How we did it

We built this the way we build everything at Intuition — with AI doing the reading and our tutors deciding what matters. Six AI models were each given identical evidence packs and asked to predict, question by question, what the 2026 HSC exams will ask. Before publishing a word, we made the same panel predict the 2025 papers blind — and scored it. This page is the whole story.

1 · The evidence packs

For each subject we assembled every official NESA document from 2019 to 2025: the exam papers, the marking guidelines, and — the underrated one — the marking-centre feedback reports, where the markers tell everyone exactly where students fail. Committees write questions there. A single subject's pack runs to roughly 150,000–250,000 tokens of source text; every model received the same pack, syllabus first, then each year's paper, guidelines and feedback in order.

2 · The panel

Six models, five labs, three different harnesses, identical inputs. No model saw another model's answer. Disagreement is signal — that's the point of a panel.

Model	Lab
Claude Fable 5	Anthropic
Claude Opus 5	Anthropic
GPT-5.6 Sol	OpenAI
Gemini 3.1 Pro	Google
Grok 4.6	xAI
DeepSeek V4	DeepSeek

90 prediction runs across the 15 subjects (including the archived pilot round). We audited all agent transcripts for outside tool use: zero web calls in any prediction run.

3 · Questions, not topic lists

Ask a model “which topics will be examined?” and it saturates — every topic appears somewhere in most papers, so everything gets a high probability and you learn nothing. So we forced predictions at question level: the concrete scenario, the marks, the section, the skill chain, the trap — each with its own probability. Question-level predictions can be graded honestly, which is what makes November's scoring meaningful.

4 · The honesty rules

Three rules protect the numbers from flattering themselves:

Same-lab discount. Two of the six models are Anthropic's. When they agree with each other within ±0.1, each vote is discounted to 0.75 weight — siblings agreeing is weaker evidence than strangers agreeing.
The corpus-echo rule. The instructive case: predicting the Maths Ext 1 factor-and-remainder question, DeepSeek and Grok independently emitted the identical invented cubic — x³ + ax² + bx − 12, factor (x+1), P(2) = −18. Those constants appear nowhere in the evidence pack. Two models inventing the same numbers isn't two independent minds converging; it's the same textbook exercise sitting in both training corpora. Under the rule, agreement on surface details not derivable from the pack counts as one vote for the question type, never as independent confirmation — and the sample paper used fresh constants.
External-knowledge anchors. Where a model cites real legislation, cases or statistics from its own knowledge rather than the pack (Legal Studies and Economics especially), the claim is flagged for human verification before it ships as fact, and the site marks it “pending human verification” until a human has checked it.

5 · The backtest: we sat the test first

Before predicting 2026, we hid the 2025 papers from the panel, had it predict them from 2019–2024 evidence only, then scored every prediction against the real papers using the official marking-guideline mapping. Two pilot subjects: Chemistry and Maths Extension 1.

Topic level saturated, as expected: every topic the panel rated ≥90% appeared — 37 of 37 calls across both subjects. That validates the pipeline but can't separate models. Question level is the honest test, graded strictly — the form must match, not just the topic:

Maths Extension 1 (2025): one dot per question-level prediction — hit, partial, grey miss; the figure at right is the weighted hit rate (hits + ½·partials) / predictions. Hover any dot for the prediction and what actually appeared.

Chemistry (2025): one dot per question-level prediction — hit, partial, grey miss; the figure at right is the weighted hit rate (hits + ½·partials) / predictions. Hover any dot for the prediction and what actually appeared.

Roughly half of all question-level predictions recognisably appeared — against a question space where naive guessing lands far lower. The stories behind the dots: five of six models called the 7-mark chemistry back-titration almost exactly; Fable's 3-mark series induction “whose step needs factorisation, not expansion” landed verbatim. And the misses teach the same lesson every time: all six models staked the chemistry flagship on identify-the-unknown-from-spectra — NESA inverted it (given the structure, predict the spectra). Content bullseye, format twist. That's why every subject page pushes skill chains over memorised questions.

The caveats, verbatim:

The 2025 papers may appear in the models' training data — leakage-guarded packs reduce but cannot eliminate this, so backtest numbers are a ceiling, not a guarantee.
Two pilot subjects only; the other thirteen inherit the method, not the measured hit rate.
November 2026 — predictions locked before the papers exist — is the true test.

The full backtest verdict (calibration tables, per-model Brier scores)

Holdout backtest verdict — predicting 2025 blind, scored against the real papers

Six models predicted the 2025 HSC Chemistry and Maths Ext 1 exams from 2019/2020–2024 evidence only (leakage-guarded packs verified free of 2025/2026 content), then were scored against the actual papers via the official marking-guideline mapping grids.

Calibration (Brier scores; lower is better)

Topic-examined is saturated — both 2025 papers touched every taxonomy topic; all models' p≥0.9 calls hit (37/37 across the panel). This dimension validates the pipeline but cannot separate models.

Topic-substantial (≥4-mark item) separates them:

model	Chemistry Brier(subst)	Ext 1 Brier(subst)
gpt-5.6-sol	0.179	0.325
opus	0.192	0.187
grok	0.200	0.259
deepseek	0.217	0.389
gemini-3.1-pro	0.251	0.512
fable	0.281	0.285
in-sample base-rate bar	0.248	0.16

Chemistry: four of six beat the bar. Ext 1: nobody beat it — the paper concentrated its ≥4-mark questions in just two topics (vectors, binomial) and the whole panel over-predicted breadth. Opus is the most consistently calibrated; the panel as a whole runs overconfident on "substantial".

Question-level hits (strict grading: form must match, not just topic)

Weighted hit rate = (hits + 0.5·partials) / predictions.

model	Chemistry	Ext 1
gpt-5.6-sol	0.600	0.571
fable	0.483	0.648
grok	0.500	0.558
gemini-3.1-pro	0.444	0.528
opus	0.438	0.517
deepseek	0.439	0.375

Roughly half of all question-level predictions recognisably appeared — against a question space where naive guessing would land far lower.

The stories

Biggest collective hit: the 7-mark chemistry back-titration (Q33) — five of six called it, DeepSeek nearly verbatim.
Sharpest single hit: Fable's Q12 3-mark series induction "whose step needs factorisation not expansion" — Q12(c) (Σk·k! = (n+1)!−1) is exactly that; its p=0.4 tan-addition/polynomial-roots fusion landed almost verbatim as Q14(e).
Most instructive partial: all six staked their chemistry flagship on identify-the-unknown-from-spectra; NESA inverted it (Q36: given the structure, predict the spectra). Content bullseye, format twist.
Biggest collective misses: the ICE-table Keq calculation (6/6 predicted, absent) and, in Ext 1, two broken streaks — pigeonhole and the vector-geometry proof (5/6 each, neither appeared).
Format migration tax: several content-correct calls were downgraded because questions moved to multiple choice (|2x−3|≥5 → MC Q1; cos2x+sinx=0 → MC Q5; the substitution integral shrank to transform-only MC Q3).
DeepSeek's breadth backfired in Ext 1: 18 of 36 predictions missed, many in off-taxonomy topics it invented.

What this means

The method's strength is what gets examined and which question families recur; NESA's freedom lives in format twists and streak-breaking — exactly the uncertainty the probabilities express. Marketing claims should be calibration-shaped, not oracle-shaped.
The ensemble is justified: different models win different dimensions (Sol: chemistry calibration + hits; Opus: substantial calibration; Fable: question-level sharpness in maths, but overconfident on breadth).
2026 implications now carry backtest support: the spectra capstone "reversion" call and the Keq "due after resting" call are strengthened by 2025's observed absences/inversions.
Caveats, stated plainly: the 2025 papers may sit in these models' training data (instructed-against but unverifiable), the in-sample base-rate bar is itself hindsight, and one year × two subjects is a small sample. The November 2026 scoring — truly out-of-sample — is the real test, and this harness is exactly what will run it.

6 · What each model is good at

The backtest is why this is an ensemble, not a leaderboard. Claude Opus 5 was the most consistently calibrated; GPT-5.6 Sol had the best chemistry hit rate; Claude Fable 5 the best Ext 1 hit rate and the sharpest single call; DeepSeek V4 nailed the back-titration nearly verbatim while grading lowest overall; and the panel as a whole runs overconfident on “substantial” topic coverage. No single model earns your trust alone — the agreement structure is the product, which is why every style card shows you how many models expect it, not just that we do.

7 · The commitment

These predictions are public and date-stamped — published and date-stamped Aug 2026, before the exams. In November 2026 we score every subject publicly against the real papers: per-model Brier scores and question-level hit rates, the same harness as the 2025 backtest. The method is public, and the metered compute for all 90 runs plus the backtest came to about US$9 — the work is in the harness and the honesty rules, not the tokens.

Independent work by Intuition Education from publicly available NESA documents. Not affiliated with or endorsed by NESA. No student data was used.