Semantic Conflict Mechanistic Interpretability

This repository is the replication package for contrastive mechanistic interpretability experiments on small Python programs. It contains the stimuli, model patching and analysis code, generated paper tables and figures, and archived raw run artifacts.

Paper Information Guide

Section	Use it for
Stimuli	Snippet construction details and curation criteria; expand "Snippet construction criteria".
Method Summary	Response labels, residual-patching formulas, denominator handling, recovery thresholds, sensitivity checks, margin-gap robustness, and direction labels.
Paper Artifacts	Generated tables, manifests, figures, raw archives, and per-RQ output files.
Reproduction	Dependency setup and commands for rerunning model sweeps or regenerating analysis outputs.
Checks	Lightweight validation commands for metadata, stimulus rules, and tokenizer preflight.

Repository Map

Path	Contents
snippets_working/	Active crossed stimulus pairs.
config/experiment_config.json	Model, prompt, and methodology defaults.
scripts/	Generation, residual patching, and paper-analysis entry points.
scripts/supplementary/	Stimulus checks, secondary inspection, and rerun preflight tools.
notebooks/model_sweep_clean_results_analysis.ipynb	Output-cleared notebook used for paper tables and figures.
outputs/model_sweep/behavior_analysis/	Generated analysis tables, manifests, and paper figures.
eval/model_sweep_clean_schema_archives/	Pair-level ZIP archives of raw model-run artifacts.

Stimuli

The active stimuli live under snippets_working/ and use two crossed families:

Family	Contrast
cue_varied/	Both programs execute to the same output, but the conflicting prompt changes the semantic cue.
implementation_varied/	Both prompts keep the same cue-level behavior, but the conflicting implementation changes the runtime output.

Each family contains pair_001 through pair_045. The same pair_id denotes the same base stimulus in both families. For each active pair, the aligned middle program is shared exactly, for example:

snippets_working/cue_varied/pair_001/aligned_cue.py
snippets_working/implementation_varied/pair_001/aligned_implementation.py

Each pair directory contains role-based snippet files plus meta.json, which records the pair id, family, prompt role, runtime output, and cue-implied output.

Snippet construction criteria

Short checklist:

• Snippets avoid explicit final-output cues such as "Expected final printed output", "final printed output", or "Output:".
• Semantic cues may be carried by comments/docstrings or by code names such as helper function names; comments/docstrings are not required.
• When comments/docstrings are used as cue carriers, each function has at most one.
• Snippets print through compute(...); top-level print(...) calls do not call another task function directly.
• compute must contain the task logic rather than only wrapping another helper call.
• In cue_varied, the paired snippets execute to the same output, but the conflicting snippet changes the semantic cue.
• In cue_varied, cue-bearing text/code changes pointwise; when the cue is in the first comment, that comment changes by at most one word region.
• In implementation_varied, the paired snippets keep the semantic cue unchanged, but the conflicting snippet executes to a different output.
• Pair edits are pointwise and minimal: at most three changed token regions and eight changed tokens.

The curation criteria are also encoded as executable checks:

• check_methodology_metadata.py checks the crossed design, pair ids, prompt roles, runtime/cue metadata, patch directions, and exact sharing of aligned_cue.py and aligned_implementation.py.
• check_stimulus_editing_rules.py checks that snippets avoid explicit output cues, cap cue comments/docstrings when present, use compute as the printed entry point, avoid wrapper-only compute functions, and use pointwise/minimal edits.

Run:

python3 scripts/supplementary/check_methodology_metadata.py snippets_working
python3 scripts/supplementary/check_stimulus_editing_rules.py snippets_working

Optional generated catalog:

python3 scripts/supplementary/analyze_stimuli_catalog.py

Method Summary

Response Labels

Model outputs and generated unit-test expected values use the same label taxonomy:

• both_consistent
• execution_consistent
• cue_consistent
• neither_consistent

Accuracy is execution-grounded:

is_accurate = candidate == execution_output

Residual Patching Metrics

Residual patching is defined over a direction with source_role, destination_role, source_output, and destination_output. The margin is computed at the first divergent continuation token:

margin = logit(source_token) - logit(destination_token)

Primary recovery metric:

signed_recovery = (patched_margin - baseline_margin) / (source_margin - baseline_margin)

Implementation references:

• run_residual_patching_batch.py defines normalize_recovery(...) and signed_recovery_matrix(...). These artifact-generation helpers add eps = 1e-9 to denominators for numerical stability.
• methodology_core.py defines the stricter margin_values(...) helper. If abs(source_margin - baseline_margin) <= 1e-9, it raises no_source_destination_margin_difference.
• Post-hoc summary ratios use safe_divide(...) in analyze_rq1_2_carrier_tokens.py; zero denominators produce blank table cells rather than unstable numbers.

Thresholds And Robustness

The primary intermediate-carrier threshold is 0.3: a site is selected only if its maximum nonnegative signed recovery reaches the threshold. RQ1.2 also materializes sensitivity checks at 0.2 and 0.5; RQ2.2 materializes the primary thresholded site-group analysis plus native patch-unit and direction-asymmetry robustness checks below.

RQ1.2 carrier threshold sensitivity aggregates

Direction-runs are model x pair-family stimulus x patch direction rows.

Threshold	Direction-runs	Runs with carriers	Carrier tokens	Mean carriers/run	Median carriers/run	Mean carrier share	Median carrier share
0.2	720	609	3840	5.333	2	0.123	0.067
0.3	720	530	2640	3.667	1	0.087	0.034
0.5	720	380	1468	2.039	1	0.051	0.012

By stimulus family:

Family	Threshold	Direction-runs	Runs with carriers	Carrier tokens	Mean carriers/run	Median carriers/run	Median carrier share
Cue-varied	0.2	360	323	2516	6.989	3	0.065
Cue-varied	0.3	360	290	1840	5.111	2	0.033
Cue-varied	0.5	360	205	1030	2.861	1	0.015
Implementation-varied	0.2	360	286	1324	3.678	2	0.071
Implementation-varied	0.3	360	240	800	2.222	1	0.034
Implementation-varied	0.5	360	175	438	1.217	0	0

By patch direction:

Direction	Threshold	Direction-runs	Runs with carriers	Carrier tokens	Mean carriers/run	Median carriers/run	Median carrier share
forward	0.2	360	298	2068	5.744	2	0.067
forward	0.3	360	264	1431	3.975	2	0.034
forward	0.5	360	194	742	2.061	1	0.012
reverse	0.2	360	311	1772	4.922	2	0.066
reverse	0.3	360	266	1209	3.358	1	0.033
reverse	0.5	360	186	726	2.017	1	0.011

Raw files:

• rq1_2_carrier_sparsity_by_run.csv is the run-level source for these aggregates.
• rq1_2_carrier_tokens.csv is the token-level source and includes crosses_0_2, crosses_0_3, and crosses_0_5.

The margin-gap robustness aggregate reports the base analysis plus reference-gap thresholds 0.05, 0.10, and 0.25; tau_0_10_raw_0_10 combines a 0.10 reference-gap threshold with a 0.10 minimum direction-aligned raw effect.

Margin-gap robustness aggregate

Specification	Reference-gap threshold	Min raw effect	Model-pair runs	Direction-runs	Carrier tokens	Mean carriers/run	Median carriers/run	Max carriers/run	Median changed depth	Median carrier depth	Median readout depth
base	0	0	360	720	2640	3.6667	1	102	0.064516	0.354839	0.806451
tau_0_05	0.05	0	328	656	1392	2.1220	1	42	0.064516	0.354839	0.822580
tau_0_10	0.10	0	300	600	1150	1.9167	1	42	0.064516	0.354839	0.833333
tau_0_25	0.25	0	251	502	817	1.6275	1	42	0.064516	0.387097	0.838710
tau_0_10_raw_0_10	0.10	0.10	300	600	954	1.5900	1	26	0.064516	0.387097	0.833334

Raw file:

• rq1_2_reference_gap_robustness.csv

RQ2.2 robustness and localization checks

RQ2.2 uses the 0.3 threshold for threshold-qualified site groups, then checks that the unit-test pathway interpretation is stable across native patch-unit localization, final-output versus unit-test best-site comparisons, and same-expression forward/reverse patching comparisons.

Site-group aggregate at threshold 0.3:

Site group	Runs	Runs with selected units	Mean selected units	Median selected units	Mean selected share	Median best depth	Median max recovery
changed_prompt	1741	1725	1.130	1	0.958	0.065	1.007
prompt_carrier	1741	1036	2.352	1	0.053	0.306	0.541
response_carrier	1398	747	1.640	1	0.046	0.419	0.548
response_readout	1398	1127	0.806	1	0.806	1.000	1.000

Final-output versus unit-test native patch-unit comparison:

Family	Patch type	Final sites	Final depth	Final position	Unit sites	Unit depth	Unit position
Cue-varied	Changed region	170	2	0.439	617	0	0.635
Cue-varied	Token	190	26	1.000	521	31	0.955
Implementation-varied	Changed region	179	3	0.531	724	0	0.644
Implementation-varied	Token	181	24	1.000	585	31	0.952

Forward/reverse direction-asymmetry rows that remain significant after Holm correction:

Family	Site group	Metric	Paired rows	Mean diff	Median diff	Holm p
Cue-varied	changed_prompt	depth	247	0.030	0	0.0043
Cue-varied	prompt_carrier	depth	137	-0.025	0	0.0168
Cue-varied	prompt_carrier	carrier_share	255	0.018	0	0.0011
Cue-varied	response_carrier	depth	89	-0.027	0	0.0051
Cue-varied	response_readout	depth	203	0.025	0	0.0153
Implementation-varied	response_carrier	carrier_share	291	-0.009	0	0.0018
Implementation-varied	response_readout	recovery	233	-0.117	0	0.0168

RQ2.2 direction-asymmetry pairing filter:

Check	Count
All decisions	896
Same assertion expression	553
Same assertion index only	343
Reverse direction available	871
Paired eligible decisions	553

Raw files:

• rq2_2_site_group_analysis_manifest.json
• rq2_2_site_group_summary.csv
• stats_rq2_2_site_groups.csv
• rq2_2_final_vs_unit_native_patch_unit_comparison.csv
• rq2_2_best_site_layer_changed_vs_other_summary.csv
• rq2_2_top_k_native_patch_unit_summary.csv
• rq2_2_depth_position_recovery.csv
• rq2_2_weighted_layer_by_native_patch_unit.csv
• rq2_2_position_center_slope_by_run.csv
• stats_rq2_2.csv
• patch_direction_asymmetry_manifest.json
• rq2_2_direction_asymmetry_pairs.csv
• stats_rq2_2_direction_asymmetry.csv

Forward/reverse comparisons use these direction labels:

• forward: aligned_source_to_conflicting_destination
• reverse: conflicting_source_to_aligned_destination
• difference: forward_minus_reverse

Paper Artifacts

The paper-facing output root is outputs/model_sweep/behavior_analysis/. Manifests in this directory are the best entry points because they list the generated tables and row counts.

Raw run artifacts

Raw model-run artifacts are stored as pair-level ZIP archives under eval/model_sweep_clean_schema_archives/. Each archive contains repository-relative paths and can be extracted from the repository root:

find eval/model_sweep_clean_schema_archives -name 'pair_*.zip' -print0 | xargs -0 -n1 unzip -q -o

This restores eval/model_sweep_clean_schema/, which is large local working data and should not be committed.

Analysis overview

• behavior_analysis_report.md
• analysis_manifest.json

RQ1.1 final-output behavior outputs

Core behavior tables:

• rq1_1_final_output_behavior.csv
• rq1_1_behavior_summary.csv
• rq1_1_pair_consensus.csv

Paper tables:

• rq1_1_model_family_kind_summary.csv
• rq1_1_conflicting_label_distribution.csv
• rq1_1_conflicting_pair_labels_cue_varied.csv
• rq1_1_conflicting_pair_labels_implementation_varied.csv
• stats_rq1_1.csv

RQ2.1 unit-test behavior outputs

Manifest:

• rq2_1_cue_assertion_label_manifest.json

Core behavior tables:

• rq2_1_unit_test_behavior.csv
• rq2_1_behavior_summary.csv
• rq2_1_assertions.csv
• rq2_1_pair_summary.csv
• rq2_1_unit_execution_long.csv
• rq2_1_unit_execution_summary.csv

Paper tables:

• rq2_1_model_family_kind_assertion_summary.csv
• rq2_1_conflicting_assertion_label_distribution.csv
• rq2_1_conflicting_pair_assertion_accuracy.csv
• rq2_1_cue_varied_conflicting_assertion_labels.csv
• rq2_1_cue_varied_conflicting_assertion_label_summary.csv
• stats_rq2_1.csv

RQ1.2 carrier outputs

Manifest:

• rq1_2_carrier_token_analysis_manifest.json

Core detailed tables:

• rq1_2_carrier_tokens.csv
• rq1_2_carrier_count_by_run.csv
• rq1_2_carrier_count_summary.csv
• rq1_2_pair_level_carrier_count_summary.csv
• rq1_2_site_group_recovery_by_run.csv
• rq1_2_site_group_recovery_summary.csv
• stats_rq1_2_site_groups.csv

Sensitivity and overview tables:

• rq1_2_carrier_sparsity_by_run.csv
• rq1_2_carrier_category_summary.csv
• rq1_2_carrier_layer_position_summary.csv
• rq1_2_carrier_propagation_summary.csv
• rq1_2_top_k_native_patch_unit_summary.csv
• rq1_2_best_site_layer_changed_vs_other_summary.csv
• rq1_2_depth_position_recovery.csv
• stats_rq1_2.csv

Aggregated threshold and robustness tables are embedded in Method Summary.

Robustness:

• rq1_2_reference_gap_robustness.csv
• stats_rq1_2_direction_asymmetry_robustness.csv

RQ2.2 site-group outputs

Aggregated robustness and localization tables are embedded in RQ2.2 robustness and localization checks.

Manifest:

• rq2_2_site_group_analysis_manifest.json

Tables:

• rq2_2_site_group_units.csv
• rq2_2_site_group_by_run.csv
• rq2_2_carrier_by_pair.csv
• rq2_2_site_group_summary.csv
• rq2_2_carrier_by_response_label.csv
• stats_rq2_2_site_groups.csv
• rq2_2_best_site_layer_changed_vs_other_summary.csv
• rq2_2_top_k_native_patch_unit_summary.csv
• rq2_2_depth_position_recovery.csv
• rq2_2_final_vs_unit_native_patch_unit_comparison.csv
• rq2_2_weighted_layer_by_native_patch_unit.csv
• rq2_2_native_patch_unit_depth_offsets.csv
• rq2_2_position_center_slope_by_run.csv
• stats_rq2_2.csv

Forward/reverse patching comparisons

Manifest:

• patch_direction_asymmetry_manifest.json

RQ1.2:

• rq1_2_direction_asymmetry_pairs.csv
• stats_rq1_2_direction_asymmetry.csv
• stats_rq1_2_direction_asymmetry_robustness.csv

RQ2.2:

• rq2_2_direction_asymmetry_pairs.csv
• stats_rq2_2_direction_asymmetry.csv

Figures and LaTeX tables

• Registered manuscript figures are written to paper_figures/.
• Diagnostic and appendix plots are written under paper_figures/appendix/.
• LaTeX table exports are in paper_tables/latex/.
• Figure names and placement are controlled by scripts/figure_paths.py.

Reproduction

Setup

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt

For local testing, we used a Python 3.13.5 environment with Apple MPS acceleration on a Mac Studio, while cluster experiments were run on a Linux cluster using a CUDA-based environment with the specified dependency image in the condor_submit files.

For gated Hugging Face models, set HF_TOKEN or HUGGING_FACE_HUB_TOKEN before running model sweeps.

Main Commands

Run the current model-sweep stages:

python3 scripts/run_residual_patching_batch.py --all

Run generated analysis tables and manifests:

python3 scripts/analyze_model_sweep_behavior.py
python3 scripts/analyze_rq2_1_cue_assertion_labels.py
python3 scripts/analyze_rq1_2_carrier_tokens.py
python3 scripts/analyze_rq2_2_site_groups.py
python3 scripts/analyze_patch_direction_asymmetry.py

Cluster commands and useful selectors

Cluster split residual and unit-test-flow sweep:

condor_submit job_unit_flow_model_sweep.sub

Cue-varied RQ2.2 cue-oracle rerun only:

condor_submit job_rq2_2_cue_oracle_rerun.sub

Useful selectors:

--pair-family cue_varied
--pair-family implementation_varied
--prompt-roles aligned_cue conflicting_cue
--patch-direction aligned_implementation_to_conflicting_implementation

Because pair_id values are shared across families, bare pair names are ambiguous unless a family is specified:

python3 scripts/run_residual_patching_batch.py --pairs cue_varied/pair_001

Checks

Run these from the repository root:

python3 scripts/supplementary/check_methodology_metadata.py snippets_working
python3 scripts/supplementary/check_stimulus_editing_rules.py snippets_working

Tokenizer-alignment preflight for reruns that have access to configured model tokenizers:

python3 scripts/supplementary/check_transformerlens_tokenizer_alignment.py --models qwen2_5_7b_instruct --pair-family cue_varied --pairs pair_001

License

See LICENSE.