Remove Phase 2 planning notes (kept on GitHub; HF holds artifacts only)

phase2_scaling/phase2_final_plan.md

@@ -1,419 +0,0 @@
-# Phase 2 Final Plan: Scaling of Cross-Relation Truth Representations
-
-**Date:** 2026-07-11  
-**Time budget:** approximately 1.5 days  
-**Purpose:** implementation and paper-writing specification
-
-## 1. Research question
-
-Phase 1 showed that breaking cross-pair truth co-occurrence removes truth decodability in the tested fully trainable synthetic model while preserving factual memorization.
-
-Phase 2 asks:
-
-> Does model scale make a linearly decodable truth representation more transferable across factual relations?
-
-The main comparison is Llama-3.1-8B versus Llama-3.1-70B. Layer-wise probing is a measurement tool; the contribution is cross-relation generalization, not the location of the single strongest layer.
-
-The final claim remains scoped to the evaluated CounterFact relations.
-
-## 2. Fixed scope
-
-### Models
-
-- `meta-llama/Llama-3.1-8B`
-- `meta-llama/Llama-3.1-70B`
-
-Use base checkpoints for both models and record exact revisions. Llama-3.1-405B is optional future work and does not block this experiment.
-
-### Dataset
-
-- 6 semantically diverse CounterFact relations
-- final sample size selected from 50, 75, or 100 factual records per relation
-- one matched true/false pair per record
-- 600, 900, or 1,200 sentences in total
-
-Select relations before probing. Each relation must contain at least 50 usable records and should contain 100 if possible. Generate false examples by replacing the true attribute with a different attribute from the same relation using a deterministic permutation and fixed seed.
-
-Each processed example preserves:
-
-```text
-example_id
-pair_id
-case_id
-relation_id
-subject
-template
-true_attribute
-used_attribute
-sentence
-label
-false_attribute_source_id
-```
-
-Prepare nested datasets with a fixed ordering:
-
-```text
-N50  = first 50 pairs per relation
-N75  = N50 + the next 25 pairs per relation
-N100 = N75 + the next 25 pairs per relation
-```
-
-The final selected subset and ordering are used for both models.
-
-### Sample-size timing gate
-
-The final sample size is selected using a disposable pilot before any full six-relation extraction.
-
-Choose one of the six relations and take its first 50 fact pairs, giving 100 pilot sentences. Run the exact four-depth activation extraction on this pilot subset for both 8B and 70B. The pilot must use the same batch size, representation position, dtype, NNsight code path, and output format intended for the final run.
-
-Record separately for each model:
-
-```text
-queue_wait_seconds
-remote_execution_seconds
-result_transfer_seconds
-total_wall_clock_seconds
-sentences_processed
-seconds_per_sentence
-```
-
-For a candidate size `N` pairs per relation, estimate full extraction time as:
-
-```text
-full_sentence_count = 6 relations * N pairs * 2 labels
-pilot_sentence_count = 100
-scaling_multiplier = full_sentence_count / 100
-```
-
-Therefore:
-
-| Candidate | Full sentences | Pilot-time multiplier per model |
-|---:|---:|---:|
-| 50 pairs/relation | 600 | 6x |
-| 75 pairs/relation | 900 | 9x |
-| 100 pairs/relation | 1,200 | 12x |
-
-Estimate 8B and 70B separately from their own pilot timings, then add them to estimate total active extraction time. Queue time is reported separately because it may not scale linearly.
-
-Choose `N_final` from 50, 75, or 100 using the remaining wall-clock budget. Reserve at least six hours after activation extraction for probe runs, validation, figures, and debugging. Save the decision and timing evidence in `sample_size_decision.json`.
-
-After validating tensor shapes, example ordering, saved metadata, and timing, discard the pilot activation files. Keep only the pilot logs and timing record. Once `N_final` is frozen, start a fresh activation extraction for all six relations on both models. The maximum wasted inference is one relation out of six at the 50-pair setting, and less proportionally if `N_final` is 75 or 100.
-
-## 3. Splits
-
-The true and false sentences from the same `pair_id` remain in the same fold.
-
-### Within-relation splits
-
-Create one fixed three-fold split independently inside each relation. Reuse it across all layers and models.
-
-### Leave-one-relation-out splits
-
-For each target relation, train on the other five complete relations and evaluate on the target relation.
-
-## 4. Activation extraction
-
-For each sentence, save the transformer-block output at the final subtoken of the target attribute.
-
-Extract four normalized depths:
-
-| Depth | 8B block index | 70B block index |
-|---:|---:|---:|
-| 25% | 7 | 19 |
-| 50% | 15 | 39 |
-| 75% | 23 | 59 |
-| 100% | 31 | 79 |
-
-Indices are zero-based. Compute them dynamically:
-
-```python
-layer_index = math.ceil(depth * model.config.num_hidden_layers) - 1
-```
-
-Save float16 activation matrices and convert them to float32 for probe training.
-
-Expected activation volume:
-
-| Pairs/relation | Sentences | Activation vectors | 8B + 70B storage |
-|---:|---:|---:|---:|
-| 50 | 600 | 4,800 | approximately 59 MB |
-| 75 | 900 | 7,200 | approximately 89 MB |
-| 100 | 1,200 | 9,600 | approximately 118 MB |
-
-Per-layer matrix shapes are `[12 * N_final, 4096]` for 8B and `[12 * N_final, 8192]` for 70B.
-
-Also save token position information, model ID and revision, layer index, normalized depth, activation shape, and dataset hash.
-
-## 5. Probe configuration
-
-Use one fixed scikit-learn pipeline throughout:
-
-```text
-StandardScaler
-L2-regularized LogisticRegression
-```
-
-Fix the solver, `C`, maximum iterations, and random seed in `experiment_config.json` before Stage 2.
-
-Primary metric: ROC-AUC.  
-Secondary metric: balanced accuracy.
-
-## 6. Stage 1A: Within-relation probing
-
-For every model, depth, and relation, run three-fold pair-grouped cross-validation using only that relation.
-
-For each fold:
-
-1. fit the probe on two folds;
-2. evaluate it on the held-out fold;
-3. record fold AUC and per-example prediction scores.
-
-The official within-relation result is:
-
-```text
-mean of the three fold AUCs +/- standard deviation
-```
-
-It is used for:
-
-- the within-relation baseline;
-- global layer selection for Stage 2;
-- the within-to-unseen generality gap;
-- the diagonal of the Stage 2 transfer matrix.
-
-Pooled OOF AUC may be stored as an additional diagnostic, but it is not the primary metric because scores from independently fitted fold models need not share an identical calibration scale.
-
-Training count:
-
-```text
-6 relations x 3 folds x 4 depths = 72 fits per model
-```
-
-## 7. Stage 1B: Leave-one-relation-out probing
-
-For every model and depth, run six experiments:
-
-```text
-train on five relations -> evaluate on the sixth unseen relation
-```
-
-Report one AUC for each held-out relation, followed by mean and variation across the six relations. Retain all four depths.
-
-Training count:
-
-```text
-6 held-out relations x 4 depths = 24 fits per model
-```
-
-### Generality gap
-
-For each model, depth, and target relation:
-
-```text
-generality_gap = mean_within_relation_auc - leave_one_out_auc
-```
-
-Interpretation:
-
-- high within and high leave-one-out AUC: relation-general decodability;
-- high within and low leave-one-out AUC: relation-specific decodability;
-- a smaller 70B gap: evidence that scaling improves transfer across the evaluated relations.
-
-## 8. Global layer selection for Stage 2
-
-Select one layer per model from Stage 1A:
-
-```text
-best_layer(model)
-    = layer with the highest mean within-relation AUC,
-      averaged first across folds and then across the six relations
-```
-
-If two depths differ by no more than 0.005 mean AUC, select the shallower depth.
-
-Selection is global per model, not per relation. The 8B and 70B matrices may use different normalized depths and are therefore results at each model's best within-relation layer.
-
-## 9. Stage 2: Six-by-six pairwise transfer matrix
-
-At each model's selected layer, fit one probe on all examples from each source relation:
-
-```text
-6 source relations = 6 full-source probes per model
-```
-
-Apply each probe to the other five relations:
-
-```text
-6 sources x 5 targets = 30 off-diagonal evaluations per model
-```
-
-Rows are source relations and columns are target relations. Transfer is directional, so `R1 -> R2` and `R2 -> R1` are separate results.
-
-For each diagonal cell, reuse the mean three-fold within-relation AUC from Stage 1A at the selected layer. The diagonal is cross-validated, whereas each off-diagonal probe is fitted on the full source relation; state this in the figure caption.
-
-Stage 2 explains whether transfer is broad, clustered among semantically related relations, or asymmetric.
-
-## 10. Compute summary
-
-| Quantity | Total after timing decision |
-|---|---:|
-| Sentences | 600, 900, or 1,200 |
-| Activation vectors | 4,800, 7,200, or 9,600 |
-| Activation storage | approximately 59, 89, or 118 MB |
-| Stage 1 within-relation fits | 144 across two models |
-| Stage 1 leave-one-out fits | 48 across two models |
-| Stage 2 full-source fits | 12 across two models |
-| Total logistic-regression fits | 204 |
-| Saved probe models | all 204 fitted probes |
-| Stage 2 off-diagonal evaluations | 60 across two models |
-
-Remote activation extraction is the main time cost. Probe training is local and inexpensive.
-
-## 11. Required artifacts
-
-```text
-reproduction/scaling/
-|-- phase2_final_plan.md
-|-- experiment_config.json
-|-- sample_size_decision.json
-|-- data/
-|   |-- raw/counterfact.json
-|   |-- processed/examples.parquet
-|   |-- processed/relations.json
-|   |-- processed/splits.parquet
-|   `-- processed/dataset_manifest.json
-|-- activations/
-|   |-- llama_3_1_8b/
-|   |   |-- manifest.json
-|   |   |-- sample_index.parquet
-|   |   `-- layer_*.npy
-|   `-- llama_3_1_70b/
-|       |-- manifest.json
-|       |-- sample_index.parquet
-|       `-- layer_*.npy
-|-- results/
-|   |-- stage1_predictions.parquet
-|   |-- stage1_metrics.csv
-|   |-- selected_layers.json
-|   |-- stage2_predictions.parquet
-|   |-- stage2_matrix_8b.csv
-|   |-- stage2_matrix_70b.csv
-|   `-- probe_weights/
-`-- figures/
-```
-
-### Data and split files
-
-`examples.parquet` contains sentence text, labels, relation and pair identity, subject, attributes, and false-attribute provenance.
-
-`splits.parquet` contains:
-
-```text
-example_id, pair_id, relation_id, within_relation_fold
-```
-
-### Activations
-
-After the pilot is discarded and `N_final` is frozen, store one float16 matrix per model and layer. `sample_index.parquet` defines row-to-example correspondence. The manifest records model/data revisions, timing, shapes, and checksums.
-
-### Predictions
-
-Store per-example scores:
-
-```text
-stage, model_id, layer_index, normalized_depth, protocol,
-source_relations, target_relation, fold, example_id,
-true_label, prediction_score
-```
-
-### Metrics
-
-```text
-stage, model_id, layer_index, normalized_depth, protocol,
-source_relations, target_relation, fold,
-auc, balanced_accuracy, n_train, n_test
-```
-
-### Probe weights
-
-Save all 204 fitted probes from Stage 1 and Stage 2. This includes every within-relation fold probe, every leave-one-relation-out probe, and every Stage 2 full-source probe.
-
-Store portable parameter arrays rather than relying only on pickled sklearn objects:
-
-```text
-coefficient
-intercept
-scaler_mean
-scaler_scale
-classes
-```
-
-Each saved probe also needs a manifest entry containing:
-
-```text
-probe_id
-stage
-protocol
-model_id
-model_revision
-layer_index
-normalized_depth
-source_relations
-target_relation
-fold
-training_example_ids
-probe_configuration_hash
-sklearn_version
-```
-
-The expected storage is only on the order of tens of megabytes when arrays are stored as float32. Saving every probe makes later prediction checks, coefficient analysis, and paper revisions possible without retraining.
-
-## 12. Required figures and tables
-
-1. **Stage 1 depth plot:** mean within-relation and leave-one-out AUC across the four depths for 8B and 70B.
-2. **Generality-gap summary:** within AUC, unseen AUC, and gap by model and depth.
-3. **Two transfer matrices:** one 6-by-6 heatmap per model using the same relation order and color scale.
-4. **Relation-level table:** within AUC, leave-one-out AUC, and gap for every target relation.
-
-Relations are the unit of generalization. Keep relation-level results visible alongside aggregate means.
-
-## 13. Implementation sequence
-
-1. Select and freeze six relations.
-2. Prepare and validate nested `N50`, `N75`, and `N100` subsets where data availability permits.
-3. Choose one pilot relation and run its first 50 pairs through all four depths on 8B and 70B.
-4. Validate the extraction and record model-specific timing components.
-5. Estimate the 50-, 75-, and 100-pair full-run costs and freeze `N_final`.
-6. Discard pilot activations while preserving logs and `sample_size_decision.json`.
-7. Generate fixed three-fold pair assignments for the final dataset.
-8. Start fresh four-depth activation extraction for all six relations on 8B and 70B.
-9. Run Stage 1 on both models.
-10. Select one global layer per model.
-11. Run Stage 2 and generate the two matrices.
-12. Produce final figures and freeze result files before interpretation.
-
-## 14. Completion criteria
-
-The experiment is complete when:
-
-- the timing benchmark and final sample-size decision are saved;
-- the balanced six-relation `N_final` dataset and fixed splits are saved;
-- four activation depths are available for both models;
-- within-relation and leave-one-out results exist for every model, depth, and relation;
-- selected layers were produced by the fixed mean-within-AUC rule;
-- both 6-by-6 transfer matrices are complete;
-- all 204 probe models, per-example predictions, metrics, manifests, and final figures are saved.
-
-## 15. Optional low-cost controls
-
-If time permits, add two local controls; neither requires additional activation extraction.
-
-### Text-only baseline
-
-Train a TF-IDF logistic classifier on the raw sentences using the same within-relation and leave-one-relation-out splits. Near-chance performance would show that the activation results are not easily explained by surface lexical cues alone.
-
-### Pair-balanced label permutation
-
-At each model's selected layer, repeat probe evaluation with randomly swapped true/false labels inside each fact pair, preserving class balance. Run approximately 10 random permutations. The resulting AUC distribution should be near 0.5 and serves as a pipeline/split sanity check.
-
-These controls are optional and must not delay the main 8B-versus-70B experiment.

phase2_scaling/phase2_notes.txt

@@ -1,842 +0,0 @@
-Phase 2: Scaling Experiment Notes
-=================================
-Authoritative specification: phase2_final_plan.md
-This file records decisions, reasoning, and data-quality observations
-for paper writing.  It does not duplicate the full protocol.
-
-
-SUPERSEDED: initial design (2026-07-11)
-----------------------------------------
-The initial plan (also in phase2_notes.txt prior to this revision and
-in the Phase 2 section of updated_plan.md) described:
-  - three models: 8B, 70B, 405B
-  - extraction at every transformer layer
-  - single relation P103 first, then top-25 relations
-  - research focus on locating the highest-AUC layer
-  - three normalised-depth AUC curves as the main figure
-
-This was superseded on 2026-07-11 by phase2_final_plan.md, which
-reframes the contribution around cross-relation generalisation rather
-than layer location.  The current design is:
-  - two models: Llama-3.1-8B and Llama-3.1-70B (405B is optional
-    future work and does not block this experiment)
-  - six finalised CounterFact relations
-  - four normalised depths: 25%, 50%, 75%, 100%
-  - Stage 1: within-relation 3-fold CV + leave-one-relation-out
-  - Stage 2: 6x6 pairwise transfer matrices at each model's best
-    within-relation layer
-  - a disposable timing pilot (1 relation x 50 pairs x 2 models)
-    to select N_final from {50, 75, 100} pairs per relation
-
-What is preserved from the initial plan:
-  - all compute runs through NNsight + NDIF (no local Llama hosting)
-  - the local RTX 5080 laptop handles only data prep, probe training,
-    and figures
-  - NDIF/NNsight citation requirement and acknowledgement
-
-
-Relation selection (finalised 2026-07-12)
-------------------------------------------
-Six semantically diverse CounterFact relations:
-
-  P19  (place of birth)          biographical, location-type target
-  P103 (native language)         biographical, language-type target
-  P101 (field of work)           biographical, domain-type target
-  P159 (headquarters location)   organisation, location-type target
-  P176 (manufacturer)            product/organisation
-  P138 (named after)             diverse entity types
-
-Design reasoning (recorded before seeing any probe results):
-
-  P19 and P103 were selected a priori, before observing any probe
-  results, as a conceptually related biographical pair: same subject
-  type (people), shared cultural/geographic
-  background, but nearly no subject overlap (3 out of ~800) and zero
-  target-vocabulary overlap (place names vs language names).  We
-  hypothesise stronger transfer between these two than between
-  unrelated pairs.
-
-  P19 and P159 share location-type targets (cities, countries) but
-  differ in subject type (people vs organisations) and predicate.
-  This allows comparison of whether conceptual similarity or shared
-  target vocabulary drives transfer.
-
-  P101 provides intermediate semantic distance from the biographical
-  pair.  P176 and P138 provide more distant relation structures.
-
-  These expectations are NOT built into the sampling or filtering
-  code.  The preprocessing treats all six relations identically
-  (except P103 French capping); hypotheses are tested only at the
-  analysis stage.
-
-
-Data statistics
-----------------
-Counts at each cleaning stage for the six target relations:
-
-  Stage 1 — Raw records in counterfact.json:
-
-    Relation  Records  Subjects  Top attribute (%)
-    P19        779       779     London (7.6%)
-    P103       919       919     French (63.9%)
-    P176       911       911     Toyota (11.0%)
-    P101       545       545     physics (13.4%)
-    P159       756       756     London (11.4%)
-    P138       279       279     Victoria (3.6%)
-
-  Stage 2 — After within-relation subject deduplication
-  (keep lowest case_id per subject):
-
-    P19   779 -> 779  (0 removed)
-    P103  919 -> 918  (1 removed)
-    P176  911 -> 900  (11 removed)
-    P101  545 -> 519  (26 removed)
-    P159  756 -> 754  (2 removed)
-    P138  279 -> 272  (7 removed)
-
-  Stage 3 — After cross-relation subject removal
-  (12 subjects appear in more than one target relation):
-
-    P19   779 -> 776  (3 removed: Michel Brault, Pierre Braunberger,
-                        Sergey Lavrov — shared with P103)
-    P103  918 -> 908  (10 removed: 7 shared with P101, 3 with P19)
-    P176  900 -> 899  (1 removed: PGM-17 Thor — shared with P138)
-    P101  519 -> 512  (7 removed: shared with P103)
-    P159  754 -> 753  (1 removed: Sheffield United F.C. — with P138)
-    P138  272 -> 270  (2 removed: PGM-17 Thor, Sheffield United F.C.)
-
-  Stage 4 — After duplicate-sentence removal:
-
-    No duplicate sentences found in any relation after prior cleaning.
-    All counts unchanged from Stage 3.
-
-  Final cleaned pool sizes:
-    P19=776, P103=908, P176=899, P101=512, P159=753, P138=270
-
-  The tightest relation is P138 with 270 usable records, well above
-  the N100 requirement.
-
-  P103 after cleaning: 908 records, French=582 (64.1%).
-
-
-Negative construction via derangement
---------------------------------------
-False sentences are constructed by permuting true attributes within
-each relation using a fixed-seed derangement: a permutation sigma of
-{1, ..., n} such that a_{sigma(i)} != a_i for every record i.
-
-Why this guarantees identical attribute marginals:
-
-  A permutation is a bijection.  The multiset of false attributes
-  {a_{sigma(1)}, ..., a_{sigma(n)}} is identical to the multiset of
-  true attributes {a_1, ..., a_n}.  Every attribute value appears
-  exactly the same number of times as true and as false.  This is an
-  exact algebraic guarantee.
-
-  Consequence: a linear probe cannot succeed merely because some
-  attribute word appears more often with one label.
-
-  The no-self-match constraint ensures no record's false sentence
-  contains the same attribute as its true sentence.
-
-  Existence condition: a valid derangement exists whenever no single
-  attribute occupies more than floor(n/2) positions.
-
-This approach is better suited to our probing control than using
-CounterFact's built-in target_new field, which was designed for
-knowledge-editing experiments and does not match attribute marginals
-across labels.
-
-
-P103 French capping
----------------------
-French represents 587/919 = 63.9% of P103 raw records, exceeding
-the derangement feasibility threshold (50%) at every sample size.
-
-Capping strategy: limit French records to 20/30/40 in the
-N50/N75/N100 subsets (= 40% of each subset), safely below the 50%
-derangement ceiling.  Non-French records are sampled with the same
-fixed seed used for all other relations.
-
-Other relations have no attribute exceeding 14% and need no capping.
-
-Scope of this control: capping eliminates unigram attribute-frequency
-difference between labels.  It does not rule out all forms of lexical
-leakage (e.g., correlated subject names or template structure).  The
-optional TF-IDF baseline described in phase2_final_plan.md Section 15
-addresses surface-lexical confounds more broadly.
-
-
-Nested subset construction
-----------------------------
-Three nested subsets per relation support the timing gate:
-
-  N50  = 50 pairs  (first selected)
-  N75  = N50 + 25 additional pairs
-  N100 = N75 + 25 additional pairs
-
-Nesting is strict: N50 subset N75 subset N100.  The same seed
-controls selection at every level.
-
-Derangements are computed independently per subset level because the
-attribute composition changes.  A given record's false attribute at
-N50 may differ from its false attribute at N100.
-
-
-Timing gate
-------------
-Before committing to a full six-relation extraction, a disposable
-pilot determines N_final.
-
-Pilot procedure:
-  1. Choose one of the six relations and take its first 50 pairs
-     (100 sentences).
-  2. Run the exact four-depth activation extraction on 8B and 70B
-     using the same batch size, token position, dtype, and NNsight
-     code path intended for the final run.
-  3. Record per model: queue wait, remote execution, result transfer,
-     and total wall-clock time.
-  4. Estimate full extraction time as:
-       N50  -> 6x pilot time per model
-       N75  -> 9x pilot time per model
-       N100 -> 12x pilot time per model
-  5. Choose N_final from {50, 75, 100} using the remaining time
-     budget, reserving at least six hours for probing, validation,
-     figures, and debugging.
-  6. Validate pilot tensor shapes and metadata, then discard pilot
-     activation files.  Keep only logs and timing record in
-     sample_size_decision.json.
-  7. Start fresh extraction for all six relations on both models.
-
-Maximum wasted inference: one relation at the 50-pair level.
-
-
-Cleaning pipeline summary
----------------------------
-  1. Filter to six target relations.
-  2. Remove duplicate subjects within each relation (keep lowest
-     case_id).
-  3. Remove subjects appearing in more than one target relation.
-  4. Remove duplicate sentences.
-  5. Apply P103 French cap.
-  6. Sample nested N50/N75/N100 subsets (fixed seed).
-  7. Construct derangement-based negatives per relation per subset.
-  8. Validate: class balance, no true==false, exact attribute
-     marginals, correct subset sizes, no duplicate examples.
-
-
-Output files
--------------
-  data/processed/examples.parquet
-    All examples with columns: example_id, pair_id, case_id,
-    relation_id, subject, template, true_attribute, used_attribute,
-    sentence, label, false_attribute_source_id, subset.
-
-  data/processed/relations.json
-    Per-relation metadata: record counts at each cleaning stage,
-    attribute distribution, French cap details (P103), template
-    statistics.
-
-  data/processed/dataset_manifest.json
-    Global seed, file hashes, counts, sampling decisions, validation
-    results.
-
-  Splits (3-fold pair-grouped CV assignments) are generated in a
-  separate step after data preparation is validated.
-
-
-Preprocessing execution record (2026-07-12)
-----------------------------------------------
-  Script: prepare_counterfact_multirelation.py
-  Seed: 20260712
-  Command: python prepare_counterfact_multirelation.py
-  Conda env: blackboxnlp
-
-  Derangement algorithm: sort records by (attribute, case_id), then
-  shift indices by max_group_size.  Deterministic given the selected
-  records — no additional random seed needed for the derangement itself.
-  Randomness enters only through subset sampling.
-
-  Per-relation RNG seeds (derived from global seed + relation index):
-    P19:  20260712 + 0*10007 = 20260712
-    P103: 20260712 + 1*10007 = 20270719 (sub-seeds +1/+2 for French/other)
-    P176: 20260712 + 2*10007 = 20280726
-    P101: 20260712 + 3*10007 = 20290733
-    P159: 20260712 + 4*10007 = 20300740
-    P138: 20260712 + 5*10007 = 20310747
-
-  Validation: 18/18 checks passed (6 relations x 3 subset levels).
-  All nesting checks passed (N50 ⊂ N75 ⊂ N100 by case_id).
-  Cross-relation subject check on N100: no overlap.
-
-  Output files:
-    data/processed/examples.parquet       (2700 rows)
-    data/processed/relations.json
-    data/processed/dataset_manifest.json
-
-  Manifest hashes:
-    raw_data:  d017056125178a13728594e66a801357a8db9ed7973a7425554bb4271de9fc6f
-    output:    5f526e216cbc5e1db494ca62340ecaad42c4515045ede33ec3658bde958a2d8d
-
-  Examples per subset:
-    N50:   600 (6 relations x 50 pairs x 2 labels)
-    N75:   900 (6 relations x 75 pairs x 2 labels)
-    N100: 1200 (6 relations x 100 pairs x 2 labels)
-
-  P103 French counts in selected subsets:
-    N50: 20/50 = 40%    (cap: 20)
-    N75: 30/75 = 40%    (cap: 30)
-    N100: 40/100 = 40%  (cap: 40)
-
-  Note: Stage 4 (sentence dedup) removed 0 records across all relations.
-  The earlier estimate of 3 duplicate sentences in P176 was based on a
-  raw-data scan before subject deduplication; those duplicates were
-  already eliminated in Stage 2 (within-relation subject dedup).
-
-  Data-quality observation: attribute marginals verified to match exactly
-  between true and false examples at every relation x subset level.  The
-  derangement shift algorithm guarantees this algebraically (permutation
-  preserves multiset), confirmed empirically.
-
-
-NNsight implementation note
-----------------------------
-NNsight 0.7.0 does not trace Python for-loops inside the
-model.trace() context manager.  Any .save() calls inside a for-loop
-are silently dropped — the trace completes but returns no tensor data
-(only ~341 bytes of metadata).
-
-Workaround: unroll layer saves as four explicit statements.  Since
-Phase 2 always extracts exactly four depths, this is straightforward.
-This behaviour was discovered during the pilot and is documented
-here so future scripts do not repeat the debugging.
-
-
-Pilot timing record (2026-07-12)
-----------------------------------
-  Script: extract_activations.py --pilot
-  Conda env: blackboxnlp-ndif
-  Pilot relation: P19
-  Pilot subset: N50 (50 pairs = 100 sentences)
-
-  Token position strategy: last real token in tokenised sequence.
-  Because the attribute is always the final word(s) in the sentence
-  (by construction in make_sentence()), the last token IS the final
-  subtoken of the target attribute.
-
-  8B (meta-llama/Llama-3.1-8B):
-    Layers extracted: 7, 15, 23, 31 (depths 25%, 50%, 75%, 100%)
-    Wall clock:       171.3s
-    Per sentence:     1.71s
-    Activation shape: (100, 4096) per layer, dtype float16
-
-  70B (meta-llama/Llama-3.1-70B):
-    Layers extracted: 19, 39, 59, 79 (depths 25%, 50%, 75%, 100%)
-    Wall clock:       231.9s
-    Per sentence:     2.32s
-    Activation shape: (100, 8192) per layer, dtype float16
-
-  Estimated full extraction times:
-    N50:   8B=17min + 70B=23min = 40min total
-    N75:   8B=26min + 70B=35min = 60min total
-    N100:  8B=34min + 70B=46min = 81min total
-
-  Pilot activations saved to activations/pilot/ for shape validation.
-  To be discarded after N_final is frozen.
-  Full timing evidence saved to sample_size_decision.json.
-
-
-N_final decision (2026-07-12)
--------------------------------
-  N_final = 100 pairs per relation
-  Total sentences: 1200 (6 relations x 100 pairs x 2 labels)
-  Estimated extraction: ~81 min (8B: 34min, 70B: 46min)
-  Rationale: well within remaining time budget; 6+ hours reserved
-  for probing, validation, figures, and debugging.
-
-  Pilot activations discarded after shape validation.
-  Timing evidence preserved in sample_size_decision.json.
-
-
-Full extraction record (2026-07-12)
---------------------------------------
-  Script: extract_activations.py --subset N100
-  Conda env: blackboxnlp-ndif
-
-  8B (meta-llama/Llama-3.1-8B):
-    Layers: 7, 15, 23, 31 (depths 25%, 50%, 75%, 100%)
-    Sentences: 1200
-    Wall clock: 1957s (32.6 min)
-    Per sentence: 1.63s
-    Shape: (1200, 4096) per layer, dtype float16
-
-  70B (meta-llama/Llama-3.1-70B):
-    Layers: 19, 39, 59, 79 (depths 25%, 50%, 75%, 100%)
-    Sentences: 1200
-    Wall clock: 2818s (47.0 min)
-    Per sentence: 2.35s
-    Shape: (1200, 8192) per layer, dtype float16
-
-  Total extraction: 4775s (79.6 min)
-  Estimated was 81 min — actual within 2% of estimate.
-
-  Dataset hash (both models): 7ef95d79b694631d
-  Token position: last real token (final subtoken of target attribute)
-
-  Output files:
-    activations/llama_3_1_8b/layer_{7,15,23,31}.npy
-    activations/llama_3_1_8b/sample_index.parquet
-    activations/llama_3_1_8b/manifest.json
-    activations/llama_3_1_70b/layer_{19,39,59,79}.npy
-    activations/llama_3_1_70b/sample_index.parquet
-    activations/llama_3_1_70b/manifest.json
-
-
-Checklist: filled after later stages
---------------------------------------
-  [x] Split seed: 2026071203
-  [x] Per-relation fold counts: 34/33/33 for all six relations
-      Same-pair constraint enforced: true/false in same fold.
-      Downstream join on example_id, not row index.
-  [x] Observed tokenisation edge cases: none encountered.
-  [x] Final figure paths: see "Output figures" section below.
-
-
-Probing execution record (2026-07-12)
----------------------------------------
-  Script: run_probing.py
-  Conda env: blackboxnlp
-  Config: experiment_config.json (hash: 667540b0fa890cca)
-
-  Probe: StandardScaler + LogisticRegression(solver=lbfgs, C=1.0,
-         penalty=l2, max_iter=1000, tol=0.0001, fit_intercept=True,
-         class_weight=None, random_state=42)
-  Scaler fitted only on each probe's training set.
-  ConvergenceWarning monitored: none triggered (all 204 probes converged).
-
-  Stage 1A (within-relation 3-fold CV):
-    144 fits (6 relations × 4 depths × 3 folds × 2 models)
-
-  Stage 1B (leave-one-relation-out):
-    48 fits (6 relations × 4 depths × 2 models)
-
-  Layer selection:
-    8B:  layer 7  (depth=0.25, mean_auc=0.9620, best=0.9660)
-         Within 0.005 tolerance → selected shallowest qualifying layer.
-    70B: layer 39 (depth=0.50, mean_auc=0.9770, best=0.9770)
-         Unique best — no tie-breaking needed.
-
-  Stage 2 (6×6 transfer matrix at selected layer):
-    12 full-source fits (6 sources × 2 models)
-    60 off-diagonal evaluations
-    Diagonal: mean 3-fold within AUC from Stage 1A
-
-  Total: 204 probes fitted, 204 npz + manifest saved.
-  Stage 2 predictions: 12,000 rows (off-diagonal only).
-
-  Output files:
-    results/stage1_metrics.csv          (192 rows: 1A + 1B)
-    results/stage2_metrics.csv          (60 rows: off-diagonal)
-    results/generality_gap.csv          (48 rows)
-    results/stage1_predictions.parquet  (19,200 rows: 1A + 1B)
-    results/stage2_predictions.parquet  (12,000 rows: off-diagonal only)
-    results/stage2_matrix_8b.csv
-    results/stage2_matrix_70b.csv
-    results/selected_layers.json
-    results/probe_weights/              (204 npz + 204 manifest json)
-
-  Relation order (fixed for all figures and matrices):
-    P19, P103, P101, P159, P176, P138
-
-
-Output figures (2026-07-12)
-----------------------------
-  Script: generate_figures.py
-  Conda env: blackboxnlp
-  Colors: 8B = #2a78d6 (blue), 70B = #1baf7a (green/aqua)
-
-  figures/fig1_depth_profile.{pdf,png}
-  figures/fig2_generality_gap.{pdf,png}
-  figures/fig3_transfer_matrices.{pdf,png}
-  figures/fig4_relation_gap_detail.{pdf,png}
-  figures/fig5_relation_profiles.{pdf,png}
-  results/table1_relation_results.csv
-
-
-===============================================================
-PAPER-WRITING REFERENCE (below this line)
-===============================================================
-
-Experiment summary
---------------------
-  Research question:
-    Does model scale make a linearly decodable truth representation
-    more transferable across factual relations?
-
-  Design:
-    - Two models: Llama-3.1-8B (32 layers) and Llama-3.1-70B (80 layers)
-    - Six semantically diverse CounterFact relations:
-        P19  (place of birth), P103 (native language),
-        P101 (field of work), P159 (headquarters location),
-        P176 (manufacturer),  P138 (named after)
-    - 100 factual pairs per relation (N100), 1200 sentences total
-    - Derangement-based negatives: a permutation that preserves
-      exact attribute marginals between true and false examples
-    - Four normalised depths per model: 25%, 50%, 75%, 100%
-    - Probe: StandardScaler + L2-regularised logistic regression
-    - Three evaluation stages:
-        1A. Within-relation 3-fold CV (pair-grouped)
-        1B. Leave-one-relation-out (train on 5, test on 1)
-        2.  6×6 pairwise transfer matrix at each model's best layer
-    - Generality gap = within-relation mean AUC − leave-one-out AUC
-      (positive = within > leave-one-out = relation-specific;
-       negative = leave-one-out > within = more general)
-    - Activations extracted via NNsight + NDIF (remote inference)
-
-
-Figure-by-figure interpretation
-----------------------------------
-
-  Fig 1: Depth profile (fig1_depth_profile.pdf)
-  ................................................
-    Shows mean ROC-AUC across all six relations at each normalised
-    depth, separately for within-relation 3-fold CV and leave-one-out,
-    for both models. Error bars = SD across 6 relations.
-    Background colour bands mark each model's selected layer depth.
-
-    Key observations:
-    - 70B achieves uniformly higher AUC than 8B at every depth in
-      both evaluation protocols.
-    - Both models peak in the first half of the network (8B at 25%,
-      70B at 50%), consistent with truth-relevant features being
-      encoded relatively early.
-    - 70B's within and leave-one-out curves track each other closely
-      (small gap), whereas 8B's curves diverge more, especially at
-      deeper layers. This is the visual signature of better cross-
-      relation transfer at larger scale.
-    - The decline at depth 100% is steeper for 8B than 70B.
-
-    Caption requirements:
-    - State that error bars are SD across 6 relations, not SE.
-    - Note background bands mark the selected layer per model.
-    - Clarify that within = mean 3-fold CV AUC, leave-one-out =
-      train-on-5-relations AUC.
-
-  Fig 2: Generality gap summary (fig2_generality_gap.pdf)
-  .........................................................
-    Grouped bar chart showing mean |gap| per depth for each model.
-
-    Key observations:
-    - 70B has smaller mean absolute generality gap at every depth:
-        depth 0.25: 70B 0.004 vs 8B 0.018
-        depth 0.50: 70B 0.003 vs 8B 0.011
-        depth 0.75: 70B 0.008 vs 8B 0.032
-        depth 1.00: 70B 0.025 vs 8B 0.048
-    - The gap advantage of 70B is most pronounced in the middle
-      layers (50–75%) and narrows at depth 100%.
-    - This directly addresses the RQ: scaling from 8B to 70B
-      reduces the difference between within-relation and cross-
-      relation probe performance.
-
-    Caption requirements:
-    - Define generality gap as within-relation mean AUC minus
-      leave-one-out AUC.
-    - State that bars show the mean absolute gap over 6 relations.
-
-  Fig 3: Transfer matrices (fig3_transfer_matrices.pdf)
-  ......................................................
-    Two 6×6 heatmaps side by side (8B left, 70B right) at each
-    model's selected layer. Rows = source relation, columns =
-    target relation. Diagonal = within-relation mean 3-fold CV AUC
-    (cross-validated). Off-diagonal = AUC from a probe trained on
-    the full source relation and evaluated on the full target.
-
-    Key observations:
-    - 70B off-diagonal values are uniformly higher than 8B (mean
-      off-diagonal: 70B ~0.972, 8B ~0.943).
-    - P19 (place of birth) is the weakest target for both models;
-      transferring into P19 yields the lowest values in both
-      matrices (8B: 0.849–0.918, 70B: 0.931–0.970).
-    - P176 and P138 are "easy" targets: nearly every source
-      relation transfers well to them at both scales.
-    - 70B shows a more uniform matrix with less row/column
-      variation, suggesting that the representation is more
-      relation-agnostic.
-    - The P19/P103 biographical pair does NOT show obviously
-      elevated mutual transfer relative to other pairs — this
-      is discussed under "Cannot claim" below.
-
-    Caption requirements:
-    - State that rows are source relations, columns are targets.
-    - State that diagonal values are mean 3-fold CV AUC (cross-
-      validated) and off-diagonal values come from probes trained
-      on the entire source relation.
-    - State which layer is shown per model (8B: layer 7 / depth
-      25%; 70B: layer 39 / depth 50%).
-
-  Fig 4: Relation-level gap detail (fig4_relation_gap_detail.pdf)
-  ................................................................
-    Horizontal diverging bar chart showing the signed generality
-    gap per relation at each model's selected layer. Positive =
-    within > leave-one-out (probe loses accuracy on unseen
-    relations). Negative = leave-one-out > within (probe gains
-    from multi-relation training). Labels show the numeric gap
-    value.
-
-    Key observations:
-    - P138 (named after) has the largest positive gap in 8B
-      (+0.048) but essentially zero in 70B (+0.001). This is
-      the single most dramatic per-relation scaling effect.
-    - P101 (field of work) in 8B also shows a notable positive
-      gap (+0.031) that shrinks to 0.000 in 70B.
-    - Negative gaps (P19, P159) indicate that multi-relation
-      training can help: probes trained on five other relations
-      slightly outperform within-relation probes for these
-      targets. This is consistent across both scales.
-    - 70B's gaps are compressed toward zero on both sides,
-      meaning it is not just that 70B is better at within-relation
-      probing — its representation is also more transferable.
-
-    Caption requirements:
-    - Define positive and negative gap direction.
-    - State that values are at each model's selected layer.
-
-  Fig 5: Relation depth profiles (fig5_relation_profiles.pdf)
-  .............................................................
-    2×3 small-multiple grid showing per-relation within-relation
-    AUC across the four depths for both models. Each panel is one
-    of the six relations.
-
-    Key observations:
-    - P176 and P138 are near ceiling (>0.99) for 70B at all
-      depths; 8B is also very high (>0.97).
-    - P19 is the hardest relation at both scales but still well
-      above chance (8B: 0.836–0.896, 70B: 0.935–0.950).
-    - The depth profiles are broadly similar across relations:
-      both models tend to peak early and decline at depth 100%.
-    - 70B shows a flatter profile (less depth-dependence) for
-      most relations compared to 8B.
-
-    Caption requirements:
-    - State that each panel shows within-relation mean 3-fold CV
-      AUC across four normalised depths.
-    - State that error bars or shaded bands (if present) are SD
-      across the 3 folds.
-
-  Table 1: Relation results (table1_relation_results.csv)
-  ........................................................
-    Columns: Relation, 8B within, 8B std, 8B leave-one-out,
-    8B gap, 70B within, 70B std, 70B leave-one-out, 70B gap.
-    Values are at each model's selected layer.
-    Mean row: mean gap uses the signed mean, not the mean of
-    absolute values.
-
-    Key numbers:
-      Mean 8B within AUC:         0.962 (SD 0.013)
-      Mean 8B leave-one-out AUC:  0.953
-      Mean 8B gap:                +0.009
-
-      Mean 70B within AUC:        0.977 (SD 0.007)
-      Mean 70B leave-one-out AUC: 0.979
-      Mean 70B gap:               −0.002
-
-    Interpretation:
-    - 70B has higher within-relation AUC (0.977 vs 0.962) and
-      even slightly higher leave-one-out AUC than within — the
-      mean gap is slightly negative (−0.002), meaning cross-
-      relation training marginally helps on average.
-    - 8B's mean gap is +0.009, meaning within-relation probes
-      are slightly better on average — there is a small cost
-      to generalising.
-    - 70B has lower within-relation SD (0.007 vs 0.013),
-      indicating more consistent performance across relations.
-
-    Caption requirements:
-    - State that values are at each model's selected layer
-      (8B layer 7 / 25%, 70B layer 39 / 50%).
-    - Define gap = within − leave-one-out.
-
-
-Core conclusions
-------------------
-  1. Cross-relation truth signal exists at both scales: leave-one-
-     out probes trained on five unrelated relations substantially
-     exceed chance when tested on the sixth, for all six relations
-     and at all four depths tested.
-
-  2. Scaling improves transferability: 70B has a smaller mean
-     absolute generality gap at every depth. At the selected layers,
-     70B's signed mean gap is −0.002 (essentially zero / slightly
-     negative) versus 8B's +0.009.
-
-  3. The 70B transfer matrix is more uniform: off-diagonal AUC
-     values are higher and less variable than 8B. The 70B
-     representation at its best layer is more relation-agnostic.
-
-  4. Relation identity still matters: P19 (place of birth) is
-     consistently the hardest target. P176 (manufacturer) and P138
-     (named after) are the easiest. This variation persists at both
-     scales, though it is attenuated in 70B.
-
-
-What CANNOT be claimed from these results
---------------------------------------------
-  1. No universal truth direction: we observe cross-relation
-     linear decodability of truth labels, not a single truth
-     direction. The probes at different source relations may use
-     different linear combinations. The transfer matrix shows that
-     these linear combinations generalise, but we have not shown
-     (nor tested) that a single vector separates true from false
-     across all relations simultaneously.
-
-  2. No causal claim: probing shows that truth-correlated
-     information is linearly decodable from the residual stream.
-     It does not demonstrate that the model uses this information
-     for prediction. Causal interventions (e.g., activation
-     patching, causal tracing) would be needed for that claim.
-
-  3. Lexical confounds not fully excluded: the derangement
-     guarantees identical attribute marginals, eliminating the most
-     obvious unigram leakage. However, more subtle lexical or
-     distributional cues (e.g., subject–attribute co-occurrence
-     patterns in pretraining data) are not controlled for. The
-     optional TF-IDF baseline described in the plan was not run.
-
-  4. No P19–P103 biographical cluster: we hypothesised ex ante
-     that the biographical pair (P19 place of birth, P103 native
-     language) might show stronger mutual transfer. The transfer
-     matrices do not support this: P19→P103 and P103→P19 are not
-     notably elevated relative to other source–target pairs. This
-     is an informative null — shared subject type does not
-     automatically yield privileged transfer.
-
-
-Limitations and caveats
---------------------------
-  1. Six relations only: the experiment covers six of the ~100
-     relations in CounterFact. Results may not generalise to
-     relations with very different structures (e.g., temporal,
-     numerical, or multi-hop relations).
-
-  2. Sample size: 100 pairs per relation (200 examples per
-     relation). This is sufficient for the linear probes used
-     here but small by the standards of some probing studies.
-
-  3. Single dataset: all data comes from CounterFact. Template-
-     generated sentences may contain distributional artifacts not
-     present in naturalistic text.
-
-  4. Linear probe only: a linear probe tests linear decodability.
-     A nonlinear probe (e.g., MLP) might find additional signal
-     or show different scaling behaviour.
-
-  5. NDIF-served models: activations were extracted via NNsight +
-     NDIF remote inference. We do not control exact batch
-     composition on the server side. Determinism is ensured by
-     per-sentence extraction and shape verification, but
-     floating-point non-determinism at the hardware level is
-     possible.
-
-  6. Leave-one-out sample size asymmetry: the within-relation
-     probe trains on ~67 pairs (2 folds of 100), while the
-     leave-one-out probe trains on 500 pairs (5 relations × 100).
-     The leave-one-out probe sees ~7.5× more training data. The
-     generality gap therefore confounds transferability with
-     training-set size. This limitation should be acknowledged
-     in the paper; it is standard for LOO designs but worth
-     stating.
-
-  7. Controls not run: the plan described two optional controls —
-     (a) a TF-IDF bag-of-words probe as a surface-feature
-     baseline, and (b) a random-labels sanity check. Neither was
-     run in this execution. The derangement design provides the
-     primary methodological control; the TF-IDF baseline would
-     strengthen the argument but is not essential.
-
-  8. Two model scales only: with just two points (8B and 70B),
-     the relationship between scale and transferability is a
-     comparison, not a trend. We cannot extrapolate to smaller
-     or larger models.
-
-
-Phase 1 ↔ Phase 2 narrative connection
------------------------------------------
-  Phase 1 (reproduction/ablation) showed that in a controlled
-  synthetic setting, breaking cross-pair truth co-occurrence
-  removes truth decodability from a fully trainable model while
-  preserving factual memorization. This established that truth
-  decodability is a function of statistical structure in the
-  training data, not an automatic byproduct of storing factual
-  knowledge.
-
-  Phase 2 asks the natural follow-up question: given that truth
-  decodability exists in pretrained language models, how does it
-  behave across relations, and does model scale affect its
-  generality?
-
-  The connection:
-  - Phase 1 shows truth decodability is contingent (can be
-    removed). Phase 2 shows it is general (transfers across
-    relations) and that this generality is scale-dependent.
-  - Phase 1 uses a controlled toy model to isolate the mechanism.
-    Phase 2 uses production-scale LLMs to test whether the
-    phenomenon matters in practice.
-  - Together they argue that truth representations in LLMs are
-    (a) not an artifact of model architecture but of data
-    structure, and (b) more transferable across factual domains
-    at larger scale, suggesting that scaling encourages more
-    abstract, relation-agnostic truth representations.
-
-  The paper should present Phase 1 as establishing the mechanism
-  and Phase 2 as testing its ecological validity at scale. The
-  shared methodological thread is linear probing of truth labels
-  with controlled negatives.
-
-
-Complete file inventory
--------------------------
-  Data:
-    data/raw/counterfact.json
-    data/processed/examples.parquet       (2700 rows, all 3 subset levels)
-    data/processed/relations.json
-    data/processed/splits.parquet
-    data/processed/dataset_manifest.json
-
-  Activations:
-    activations/llama_3_1_8b/layer_{7,15,23,31}.npy
-    activations/llama_3_1_8b/sample_index.parquet
-    activations/llama_3_1_8b/manifest.json  (includes file_sha256)
-    activations/llama_3_1_70b/layer_{19,39,59,79}.npy
-    activations/llama_3_1_70b/sample_index.parquet
-    activations/llama_3_1_70b/manifest.json (includes file_sha256)
-
-  Results:
-    results/stage1_metrics.csv            (192 rows)
-    results/stage2_metrics.csv            (60 rows)
-    results/generality_gap.csv            (48 rows)
-    results/stage1_predictions.parquet    (19,200 rows)
-    results/stage2_predictions.parquet    (12,000 rows)
-    results/stage2_matrix_8b.csv
-    results/stage2_matrix_70b.csv
-    results/selected_layers.json
-    results/table1_relation_results.csv
-    results/probe_weights/                (204 npz + 204 manifest json)
-
-  Figures:
-    figures/fig1_depth_profile.{pdf,png}
-    figures/fig2_generality_gap.{pdf,png}
-    figures/fig3_transfer_matrices.{pdf,png}
-    figures/fig4_relation_gap_detail.{pdf,png}
-    figures/fig5_relation_profiles.{pdf,png}
-
-  Scripts:
-    prepare_counterfact_multirelation.py
-    extract_activations.py
-    run_probing.py
-    generate_figures.py
-
-  Configuration:
-    experiment_config.json
-    sample_size_decision.json
-    phase2_final_plan.md
-    phase2_notes.txt                      (this file)