Year 12 · 55 minutes · Synthetic evaluation and selection
Select candidates using a proxy reward, then audit those same selected answers against independent labels. Vary candidate quality, pool size, audit size and the proxy rule.
Student paper journal · Interactive notebook · Editable Python
Try the starting controls, save two runs and reopen a downloaded project. Print the student journal before class. If Python cannot load, use the supplied example below for a prediction and evidence critique; rerunning changed inputs requires the live or native notebook. Do not present paper discussion as a new model experiment. Keep an individual record even when you investigate in pairs. Record the exact settings, source and data split.
Start with 5–8 minutes of prediction, use about half the lesson for paired comparisons, then reserve at least 10 minutes for a counterexample and an individual explanation. Selection and the uncertainty interval now belong to the same experiment. Compare paired baselines on the same generated pools. The cue changes a known scoring rule; it does not measure a model’s mental state.
Settings: {'sample': 200, 'failure': 0.01, 'best': 8, 'cue': 0, 'seed': 43}
The selected policy fails on 0/200 independently generated candidate pools. Wilson 95% interval [0.000%, 1.885%]. Compare with the paired first-candidate baseline.
| condition | correct | failures | n |
|---|---|---|---|
| First candidate | 197 | 3 | 200 |
| Highest proxy reward | 200 | 0 | 200 |
The table shows up to eight rows. Inspect the notebook for all values, controls and denominators.
Method and scope: Generate N independent candidate pools using a known per-candidate failure probability. Select by truth+style or style alone, then audit the selected candidates against their independent labels. The interval applies to selected-policy failure under this simulator, not to the input candidate failure rate or an LLM. After tuning on a seed, use a fresh seed for final evaluation. A zero-failure sample cannot certify safety.
Assess four criteria, each 0–2: testable prediction; comparison identifying what stayed fixed; accurate use of exact evidence; counterexample and limited conclusion. 0 means absent or contradicted by the record, 1 means partly supported, and 2 means clear and supported. A surprising result earns no penalty.
Beginning response: ‘The picture looks right.’ Ask for a particular case. Developing response: names one number without its setting. Ask what comparison supports it. Secure response: states the setting and measured change and separates the notebook's result from the source paper's claim. Extending response: authors or reserves a new case, tests an alternative explanation, and revises the claim if needed. These are marking examples, not pupil data.
Extension: Repeat paired seeds, add a third unseen proxy condition, and pre-register the acceptance threshold.
System and data: Engineering perspective; examples of safety infrastructure
Method: Reason through deployment, testing and operational failure modes; no new controlled model experiment.
Task: Audit data provenance, monitoring and reproducibility alongside model metrics. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: An engineering argument, not experimental proof of model safety.
System and data: Research funding programme
Method: Describes eligibility and support; assess proposed experiments separately from the announcement.
Task: Write a bounded research proposal with a falsifiable hypothesis and compute budget. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: A grants announcement supplies opportunities, not a new scientific result.
System and data: Company funding and strategy
Method: Reports financing and plans; investment is not a model evaluation.
Task: Turn a product claim into an operational definition and an independent test. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: Funding is not validation. The older Ember service is deprecated.
System and data: Research fellowship programme
Method: Describes research directions and participation; does not test a scientific hypothesis.
Task: Design a collaborative experiment with reviewable code and explicit ownership. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: A fellowship announcement is careers context, not a research finding.
System and data: Genomic research collaboration
Method: Describes planned interpretation of biological model representations; requires later independent scientific validation.
Task: Separate exploratory biomarkers, external validation and clinical use. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: A collaboration announcement is not clinical validation. No pupil health data are used.
System and data: Policies for interpretability tools
Method: Describes moderation, feature access and research processes; organisational policy and model capability are separate claims.
Task: Separate the tool's access policy from evidence about the underlying model's behaviour. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: A stated safety process is not proof that all failure modes are covered.
System and data: Company financing and intentional-design strategy
Method: Reports funding and proposed applications; follow the underlying experiments for empirical evidence.
Task: Construct a claim–evidence table for research and product statements. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: A company announcement provides strategy and context, not an independent benchmark.
System and data: Materials-discovery partnership
Method: Announces collaboration; a predicted material must still be independently evaluated.
Task: Separate a predicted material property from a measured property. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: The announcement is not a materials experiment or an LLM result.
System and data: Organisational security-control audit
Method: Reports independent control assurance over an audit period; does not evaluate the truth of generated answers.
Task: Match assurance evidence to the precise system and claim it covers. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: SOC 2 is not a certificate of LLM truthfulness or absence of harmful behaviour.
System and data: Scribe experiments on Evo 2 and GPT-2 examples
Method: Reports agent-assisted research tasks and failure cases; provenance checks are needed when agents can shortcut work.
Task: Audit notebook state, generated code, shortcuts and result provenance. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: The shared task suite is described as directional, not a rigorously audited benchmark.
System and data: Kimi K2.5 and Qwen3-32B; Fortress/Petri evaluation settings compared with WildChat conversations
Method: At sentence boundaries, scores a prespecified evaluation-awareness phrase from token probabilities and compares the monitor with explicit verbalisation and judge-based monitoring. Reported rollout savings are specific to these experiments; the score is a proxy for awareness, not a direct mental-state measurement.
Task: Distinguish monitor discrimination, calibration and a causal claim about awareness. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: Phrase choice and prompt framing matter. High AUROC does not establish intent or universal reliability.
System and data: Qwen3-4B and an abliterated variant on HarmBench, with a 100,000-rollout reference
Method: Interpolates the paired models in logit space, measures compliance along the path, fits an empirical log-linear trend below 50% compliance and extrapolates to the original model. Requires a related variant and a suitable trend; the classroom Wilson interval uses a different, direct-sampling method.
Task: Study sampling intervals and distinguish extrapolation assumptions from directly observed failures. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: The reported efficiency is setting-dependent. Our binomial experiment explains uncertainty; it does not implement the paper's extrapolation method.
System and data: Gemma-3-12B-IT and LongFact++ with 999 held-out prompts
Method: Trains factuality/correction probes and uses a frozen model for rewards; compares RL, inline intervention and best-of-N contributions with independent labels.
Task: Vary best-of-N selection under an imperfect reward and test against independent truth labels. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: The paper combines methods; its headline reduction is not attributable to feature rewards alone. Our reward pool is synthetic.
System and data: Band-gap-conditioned MatterGen diffusion
Method: Uses an activation probe to accept or reject proposed denoising steps and evaluates targeting, stability, uniqueness and novelty.
Task: Examine selection pressure, proxy errors and external property validation. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: A model's predicted material property is not a physical measurement; this is not text-LLM research.
System and data: Eight models, nineteen benchmarks; causal work on Kimi K2.5/Fortress
Method: Manually verifies verbalised awareness, compares matched cues and performs interventions; broad correlations and narrower causal evidence have different scope.
Task: Use matched conditions, independent labels and explicit limits on causal claims. Record the source model/task, a baseline, the changed factor, a measurement and an alternative explanation. State exactly which part your notebook investigates and which part it does not reproduce.
Boundary: Correlations span more models than the causal intervention study. Silence about a test does not prove absence of awareness.
Independent Brightlab adaptations; no Goodfire endorsement or full reproduction claim. Source methods reviewed 7 September 2026. Curriculum connections are selected planning links; confirm your school syllabus.