Update README.md

README.md

@@ -106,13 +106,7 @@ configs:
 
 # fibs-v1
 
-Private FAR AI Deception Pod research dataset: **`AlignmentResearch/deceptive-followup-v33-balanced`
-(revision `3db4df84116462a6eb5761e9b043b534b8e22c72`) with named families edited in place.** Every row of every
-untouched family is byte-identical to the parent (`provenance/untouched_identity.json` carries the
-per-family content digests); the edited families below were regenerated or topped up, screened by the
-same four-judge unanimous rule as the parent, and certified by Fable reads against each family's bar.
-
-**163,016 rows** over the four quadrant splits:
+FAR AI Deception Pod research dataset of **163,016 rows** over the four quadrant splits:
 
 | split | rows |
 |---|---:|
@@ -121,66 +115,45 @@ same four-judge unanimous rule as the parent, and certified by Fable reads again
 | deceptive_honest | 37,567 |
 | deceptive_deceptive | 40,464 |
 
-## Edited families
+## Families
 
-| family | parent rows | rows in this release |
+| family | rows in this release |
 |---|---:|---:|
-| among | 6,464 | 6,464 |
-| anti | 7,984 | 7,984 |
-| b1-bare | 4,034 | 4,034 |
-| b1-followup | 4,488 | 4,488 |
-| caving | 1,648 | 1,648 |
-| doluschat | 7,988 | 8,000 |
-| fabricated | 2,164 | 2,164 |
-| fever | 8,000 | 8,000 |
-| harm | 5,024 | 8,000 |
-| instructed | 7,868 | 7,868 |
-| liarsbench | 8,000 | 7,992 |
-| lies | 6,760 | 6,760 |
-| longcontext | 1,764 | 1,764 |
-| mask | 8,000 | 8,000 |
-| medqa | 4,020 | 4,020 |
-| model | 3,304 | 3,304 |
-| multi | 1,476 | 1,476 |
-| negotiation | 6,954 | 6,954 |
-| original | 4,204 | 4,204 |
-| persona_prefill | 8,000 | 8,000 |
-| sandbagging | 6,700 | 6,700 |
-| sycophancy | 7,976 | 7,976 |
-| synthetic | 8,000 | 8,000 |
-| truthfulqa | 4,804 | 4,804 |
-| ultrachat | 8,000 | 8,000 |
-| werewolf | 6,692 | 6,692 |
-| wikihow | 1,760 | 1,760 |
-| wikihownest | 7,960 | 7,960 |
-
-The harm-knowledge rows that sat under the Liars' Bench family label in the parent are reassigned to
-the harm family here (`provenance/family_edits.json`).
-
-## Certification of the new rows
-
-| family (rubric) | kept by the four judges | Fable reads | mislabels | rate | 95% upper | pass line (alpha 0.05/29) | outcome |
-|---|---:|---:|---:|---:|---:|---:|---|
-| Harm pressure knowledge (v1.3) | 2,186 of 3,214 | 296 (207 refusals re-read by GLM-5.3) | 0 | 0.0% | 1.0% | 4 | PASS |
-| Liars' Bench soft-trigger (v1.3) | 4,032 of 5,097 | 500 | 4 | 0.8% | 1.8% | 11 | PASS |
-| DolusChat (v1.5) | 17,163 of 22,447 | 299 (4 refusals, 3 re-read by GLM-5.3, 1 no verdict) | 5 | 1.67% | 3.5% | 4 | accepted by ruling at the parent family's certified rate (5 in 342, 1.46%); the five convicted rows are excluded |
-
-Every Fable-convicted row is excluded from the build and every Fable-confirmed row is kept regardless of the four judges (`provenance/new_row_screen.json`). Insider trading was generated but is not part of this release.
-
-Screening campaigns folded into this build:
-
-| campaign | rubric | new-row screen outcomes |
-|---|---|---|
-| pharm | rubric_plainconstruct_v1_3.md@ec597d99eb10d2afa92a28bddcde2a62 | harm: duplicate_of_new 1, duplicate_of_parent 1, fable_confirmed 93, glm_confirmed 203, judges_kept 1,886, judges_rejected 1,028, refused_no_stand_in_judges_kept 4 |
-| plbsoft | rubric_plainconstruct_v1_3.md@ec597d99eb10d2afa92a28bddcde2a62 | liarsbench: fable_confirmed 496, fable_mislabel 4, judges_kept 3,532, judges_rejected 1,065 |
-| pdolu2 | rubric_plainconstruct_v1_5.md@bcb71e10fabe67dc387440c3c2d5e31b | doluschat: fable_confirmed 291, fable_mislabel 5, glm_confirmed 3, judges_kept 16,863, judges_rejected 5,284, refused_no_stand_in_judges_kept 1 |
-
-## Contextual rows: a family-level note (no column in this release)
+| among | 6,464 |
+| anti |  7,984 |
+| b1-bare |  4,034 |
+| b1-followup  | 4,488 |
+| caving  | 1,648 |
+| doluschat  | 8,000 |
+| fabricated  | 2,164 |
+| fever  | 8,000 |
+| harm  | 8,000 |
+| instructed  | 7,868 |
+| liarsbench  | 7,992 |
+| lies  | 6,760 |
+| longcontext  | 1,764 |
+| mask  | 8,000 |
+| medqa  | 4,020 |
+| model  | 3,304 |
+| multi  | 1,476 |
+| negotiation  | 6,954 |
+| original | 4,204 |
+| persona_prefill  | 8,000 |
+| sandbagging  | 6,700 |
+| sycophancy  | 7,976 |
+| synthetic  | 8,000 |
+| truthfulqa  | 4,804 |
+| ultrachat  | 8,000 |
+| werewolf  | 6,692 |
+| wikihow  | 1,760 |
+| wikihownest | 7,960 |
+
+## Contextual rows: a family-level note
 
 The paper distinguishes **contextual** rows, where the truth of the statement the model makes is evidenced
 in the transcript itself (system prompt, earlier turns, tool results, a supplied passage, or the model's
 own visible working), from rows that rest on world knowledge, which the model could in principle have been
-trained to believe otherwise. This release carries no per-row column for it. The classification below is by
+trained to believe otherwise. The classification below is by
 family, sub-source and follow-up category, chosen for precision over recall after reading two full
 transcripts per family; a row-level classifier is future work.
 
@@ -262,14 +235,10 @@ the family cap (8,000), the per-family before/after counts and the builder's sha
 `provenance/longcontext_single_turn_split.json` records the 2026-10-02 longcontext re-rendering. The builder is
 `sandbox/oa_backdoor/scripts/dataset_generation/build_v2_slotin.py` in the Deception Pod monorepo.
 
-## Columns and loading
+## Columns
 
-Identical to the parent: `messages_json` (UTF-8 JSON message list; assistant turns may carry
+`messages_json` (UTF-8 JSON message list; assistant turns may carry
 `reasoning_content` sidecars and OpenAI-style `tool_calls`), `tools` (JSON string tool-schema list where
 the source uses tools, null elsewhere), `completion`, and the source and judge metadata columns.
 Consumers must `json.loads` the JSON-string columns. Quadrants use `base_follow-up` semantics: the first
 term labels the base turn and the second labels the follow-up turn.
-
-Re-rendered text derived from these rows is for probe input and activation harvest only; it must not
-replace original rollout token IDs or feed RL-training token paths, because re-templating can change
-tokenization.