A queue model built from all 24 monthly SkillSelect snapshots (Sep 2024 – Aug 2026), read straight from the Qlik engine behind the Department of Employment's public EOI dashboard. Every figure here is a distinct-EOI count, not a row count.
that the cut-off for ANZSCO 2349 reaches 85 points in a round of 10,000 — the size implied by the 2026–27 planning levels. Central forecast 75 points, 80% interval 70–81. At a 5,000 round it falls to 77%; at 7,500, 90%.
Calibrated on this method's own errors across held-out rounds, not on a judgement call. Run it for your own occupation and score →
That number came from asking how many of the five historical allocations would have covered your queue rank. Three of those five are not plausible draws for the next round — they come from a regime when this occupation group received a far smaller slice.
| Round | 2349 allocation | Share of the round | Comparable to today? |
|---|---|---|---|
| Sep 2024 | 3 | 0.07% | stale regime |
| Nov 2024 | 11 | 0.13% | stale regime |
| Aug 2025 | 15 | 0.42% | transitional |
| Nov 2025 | 24 | 0.46% | yes |
| Jun 2026 | 35 | 0.67% | yes |
Treating a 0.07% share from September 2024 as an equally likely outcome for the next round is what produced the pessimistic figure. Share autocorrelation on the recent transitions is 0.92–0.96, so the recent share is the forecastable quantity and the 2024 values are a different regime.
A later revision also retired the interim “1.6× headroom” framing that replaced it. Headroom is a ratio, not a probability, and it could not say how likely the cut-off was to land where the model said. The verdict above is now a calibrated probability taken from the distribution of this method's own out-of-sample errors — see UNCERTAINTY.md. Both superseded numbers are recorded here rather than quietly amended.
Nearly half this pool did not lodge at the score it now holds. The table records every version of an EOI with its own validity interval, so an upgrade is directly observable — and an upgrade resets the date of effect, which is what the queue is actually ordered on.
This works in your favour. Someone sitting at 80 points who reaches 85 after 10 Sep 2026 takes a date of effect later than yours and joins the queue behind you, not ahead. Because the 85+ band turns over almost completely within a year, your position stops eroding the moment you lodge — it only improves as those ahead are invited away or hit the two-year expiry. Of the 31 people ahead of you in 2349, every one has a date of effect before 10 Sep 2026, verified directly rather than inferred from lodgement month.
Five rounds, 181 occupations, cut-offs from 65 to 100 in a single round, a published tie-break that appears to exclude recent applicants, and a ceiling that never binds — all of it follows from one rule.
Residual. 13 of 58 rationed cells sit above their group's lowest invited score, which strict points-then-date forbids. Cause: submission month is not date of effect. An EOI whose points changed carries an old submission month but a fresh date of effect, so the exhaustion probe misreads it. Confirmed directly — 2349's 85-point cell fell from 10 to 1 across the Jun-2026 round, then refilled to 13 by August including a 2024-submitted EOI absent in June and July, which can only have re-entered at 85 with a 2026 date of effect.
The Department published per-occupation minimum scores for the 4 June 2026 round. Those figures were not used to build this model — they are an independent answer key for cut-offs derived from raw EOI records.
Every one of the 12 disagreements is positive (+5 in eleven cases, +10 in one) and none negative. That is the signature of the single-leg filter: where the marginal invitee was a multi-leg EOI, the derived cut-off lands one tier high. The method is conservative by construction — it cannot overstate a candidate's chances. The round-size estimate is validated too: 9,761 derived against 10,000 official, a 2.4% error. For Physicist the model derived 80 and the Department published 80.
Within a unit group, a round invites strictly down the points order. 31 people in ANZSCO 2349 sit ahead of you on 85+ points — every one of them verified to hold a date of effect before 10 Sep 2026. You are invited if and only if the round allocates at least that many invitations to 2349.
| If the round is held | Ahead of you lapsed | Your rank | Allocation needed |
|---|---|---|---|
| by 30 Sep 2026 | 2 | 32 | ≥ 32 |
| by 31 Dec 2026 | 4 | 30 | ≥ 30 |
| Round | Allocation to 2349 | Recency weight | Covers a Sep-2026 round? |
|---|---|---|---|
| Sep 2024 | 5 | 0.032 | short |
| Nov 2024 | 21 | 0.065 | short |
| Aug 2025 | 29 | 0.129 | short |
| Nov 2025 | 43 | 0.258 | covered |
| Jun 2026 | 87 | 0.516 | covered |
Allocation history 5 → 21 → 29 → 43 → 87 — monotone increasing, 17.4× over five rounds, and 2349's share of the round grew from 0.12% to 1.67%. The downside risk is not your score. It is a policy cut returning this stratum's allocation below 32, as in Sep-2024, Nov-2024 and Aug-2025.
The same model run for all 82 unit groups across all five rounds — 315 group-rounds — rather than the single round it was first checked on. Predict each group's cut-off by walking its pool from the top until its allocation runs out, then compare with what the round actually did.
| Round | Panel coverage | Groups | Exact | Within ±5 | MAE (pts) |
|---|---|---|---|---|---|
| Sep 2024 | 12% | 63 | 13% | 38% | 9.92 |
| Nov 2024 | 20% | 73 | 15% | 45% | 8.70 |
| Aug 2025 | 69% | 56 | 88% | 96% | 0.80 |
| Nov 2025 | 81% | 61 | 75% | 100% | 1.23 |
| Jun 2026 | 98% | 62 | 85% | 98% | 0.81 |
The model is not uniformly good, and the pattern is the point. Accuracy tracks how complete the panel was: correlation between panel coverage and error is −0.962. Where the panel is complete the model is 83% exact with a mean error of 0.95 points; on the two 2024 rounds, when the dashboard had captured only 12–20% of the eventual pool, it drops to 14%. The 2024 errors are all negative — with the pool understated the top-down walk runs out of people too early and dives too deep, which is exactly what under-coverage produces.
Out-of-sample means the allocation is forecast from the previous round only, with round size taken as given. The gap between knowing the allocation (MAE 4.54) and forecasting it (5.64) is just 1.1 points — the ranking rule is the reliable part; the allocation is the uncertain input.
The first version reported “49 of 49 occupations within ±5 points”. That compared single-leg invitations against an all-leg pool. On a consistent all-leg basis the model collapses, because the all-leg minimum invited score is dragged down 7.3 points on average (11.3 in large groups) by EOIs invited for 190/491. Everything above is on a consistent single-leg basis, the one that matched the official table.
2349's allocation series is monotone increasing, but only 18% of groups are — it is unusual, so don't read the trend as typical. And its clean-step boundary is significantly associated with having a small pool (p < 0.0001), so that favourable property is partly a small-numbers artefact.
Every round invites someone at 65 points, so there is no such thing as "the 189 cut-off". Each ANZSCO unit group clears to its own depth. Physicists sit in 2349 Other Natural and Physical Science Professionals — a low-volume group whose cut-off has dropped 15 points in five rounds while the national floor never moved.
Minimum points invited in unit group 2349 per round. Shaded band = scores above 85. Jun 2026 went below your score, inviting physicists at 80.
Only five months in the whole panel contain genuine 189 invitations. The other 19 months show nothing but 190/491 state nominations leaking through a shared status field — see the methodology note below.
| Round | 189-only invites | Est. total | To 2349 | 2349 stock ≥85 | 2349 cut-off |
|---|---|---|---|---|---|
| Sep 2024 | 4,274 | ~7,814 | 5 | 5 | 95 |
| Nov 2024 | 8,230 | ~14,507 | 21 | 14 | 90 |
| Aug 2025 | 3,574 | ~6,705 | 29 | 62 | 90 |
| Nov 2025 | 5,207 | ~9,821 | 43 | 47 | 85 |
| Jun 2026 | 5,198 | ~9,761 | 87 | 50 | 80 |
Allocation to 2349 and its ≥85 stock are both on the all-leg basis, so they are directly comparable; round size is quoted single-leg with a difference-in-differences estimate of the true total. The 2024 rows look self-contradictory — Nov-2024 allocated 21 against a recorded stock of 14 yet held the cut-off at 90 — because the panel was still filling in late 2024 and understates the true standing pool then. The backtest below is run on Jun-2026 only, where coverage is complete. "189-only" counts EOIs whose only visa leg is 189, so the invitation is unambiguously a 189 invitation. "Est. total" adds multi-leg EOIs via a difference-in-differences correction against non-round months.
What moves the answer is not any one applicant's attributes but the balance between a pool growing faster than it is cleared, and an allocation to small scientific occupations that has been expanding sharply.
June 2026's published tie-break was April 2026, which sounds fatal for a 10 Sep lodgement. It is not: 20.2% of invitations at 90+ points went to EOIs dated after that tie-break. The date binds only at the marginal position within a score, not across all scores.
Cohort survival runs 96% at 0–6 months, 82% at 12–18, 77% at 18–24, then collapses to 0.8% beyond 24 months. Four of the 31 people ahead of you in 2349 lapse by the end of 2026, which is why your rank improves simply by waiting.
ANZSCO 2349's annual ceiling is 500, and the largest round used only 87 of it. Nothing caps this group from going deeper — a downside risk this model had flagged and can now discharge.
Invitations are highly concentrated: ten unit groups absorb two thirds of a round, thirty absorb 94%. Being in a small, low-competition group is an advantage here — 2349's cut-off fell while the national floor stayed at 65. Not visible in this data: annual planning levels and ministerial direction on occupation priority. Those set round size and cadence, and no amount of EOI data substitutes for them.
State-nominated subclasses are not points-ranked federally — a state selects against its own criteria, so a queue model does not transfer. For context only, here is where physicists stand in those pools.
| Subclass | Physicist EOIs | At 90+ pts (= 85 + nomination) |
|---|---|---|
| 190 State Nominated | 128 | 31 |
| 491 Regional Nominated | 125 | 44 |
A 190 nomination adds 5 points and a 491 adds 15, so an 85-point 189 profile appears at 90 and 100 in those pools.
The points cut-off each ANZSCO occupation actually reached in each of the five rounds, with its current standing pool. Green means that round would have reached your score; set your own points below. A dash means the occupation received no invitation in that round. Small number beside each cut-off is that round's invitation count.
| Occupation | Sep 2024 | Nov 2024 | Aug 2025 | Nov 2025 | Jun 2026 | Next round | Pool @85 | Pool >85 |
|---|
Filtering Visa Type = 189 and counting INVITED overstates 189 invitations by roughly 2×,
because an EOI invited for 190 still carries a 189 leg. Across the panel the three visa filters overlap ~50%. Isolating
single-leg EOIs is what reveals that there were five rounds, not twenty-four.
The Qlik front end masks small cells, but the websocket engine underneath does not. Verified by complement differencing — total minus total-excluding-the-occupation reproduced the direct figure exactly (72 = 72) down to cells of 1.