idirectships/abacus-cheat-tell-eval-v2

ABACUS Cheat-Tell Eval v2 — Surgical Anachronism

Why v2 Exists

The W7.2 v1 classifier (idirectships/abacus-cheat-tell-v1) achieved perfect accuracy (1.0/1.0/1.0) on a degenerate task. Its negatives were FineWeb-Edu post-1930 text — stylistically obvious compared to pre_modern TEAs. The classifier learned modernity detection (modern vocab → anachronism), not knowledge leakage detection.

The W10 AGI verification verdict requires a classifier that can detect a pre-modern reasoning model occasionally injecting a post-1930 scholarly reference into otherwise period-faithful prose. That is a fundamentally harder task.

Design: Surgical Anachronism in Synthesized Prose

• 200 rows total: 100 authentic (label=0) + 100 anachronism (label=1)
• Source: idirectships/abacus-tea-text-v0.3 — 487 pre_modern TEAs across traditions
• Stratified: chinese / greek / islamic / vedic / babylonian / other
• Passages: ~400–800 token prose synthesized from TEA metadata (person, topics, tradition)

Note on prose synthesis: The abacus-tea-text-v0.3 dataset stores TEA descriptor headers (Person/Difficulty/Topics metadata, 115–252 chars). To produce actual mathematical prose suitable for anachronism insertion, each passage is synthesized from tradition-specific templates populated with the TEA's person name and topics. Templates are modeled on actual pre-modern scholarly translation register (Islamic algebra commentary, Greek geometrical demonstration, Vedic series expansion notation, Chinese counting-board method, Babylonian tablet procedure).

Each anachronism row has one surgical insertion:

• A single em-dash parenthetical clause inserted mid-passage (not appended at end)
• A post-1930 mathematical/historical reference — not modern vocabulary like "AI" or "neural network"
• 20 distinct famous results in the pool: Gödel (1931), Bourbaki (1935), Cohen (1963), Grothendieck (1960s), Langlands (1967), Wiles (1995), Perelman (2003), Green-Tao (2004), Tao (2008), category theory (1945), Ramsey (1930), ZFC (1930), etc.
• Style-matched: insertion uses scholarly hedged language consistent with translated academic prose

What Makes This Hard

The insertions reference mathematical figures and results that could plausibly appear in a period-faithful reasoning chain about mathematics. They are not obvious via vocabulary scan. A naive N-gram or vocabulary detector would fail — it requires understanding that the referenced result or person postdates the document's claimed epistemic horizon.

Schema

Field	Type	Description
row_id	string	ct-eval-v2-{0001..0200}
source_tea_id	string	Tea ID from abacus-tea-text-v0.3
tradition	string	chinese / greek / islamic / vedic / math / other
label	string	"authentic" or "anachronism"
label_id	int	0=authentic, 1=anachronism
text	string	400–800 token passage (with insertion for label=1)
inserted_anachronism	string or null	Exact insertion text
insertion_position	int or null	Character offset of insertion
original_text_hash	string	SHA-256 of unchanged passage

Splits

Split	Authentic	Anachronism	Total
train	80	80	160
eval	20	20	40

Anachronism Reference Pool (20 distinct results)

Gödel incompleteness (1931), Bourbaki formalism (1935), Bourbaki structure (1939), Cohen forcing (1963), Grothendieck schemes (1960), Grothendieck topos (1963), Langlands program (1967, 1970), Wiles FLT proof (1995), Perelman Poincaré (2003), Perelman geometrization (2003), Green-Tao theorem (2004), Tao structure/randomness (2008), ZFC axioms (1930), ZFC choice (1930), category theory — Mac Lane & Eilenberg (1945), Ramsey theory (1930).

Usage

from datasets import load_dataset
ds = load_dataset("idirectships/abacus-cheat-tell-eval-v2")

License

Apache-2.0 — consistent with the ABACUS codebase license.