What the first feedback snapshot tells us
Evidence cutoff: 2026-09-26 18:45:09 UTC / 20:45:09 Warsaw. Analysis uses the frozen snapshot 02 report and its hashed input manifest. Grading continued during inspection; this report deliberately stays at 37/40 exams. Model: Gemma 4 12B Q4, thinking off, temperature 0, seed 42. Judge: Astra, medium. This is one solver configuration, not a comparison of models.
Measured results
Of 1,289 items, 1,268 have a score: 643 full-credit, 97 partial-credit and 528 zero-credit. The assessed score is 1,011/1,910 (52.93%). There are 899 lost assessed points. Twenty-one items remain unassessed: 17 tagged rubric ambiguity and four infrastructure failures. Their combined maximum is 40 points. Three exams are absent from this snapshot. Neither unknown items nor missing exams have been treated as zero.
| Existing judge tag | Scored failures | Share of the 625 scored failures | Associated lost points |
|---|---|---|---|
| Knowledge | 530 | 84.8% | 778 |
| Source reading | 337 | 53.9% | 459 |
| Reasoning | 189 | 30.2% | 430 |
| Instruction following | 57 | 9.1% | 123 |
Labels overlap. These are descriptions assigned by the same judge that produced the grades, not independent causal findings. A knowledge tag does not establish that memorizing more facts will fix the answer: source selection or a misleading interpretation may produce the same wrong identity.
Essays deserve a dedicated workstream: all 36 assessed essays lost points, scoring 222/494 (44.94%) and losing 272 points, 30.3% of all assessed loss. The remaining essay is unassessed. Across the inspected essay rationales, the recurring pattern is fluent but general prose with too few specific events, weak chronological development, shallow coverage of requested dimensions, and occasional factual errors. Length compliance also fails in some cases, but is not the general explanation. For example, the 2018 May old-format essay scores 2/20 and omits the concrete events needed to develop the Chrobry/Otto III relationship; the 2023 May essay exceeds 300 words but still scores only 4/15.
Actionable mechanisms seen in examples
These examples illustrate mechanisms; their prevalence has not yet been manually annotated.
Item IDs refer to full records in snapshot 02 items.jsonl.
| Candidate mechanism | Evidence example | What to investigate |
|---|---|---|
| Wrong historical identity with confident elaboration | 2017 June, 6: Sobieski/1686 instead of Kazimierz Jagiellończyk and the Thirteen Years' War | Fact retrieval versus use of discriminating source clues |
| Evidence noticed, then explained away | 2017 May, 2: answer observes a ship's bow but chooses the land battle of Thermopylae; judge expects Salamis | Evidence-to-conclusion consistency |
| Required answer elements missing | 2017 June, 5.2: Kraków and Wrocław given, Kołobrzeg omitted; all three required | Completeness checking and exact rubric conditions |
| Repetition or paraphrase instead of reasoning | 2019 Nowa Era, 24.1: names Operation Tempest but repeats the order instead of explaining why not to fight the USSR | Causal explanation versus restating the source |
| Incorrect chart interpretation and non-distinct arguments | 2019 Nowa Era, 22: regional shares of industry employment treated as a region's industrial composition; two arguments repeat one claim | Denominator reading and independent evidence |
| Answer outside an explicit time boundary | 2019 May, 20.1: supplies 1806 for a question asking about the late eighteenth century | Constraint checking |
| Wrong source selection | 2017 May old-format, 25: compares E/F instead of H/I; 2019 May old-format, 23: compares A/B instead of F/G | Audit source-to-question packaging before attributing this to the model |
Visual interpretations above are reported judge findings; this analysis has not independently regraded the original images. Short-answer quotations and source/answer relationships were inspected directly.
Input and evaluation quality need their own lane
The wrong-source examples have a concrete confound confirmed in the saved solver prompts:
2017 May old-format item 25 receives sources A–J together, and 2019 May old-format item 23 receives
A–I. The question text does not explicitly bind the required pair. Several source-specific instructions
appear inside the aggregated source text without their original question adjacency. Current
harness/matura.py:128 serializes that entire source_text into the single-item prompt. The answer
chooses a wrong pair, but this evidence is insufficient to assign the whole error to instruction following.
The next step is an original-page mapping audit and a controlled rerun with explicit source binding;
do not silently alter existing grades or prompts.
There are also 130 exact question/source-text duplicate groups across this snapshot. This is only a text match: images can differ, and paraphrases are missed. Exam counts therefore overstate the amount of independent task evidence. Resolve exam families and near-duplicates before building train/holdout splits or estimating uncertainty. Current percentages are descriptive item/point aggregates; no independence-based confidence interval or model superiority claim has been made.
Pipeline now available
saved attempts + validated grades → versioned joined item records → normalized tags and loss tables → evidence-linked review queues → confirmed mechanisms → paired probes → regression and training splits
The implemented stage ends at the review queues: 609 model-diagnosis cases, 40 grade-adjudication cases and four infrastructure cases. All annotations are pending. Original tags, scores, complete answers, rubric text, evidence paths and model identities remain available locally. Human-reviewed mechanisms, cross-model comparisons and diagnostic model reruns are proposed next steps, not completed work.
Start by auditing the source-binding examples and resolving the 21 unknown items. In parallel with that work, review a small stratified set of failures and successful controls, including essays, before expanding the taxonomy. Prioritize essay specificity, source grounding and answer completeness after the review confirms each mechanism. Do not train on these evaluation answers and then report gains on the same papers as held-out improvement.
Five offline tests passed. The production run freshly revalidated every selected exam's input hashes, quotation evidence, coverage and score arithmetic. Semantic grading correctness remains provisional. Aggregate metrics and an aggregate-only report were published to ClearML. Detailed evidence and review records remain local.