Series Evidence Synthesis

Series Evidence Synthesis — What Four Webinars Actually Said

Stream 1 · Project Output · Asset Zero. Drafted 2 Aug 2026 (AI-drafted, fact-checked against the results files; provisional until a named human signs it off). This is the spine document: every framework section drafts from here, and every claim here traces to a results file. Nothing in this document is an official output until a named human signs it off.

0. Sources and how to read this document

Series total: 2,796 recorded responses (663 + 524 + 625 + 956 in the live bundles, plus 28 W3 sector answers logged outside the cue script). Use "almost 2,800 responses" in comms; do not round to 3,000 in anything citable.

Reading rules (they are printed in the files; they bind here too):

1. Every figure carries its own n. Rooms shrink across an hour; compare within a block, not across blocks.
2. Where a paired figure exists (same participants matched on anonymised token), it is the honest figure and the one to quote: W4 boundary ladder (+0.44, n=25), W4 retained-rights re-price (n=25), W4 comfort (+0.18, n=17), W3 trust (+0.43, n=14).
3. Multi-select percentages are % of that block's voters and sum past 100%.
4. Percentages below use each file's own rounding. Ratios computed on raw counts.
5. Cross-webinar comparisons are cross-room (different people, different sector mixes) except where an instrument was deliberately re-run verbatim ([W4] §7 re-runs [W2] dr-poll-c_trig; [W4] C04/C21 re-runs the series boundary vote twice in one room).

1. The series at a glance

Sources: [W1] q5-ai-role, q9-first-principle · [W2] boundary-vote-green-score, principle-test · [W3] §1a, §2 · [W4] §1a–1b, §9.

Room composition where measured: W3 was 43% transport/roads, 36% consulting (n=28, [W3] §6); W1 skewed government/infrastructure/transport ([W1] q3-sector). W4 did not capture sector (its C06 slot went to the org-reality questions, [W4] §3). Treat sector mix as a confound on all cross-webinar reads.

2. Finding 1 — The ladder holds at Recommend, until the room watches bounded action work

For three webinars, on three different decision domains, the community capped AI at analyse-or-recommend by a stable supermajority: 68% on facts about the asset ([W1] q5-ai-role, n=78), 67% on acting on a prediction ([W2] boundary-vote-green-score, n=42), 72% on money ([W3] §1a, n=18). Autonomy peaked at 13% (W4 cold, 5/39) and hit 0% on money. And in all four series boundary votes (W2, W3, W4 cold & warm), "no role" and "should not be used" took 0% every time (W1's poll, which offered only "should not be used", likewise 0%); across the eight decision-rights polls (W2×5, W4×3), "don't use AI" never drew more than one vote (≤3%). The community, in the abstract, never votes to ban AI (concrete "nevers" do exist — see Finding 3).

W4 ran the series' controlled experiment: the identical boundary vote, cold at C04 and warm at C21, after the room spent ~40 minutes watching a bounded agent act (the storm, the envelope, the 2 a.m. signature). Result:

Paired (n=25): mean ladder move +0.44 steps — 10 moved AI up, 5 down, 10 stayed ([W4] §1c). The warm room did not move toward autonomy (13%→9%); it moved from analyse/recommend to act with approval.

Spine claim: the community's default is Recommend; demonstrated, bounded, stoppable, ledgered action moves it one step — to approval-gated action, not autonomy. The ladder's load-bearing rung is Act with approval, and it is earned by showing the controls, not by asserting them.

Corroborating trend on the one verbatim re-run instrument — the intervention trigger ([W2] dr-poll-c_trig n=41 → [W4] §7 n=33, different rooms, 4 weeks apart): Recommend 71%→33%, Act-with-sign-off 5%→24%, autonomous 0%→0%. Same shape: the middle of the ladder is migrating one rung, autonomy is not moving.

3. Finding 2 — Approval over interruption: the gap that widened for three webinars, and W4's re-price of the stop

The community consistently prices approving before far above interrupting after. Define the gap as Approve the decision ÷ Hold stop authority (W1 offered no stop option; its nearest interrupt right, Override, is used for W1 only):

Sources: [W1] q6-humans-retain · [W2] boundary-vote-green-score (humans retain) · [W3] §1b · [W4] §1d–1e. W3's n=18 is small; treat its ratio as directional.

Stop authority and override ranked at or near the bottom of every humans-must-retain vote from W2 to W3 (W2: 12–13 of ~42 voters; W3: 17%). This was the tension W4 was built to test — the session showed what the stop costs (the storm; EEMUA-style flood; the quarantine call), then re-ran the vote.

The re-price ([W4] §1e, paired n=25): stop authority 48% → 64% — 4 added it, 0 dropped it. It is the only retained right that nobody dropped after watching the machine act. Approval also rose (68%→80%); risk acceptance rose (40%→52%); interpret-context doubled (20%→40%). What fell: govern-the-knowledge (60%→48%) and own-the-value-judgement (48%→44%) — under pressure the room traded abstract stewardship rights for concrete interrupt-and-accountability rights.

Spine claim: approval is the community's currency, but the stop was systematically under-priced until the room saw a storm. The framework must price the stop explicitly (P4-B) — and the evidence says demonstration, not exhortation, is what re-prices it.

Reality check (Finding 7 has the full set): only 36% have a named person who could pause machine-generated work within the hour today; the pause has been drilled in the last year in 22% of organisations, and 57% don't know ([W4] §3).

4. Finding 3 — Reversibility is worth 38 points; "never" only appears when the case is concrete

W4's envelope exercise put six concrete work classes on an allow / approval-gate / never ballot ([W4] §6):

The reversibility flip ([W4] §6, the premium experiment): the same pothole inspection drew 59% allow when cancellable for two hours (n=29) vs 21% allow when cost commits on creation (n=33) — a 38-percentage-point reversibility premium, measured on this community for the first time. (Note the two flip variants had different n; the premium is a cross-block read within one session and one instrument family.)

Two structural observations:

1. Abstract bans never win; concrete bans exist. Across every abstract boundary vote, "should not be used" took 0%. Put a live 66kV isolation in front of the same community and 19% say never, and allow-auto collapses to 11%. The framework's "never automate" list must be written in concrete work classes, not categories of principle.
2. Reversibility moves votes more than work type does. The six work classes span water, power, roads, health — allow-auto varies only 42–53% across five of them. One reversibility change moves the same class by 38pp. This validates the W5 brief's axis — consequence × reversibility × evidence ([W5B] §1) — with reversibility as the strongest measured lever.

Spine claim: the Decision Rights Matrix's "Act (bounded)" column is really a reversibility column. Envelope anatomy (scope, magnitude caps, rate limits, reversibility window, owner, expiry) is what the community is actually pricing.

5. Finding 4 — The community trusts records over promises: ledger shapes rate highest

Eight principles were rated live across the series (1 = reject, 5 = adopt):

The ledgers win their rooms: the highest-rated principle of the whole series is the action ledger (P4-C, 4.27 — every machine action logged with an owner's signature), and W3's highest was its twin, the deferral ledger (P3-C, 4.07 — every deferral logged as a named, dated risk acceptance); each beat every alternative rated alongside it. (P4-A, the trigger gate, sits between them at 4.15.) The weakest of the series, P3-B (3.39), is the one principle that grants a right to see rather than requiring a record to exist.

Consistent signals from W4's honesty poll ("which clause would your organisation quietly skip first?", n=27, [W4] §9): re-signing schedule 0% — nobody thinks the signature ritual would be skipped; what dies first is the close-out feedback loop (30%) and the practised stop drill (22%) (37% claim they'd keep all). The community believes in signatures and doubts follow-through — precisely the ledger-over-promise pattern.

Spine claim: wherever the framework must choose a mechanism, choose the recording mechanism: named, dated, signed, auditable. The community endorses accountability artefacts more strongly than access rights, oversight declarations, or process promises. (A consistent reading, offered as interpretation: W2's principle — strong content, no ledger clause — landed mid-pack at 3.76.)

6. Finding 5 — The community set its first number: 7.5 machine-created work orders per planner per week

Asked where auto-pause should kick in — machine-created work orders per planner per week — the room drafted the missing volume budget ([W4] §5, n=30):

• Median 7.5 · mean 37.3 (the mean is dragged by a heavy tail: answers ran 0→300)
• 30% answered 0 — never auto-create
• Full distribution is published in [W4] §5 (all 30 answers, sorted)

This is the community's first quantified boundary parameter — the seed of an "EEMUA-for-work-orders" style volume budget (the alarm-management analogy: EEMUA 191's steady-state budget of ~1 alarm per operator per 10 minutes; see the W4 research pack, dossier 05). Quote it as "the room's median was 7.5; a third said never" — the median and the zero-share together are the finding; the mean alone misleads.

Spine claim: the framework's envelope section can ship a worked example with a community-sourced default rather than a blank. Status: draft parameter for validation in cycles (OQ2 in the open-questions register), not a recommendation yet.

7. Finding 6 — Money is where values hide: weights, defaults and the signed objective

W3 put the value-framework question directly to the room:

• What this room valued (100 points across five criteria, n=45, [W3] §4): medians — cost 26 · risk 25 · service 24 · equity 15 · carbon 10, last. Cost/risk/service take ~three-quarters of the weight. (No is/ought baseline exists — the pre-survey never opened; flag when quoting.)
• 85% would not have signed the vendor-default weights nobody in the room had chosen (n=39, [W3] w3-s1-r5-sign-default).
• Where are your organisation's value weights written down? (n=36): a named person could show me them today 39% · a team could reconstruct them 25% · we don't use a weighted model 19% · nobody 11% · only the vendor 6% ([W3] w3-s1-r6-weights-written).
• Shown what an optimiser quietly defunded, 64% rejected the trade (n=42, [W3] w3-s2-r5-accept-trade); asked who owns checking that today: we don't use an optimiser 37% · a named person 32% · a team could 17% · nobody 12% · only the vendor 2% (n=41).
• Median guess for the share of AI-generated plans that get audited: 0% (mean 10%, n=35, [W3] w3-p2-audit-guess).
• Sim 3 (sign / send back / override on four AI-assembled plans): no plan reached a majority to sign; the strongest response to the "beautiful but gamed" plan was send-back/override 72% combined ([W3] §5). The room's instinct on machine-assembled plans is scrutiny, not signature.

Spine claim: P3-A (only a named human sets and re-signs the weights) is evidenced not by what the room endorsed (3.73) but by what it did: refusing default weights 85%, rejecting the optimiser's silent trade 64%, and reporting that in most organisations the weights and the check have no named owner. The matrix rows for prioritisation/valuation must carry the named-accountability column as their load-bearing cell.

8. Finding 7 — The is/ought gap: rooms demand gates their organisations don't run

Put the boundary votes (ought) beside the org-reality polls (is):

Also W1's field-validation poll ([W1] q8-patterns, n=71, multi-select): 75% have seen "bad data → confident output", 45% the authority illusion, 28% silent fabrication, 25% data laundering, 17% rare-event blindness, 14% rubber-stamp oversight; only 11% "none yet". The failure-modes library is not hypothetical — the room has met it.

Trust starts low and moves slowly: trust in an AI-generated capital plan 2.64/7 before → 3.12 after, paired +0.43 (n=14) ([W3] §3); comfort with the overnight work order 2.86/7 → 3.11, paired +0.18 (n=17) ([W4] §2). One session of evidence and controls buys about a third of a point on a seven-point scale. Trust is earned in inches — which is the argument for the living framework's telemetry over a one-off PDF.

Spine claim: the framework's job is to close an is/ought gap the community itself measured. That is the "getting started → maturing" path's evidence base ([W5B] §7.1, §7.6): most member organisations are at the left edge (no auto-creation, no drilled stop, unwritten weights), while their practitioners already demand the right-edge controls. Both facts are in the data; the framework must serve both.

9. Finding 8 — The human core is stable across the lifecycle — and it sharpens under pressure

What humans must retain, across every instrument that asked (multi-select; % of that block's voters; instruments evolved, so read ranks within a column, not levels across columns):

Sources: [W1] q6-humans-retain (7-option list: no stop-authority or own-value options; had define-purpose) · [W2] boundary-vote-green-score (8-option list stabilises here) · [W3] §1b · [W4] §1d–1e.

Stable structure: Approve is top-two in every room that voted it. Validate is always high. The domain flavours the top right — W3 (money) uniquely elevates own the value judgement to #1; W1 (data) elevates govern the knowledge. And W1's weakest right — accept risk, 47%, the lowest of its seven — stayed weak in W2–W3 (33%, 22%) until W4's action session re-priced it (+12pp paired). The pattern of Finding 2 generalises: exposure to consequence shifts the community from stewardship rights to interrupt-and-own rights (interpret-context +20pp, stop +16pp, approve +12pp, accept-risk +12pp; govern −12pp, own-value −4pp).

W1's open text corroborates the core: top-upvoted hopes/worries are bad data and results (▲29), lack of human oversight (▲24), data security (▲19), and one answer asks, in effect, who goes to jail for an issue controlled by AI ([W1] q2-hope-worry) — accountability by name, again. The questions the room left with centre on accuracy and trust ([W1] q10-leaving-question).

Spine claim: the framework's human-role set (define purpose · validate evidence · approve the decision · own the value judgement · accept & record residual risk · hold stop authority · override · interpret context · govern the knowledge · remain accountable by name) is fully evidenced across four rooms — with measured guidance on which rights need defending in the text (the ones rooms under-price cold: stop, risk acceptance, context) versus which defend themselves (approval, validation).

10. What the spine means for each framework section

11. Caveats, exclusions and data hygiene (bind on every downstream use)

1. Instruments evolved. W1 ran on Mentimeter with different option labels ("Make corrections with approval" ≈ act-with-approval) and a 7-option retained-rights list without stop authority or own-the-value-judgement. Cross-webinar tables in §1 and §9 are alignment reads, not identical instruments. The boundary vote is verbatim-stable from W2 onward; the decision-rights 0–5 poll is verbatim-stable across W2→W4.
2. Small/late blocks. W3's boundary vote ran at C22 with n=18 after attrition (39 answered the opening question, 26 the closing one) — its percentages are directional. W4 fixed this by design (vote twice, C04+C21), which is itself a method finding for the paper.
3. Unusable blocks — do not quote: W3 w3-s1-r4-untouchable word cloud (prompt misfired; 6 of 17 answers say they didn't understand); W3 w3-monday-action (n=4). Both are flagged in [W3] and withheld from the public page.
4. Withheld data: W2's 11-section volunteer questionnaire drew 4 responses — below the MIN_N=10 privacy gate; its contents are deliberately not in the results file nor here.
5. Paired beats unpaired. Where this document gives a paired figure, comms must use it (W4 ladder +0.44 n=25; W4 stop 48→64 n=25; W4 comfort +0.18 n=17; W3 trust +0.43 n=14). Unpaired shifts mix a change of mind with a change of room.
6. Cross-room reads (W2 71% → W4 33/24% on the trigger instrument) compare different audiences four weeks apart; say "the community", not "participants changed their minds".
7. Percentages reproduce each file's rounding; multi-selects sum >100%; ratios in §3 are computed from raw counts.
8. Consent & anonymity: figures are consented responses; W3/W4 tokens are HMAC-anonymised and irreversible. No individual is identifiable; keep it that way in derived work.
9. Sector mix differs per room (W3 transport/consulting-heavy; W1 government-heavy; W4 unmeasured). No claim in this document is sector-specific.
10. Claim-language discipline for anything derived from this document: "the world's first Human-AI Boundary Framework for asset management"; cite PMI's June 2026 standard (world's first global standard for AI in project work) rather than competing with it; the Asset Zero claim keeps its "to our knowledge" qualifier. See Phase 2 - AI Native Project/research/01-ai-native-pm-landscape.md §(d).

12. One paragraph for the paper (the spine in prose)

Across four webinars, June–July 2026, an Australian and New Zealand asset-management community (2,796 recorded responses) drew a consistent boundary: AI may analyse and recommend almost anywhere, may act only inside engineered, reversible, rate-limited envelopes behind human approval, and may never own purpose, values, risk acceptance or the stop. The ceiling held at analyse-or-recommend (67–72%) across data, predictions and money; it moved — once, by +0.44 rungs paired — when a room watched bounded, ledgered, stoppable action for forty minutes, and it moved to approval-gated action, not autonomy. The community priced approval above interruption for three sessions (1.1×→2.7×), then re-priced the stop the moment it saw a storm (48%→64% paired, with zero drops). It valued reversibility at 38 percentage points on the same work order. It rated ledgers — named, dated, signed records of deferrals and actions — above every other mechanism offered (4.07, 4.27 of 5). It drafted its first quantitative boundary parameter: a median of 7.5 machine-created work orders per planner per week, with a third saying never. And it reported an is/ought gap the framework exists to close: 36% could pause the machine within the hour today, 22% have drilled it, and the median practitioner guesses that 0% of AI-generated plans are audited. The boundary, on this evidence, is a governance choice — and this community has now specified it, decision by decision.

Provenance: AI-drafted (Stream 1, 2 Aug 2026) from the four results files listed in §0; verified figure-by-figure against those files; fact-check pass run before repo commit. Interim boundary registration: Recommend/draft-only — a human (Kai) reviews and commits. Status: PROVISIONAL until WG cycle 1 opens and a named human signs off.