibly analysed, and the comparison between rounds is the method itself.

Item 5 is not only guidance: the server refuses to change a closed round's dialogue, phase rules and schedule. That is why Collection shows only the open round, and a closed round's material is read in Analysis.

Everything else in this guide is reversible, and most of it is reversible by pressing the same button again.

13. What is folded, and when

Some sections are folded by default when they do not belong to the phase you are in. This is a promise, not a surprise: the table below is the whole rule, and it is generated from the same module the program reads (lib/sectionFolding.mjs). A guard test fails if the two disagree.

Three things are worth knowing before you read it.

1. Your own choice always wins. Open a section by hand and it stays open, whatever the phase says.
2. A folded section is not a hidden section. It leaves a chip and a line telling you how many sections the phase folded, so you can open them. In the CONTENTS row at the top of the view, a chip's name takes you to the section and opens it.
3. A section not listed here is never folded. The rule only removes; it never adds.

task is your work in the current phase (build · monitor · prepare · report), phase is the round's phase (response · dialogue · revision · closed), and mode is the Delphi variation. An empty column means that dimension does not restrict the section.

If a section you needed was folded, that is a fault in this table rather than in your work — and it is worth reporting, because the table is the rule.

14. The line that says where on the arc you are

Above the state card there is one short line. It is the only place that answers the question where in the whole process am I — not which screen you have open, but which stage of the study you are standing in.

It says one of these, and nothing else:

Real-Time Delphi and Barometer Delphi speak differently, because they do not walk the same arc (Classic Delphi and Argument Delphi walk it together):

Why it matters that this is a sentence and not a code. Internally each stage has a short code (A1, B1, A2 …), and it is in the page's markup where a test can read it. It is deliberately not on the screen: a code would be a third vocabulary beside the sentence and the guide, and the codes only exist in round-based Delphi — in the Barometer and in Real-Time Delphi the code is empty, which would make the screen look as though the code were missing rather than as though it does not exist.

So the line is written to be read, not decoded. If it disagrees with what you believe the study is doing, believe the line: it is computed from the study's own state, not from the round number.

15. Retirement is the Barometer's way, not a general one

A thesis whose time has passed can be retired — but only in the Barometer, and only at the turn of a wave. Everywhere else the program refuses, and the refusal says why: "Retirement is possible only in the Barometer, between waves. Let the thesis run the round to its end."

This is a methodological rule, not a limitation of the software.

• In a round-based Delphi a thesis runs the round to its end. The whole point of the method is the comparison between rounds: the same thesis, asked again after the panel has seen each other's reasoning. Removing a thesis mid-round would delete the very measurement the round exists to produce.
• In the Barometer the waves differ only by their date. The same theses are rated from one time to the next, so a thesis can genuinely reach the end of its usefulness — the future it asked about has arrived, or the question no longer divides anyone. The facilitator decides, and a new thesis can take its place.

What retirement does, exactly:

1. The thesis stays in the current wave to the end. Panelists can still rate it, and the wave's results are archived into the wave series as usual.
2. At the wave turn it moves to the archive and is no longer asked from the next wave onwards.
3. Its responses are kept for analysis. The export marks it lifecycle = retired, so the analysis can see that the thesis is no longer being asked — a retired thesis is not a deleted thesis, and the difference matters when someone later wonders why a series stops.
4. The decision can be cancelled before the wave turn.

The reason is asked for and stored ("time is up", "the future is visible"). It is worth writing properly: at the turn of the next wave it is the only thing that explains why the series ends there.

Written 26.8.2026, against the interfaces as rebuilt on 23.–25.8.2026. Button and tab names are taken from the application's own English strings, and a guard test (scripts/test_guides_coverage.py) fails if this guide names a control differently from the program. The specification behind this text is docs/UI_Vaihearkkitehtuuri_IA_v2.md; where it and this guide disagree, it is right and this is stale.