arm large_stringclasses 1
value | intervention large_stringclasses 10
values | alpha float64 0.5 4 | kind large_stringclasses 3
values | source int64 -1 7 | delta float64 -0.54 0.03 |
|---|---|---|---|---|---|
A1 | insub | 0.5 | control | -1 | -0.001389 |
A1 | insub | 1 | control | -1 | -0.016667 |
A1 | insub | 2 | control | -1 | -0.05 |
A1 | insub | 4 | control | -1 | -0.133333 |
A1 | rand | 0.5 | control | -1 | 0.004167 |
A1 | rand | 1 | control | -1 | 0 |
A1 | rand | 2 | control | -1 | -0.005556 |
A1 | rand | 4 | control | -1 | -0.004167 |
A1 | src0 | 0.5 | diag | 0 | -0.033333 |
A1 | src0 | 0.5 | offdiag | -1 | -0.001587 |
A1 | src0 | 1 | diag | 0 | -0.111111 |
A1 | src0 | 1 | offdiag | -1 | -0.009524 |
A1 | src0 | 2 | diag | 0 | -0.144444 |
A1 | src0 | 2 | offdiag | -1 | -0.012698 |
A1 | src0 | 4 | diag | 0 | -0.222222 |
A1 | src0 | 4 | offdiag | -1 | -0.028571 |
A1 | src1 | 0.5 | diag | 1 | -0.011111 |
A1 | src1 | 0.5 | offdiag | -1 | -0.004762 |
A1 | src1 | 1 | diag | 1 | -0.044444 |
A1 | src1 | 1 | offdiag | -1 | -0.004762 |
A1 | src1 | 2 | diag | 1 | -0.044444 |
A1 | src1 | 2 | offdiag | -1 | -0.006349 |
A1 | src1 | 4 | diag | 1 | -0.155556 |
A1 | src1 | 4 | offdiag | -1 | -0.098413 |
A1 | src2 | 0.5 | diag | 2 | -0.044444 |
A1 | src2 | 0.5 | offdiag | -1 | 0.001587 |
A1 | src2 | 1 | diag | 2 | -0.088889 |
A1 | src2 | 1 | offdiag | -1 | -0.006349 |
A1 | src2 | 2 | diag | 2 | -0.255556 |
A1 | src2 | 2 | offdiag | -1 | -0.004762 |
A1 | src2 | 4 | diag | 2 | -0.4 |
A1 | src2 | 4 | offdiag | -1 | -0.039683 |
A1 | src3 | 0.5 | diag | 3 | -0.022222 |
A1 | src3 | 0.5 | offdiag | -1 | 0.003175 |
A1 | src3 | 1 | diag | 3 | -0.066667 |
A1 | src3 | 1 | offdiag | -1 | -0.009524 |
A1 | src3 | 2 | diag | 3 | -0.188889 |
A1 | src3 | 2 | offdiag | -1 | -0.001587 |
A1 | src3 | 4 | diag | 3 | -0.255556 |
A1 | src3 | 4 | offdiag | -1 | -0.02381 |
A1 | src4 | 0.5 | diag | 4 | 0 |
A1 | src4 | 0.5 | offdiag | -1 | -0.004762 |
A1 | src4 | 1 | diag | 4 | 0.011111 |
A1 | src4 | 1 | offdiag | -1 | 0.007937 |
A1 | src4 | 2 | diag | 4 | -0.044444 |
A1 | src4 | 2 | offdiag | -1 | -0.01746 |
A1 | src4 | 4 | diag | 4 | -0.111111 |
A1 | src4 | 4 | offdiag | -1 | -0.01746 |
A1 | src5 | 0.5 | diag | 5 | -0.044444 |
A1 | src5 | 0.5 | offdiag | -1 | 0.001587 |
A1 | src5 | 1 | diag | 5 | -0.066667 |
A1 | src5 | 1 | offdiag | -1 | 0.019048 |
A1 | src5 | 2 | diag | 5 | -0.277778 |
A1 | src5 | 2 | offdiag | -1 | 0.028571 |
A1 | src5 | 4 | diag | 5 | -0.377778 |
A1 | src5 | 4 | offdiag | -1 | 0.014286 |
A1 | src6 | 0.5 | diag | 6 | -0.033333 |
A1 | src6 | 0.5 | offdiag | -1 | 0.003175 |
A1 | src6 | 1 | diag | 6 | -0.1 |
A1 | src6 | 1 | offdiag | -1 | 0.004762 |
A1 | src6 | 2 | diag | 6 | -0.2 |
A1 | src6 | 2 | offdiag | -1 | 0.009524 |
A1 | src6 | 4 | diag | 6 | -0.211111 |
A1 | src6 | 4 | offdiag | -1 | 0.001587 |
A1 | src7 | 0.5 | diag | 7 | -0.088889 |
A1 | src7 | 0.5 | offdiag | -1 | 0.009524 |
A1 | src7 | 1 | diag | 7 | -0.188889 |
A1 | src7 | 1 | offdiag | -1 | 0.004762 |
A1 | src7 | 2 | diag | 7 | -0.433333 |
A1 | src7 | 2 | offdiag | -1 | 0.004762 |
A1 | src7 | 4 | diag | 7 | -0.544444 |
A1 | src7 | 4 | offdiag | -1 | -0.014286 |
Provenance-codeword readout records
Per-experiment readout artifacts from the provenance codewords program:
source decodes, generation-time attribution records, model answers,
suppression/ablation sweeps, calibration files, capability checks (MMLU /
perplexity), and comparison matrices — the evidence behind every number in the
paper. JSON/parquet only; contents are synthetic or model-generated text plus
scores (no raw news text, no activation tensors — those stay on-cluster and
are indexed in the GitHub repo's ARTIFACTS.md).
Code, paper, and reproduction scripts:
https://github.com/Sid-MB/mats-gf-provenance-codewords — in particular
scripts/repro_c2_attribution.py (reproduces attribution 0.9957 from
exp02/results/c2/) and scripts/repro_adoption.py (within-pair adoption
0.9878 from #8's records).
Contents
| Folder | Experiment | Highlights |
|---|---|---|
exp01/ |
#1 provenance readout | decode/attribution sweeps across λ recipes; lam-0 probe (probes/) |
exp02/ |
#2 causal suppression | c2/answers_A1.parquet + c2_A1.json (attribution 0.9957), suppression sweep (c3/), LOSO, capability |
exp03/ |
#3 lever shore-up | seed-43 replication + 10-exposure ablation readouts, arm comparisons |
exp04/ |
#4 capacity 1.7B | 16/64/256-tier decodes, guessing floors, capability |
exp07/ |
#7 sign flip 8B | 256-way decode, margin records, MMLU |
exp08/ |
#8 adoption under conflict | contested-pair records (r2/), beta-null calibration + empirical threshold (r4/), demo equivalence-gate reports (demo/) |
exp09/ |
#9 natural substrate | R1/R3 comparison matrices, BM25 baselines, suppression + calibration, scored answers (tracer/phantom/mined/off-corpus) |
Provenance
Produced 2026-06/07; experiment IDs and
full artifact index in ARTIFACTS.md of the GitHub repo. Companion repos:
adapters,
demo checkpoint,
corpora.
CC-BY-4.0.
- Downloads last month
- 286