---
license: cc-by-4.0
language: [en]
tags: [interpretability, activation-oracles, knowledge-cutoff, percepts, nla]
size_categories: [100K<n<1M]
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: heldout_entity
    path: data/heldout_entity.parquet
  - split: test
    path: data/test.parquet
---

# Chronopercept (merged v1 + v2 + v3)

Union of [chronopercept-v1](https://huggingface.co/datasets/cds-jb/chronopercept-v1) (4,547 curated,
twin-certified scenes), [chronopercept-v2](https://huggingface.co/datasets/cds-jb/chronopercept-v2)
(116,452 generated scenes), and chronopercept-v3 ([cds-jb/chronopercept-v3](https://huggingface.co/datasets/cds-jb/chronopercept-v3),
660,223 generated scenes): **781,222 rows, 101,605 entities, 16 L0 task types
(14 generated mechanisms + 2 v1 control classes), 903 L1 domains.**
Splits are ENTITY-disjoint across the whole union — colliding entities adopt the earlier
version's split assignment (verified 0 entities span splits).

Each scene is 1–2 sentences of plain pre-1931 English in which an entity appears innocently:
a reader in 1930 finds the text unremarkable, while a well-informed modern reader perceives a
decisive second meaning (the "percept") that requires post-1931 knowledge.

## v3 additions (2026-08-08)

- **5 new L0 mechanisms** (Fable-agent taxonomy expansion): `retrospective_unmasking`,
  `fateful_conjunction`, `standing_mystery_resolved`, `latent_utility`, `legality_flip`.
- **625 new L1 domains** across all 14 generated L0s (Fable agents, each domain proven mineable
  with 3 real examples; global dedup; capped at 60 L1/L0).
- **`solvable_from_year` column (all rows)** — the smallest year Y such that a well-informed reader
  in year Y could already know the fact(s) needed to perceive the percept. Judged per-scene at QC
  for v3 rows and backfilled for all v1/v2 rows via a dedicated Sonnet 5 batch pass.
  `solvable_pre1931=True` (7,649 rows, ~1%) flags rows judged solvable before 1931 — these violate
  the corpus premise and should be filtered for training.
- **New entity-verification gate**: mined entities are batch-judged (real? evidence anchor?
  solvable_from_year >= 1931?) before scene generation; 85% pass rate. `evidence` column carries
  the anchor fact for v3 rows.
- **Decade flattening**: entity mining was decade-steered (2 of 6 calls per domain target
  1955-1990 / 1990-2025); merged median `solvable_from_year` is 1960 (10th-90th pct: 1935-2000),
  vs. the strong 1931-1945 skew of v1/v2.
- Scenes: rotating per-request few-shot from v1 (v2 used 2 fixed examples). QC keep rate 94%.

![composition](fig_v3_composition.png)

Columns: `scene`, `gold_percept` (modern-reader insight), `gold_secret` (the insight re-expressed
within a strict 1930 knowledge horizon), `secret_atoms`, `implication_class`, `intended_valence`,
`L0`/`L1`/`L2_entity`, `split`, `source` (v1|v2|v3), `solvable_from_year`, `solvable_pre1931`,
`twin_certified` (True only for v1 rows — generated rows are QC-screened but not certified against
the talkie twins; invented entities may carry no real knowledge gap), plus `gloss`/`year` (v2/v3)
and `evidence` (v3).

Generation pipeline: `code/chrono_v3_gen.py` (taxonomy merge from Fable-agent JSON, entity mining
with per-pair avoid-lists, verification gate, scene generation, QC, backfill, assemble, publish —
Sonnet 5 Message Batches throughout).