When a public model was fine-tuned by us on four consecutive decoder blocks at a known depth (LoRA r=8 on blocks LO..LO+3, merged; three depths), the reported onset matched the first fine-tuned block in 3 of 3 cells (blocks are numbered from 1). This does not establish behaviour on fine-tunes produced by others, whose depth and magnitude are unknown.
Frozen B7 preregistration SHA-256: 1028eedb853fb662aeb85a6d5f0404df1dcf99d367088bc2c94a9000a3243977.
Two controls were run first: one artifact, re-serialized into a different shard layout with identical tensor values, was reported as no difference; a known lesion reported a difference at the expected depth. In 3 controlled micro-fine-tune pairs (single model family), every difference this instrument reported was also visible in the model's output text. We have not yet produced a case where the instrument reports a difference that a plain output comparison would miss. Until we do, we make no claim that it can.
Frozen B8-SD preregistration SHA-256: d2dae32fa361f16a9f660241181b4a5bd19ca44338d284b15dbfc57e0c7a6d9d.
Onset localization works on our controlled fine-tunes; added value over plain output comparison is not yet shown.
The reading tells you where a change first becomes observable. It does not tell you how large the change is, or what kind of change it was: in our controlled tests, a fine-tune and a structural lesion at the same location produced the same reading.
Configuration comparison is limited to the compared fields. Numeric-precision fields are not compared here; any file difference is recorded in the artifact identity.
Reports now show categorical onset depth, a depth map and the extent of the observable difference. For all ten reports the difference is already present at the first evaluable position, so it may begin earlier, below the instrument floor; no starting depth is reported for these pairs.
Calibration: exact onset predicted and observed in 6 of 6 controlled cells on each of three public models (Qwen2.5-0.5B-Instruct, SmolLM2-1.7B, TinyLlama-1.1B-Chat-v1.0); cross-board agreement on onset 4 of 4 cells. These controlled changes do not establish performance on changes that arise naturally. The first four profile positions, the embedding output and the outputs of blocks 1 to 3, are below the instrument floor.
16 of 16 A = B comparisons completed across four models and two card classes: 0 observed differences and 0 output-text changes; two-sided exact 95% Clopper–Pearson interval 0–20.59%. This describes the tested matrix, not a population error rate.
| # | Pair · revisions | Card · runtime | Pair relation | Difference observed | Output text changed | Onset | Probes evaluated | Report |
|---|---|---|---|---|---|---|---|---|
| 1 | Qwen/Qwen2.5-0.5B@ 060db6499f32 → Qwen/Qwen2.5-0.5B-Instruct@ 7ae557604adf | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configs differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 2 | Qwen/Qwen2.5-0.5B@ 060db6499f32 → Qwen/Qwen2.5-0.5B-Instruct@ 7ae557604adf | RTX 4060PyTorch 2.9.1+cu128 / Transformers 5.5.4 | combined artifact change (configs differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 3 | Qwen/Qwen2.5-1.5B@ 8faed761d45a → Qwen/Qwen2.5-1.5B-Instruct@ 989aa7980e4c | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configs differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 4 | Qwen/Qwen2.5-1.5B@ 8faed761d45a → Qwen/Qwen2.5-1.5B-Instruct@ 989aa7980e4c | RTX 4060PyTorch 2.9.1+cu128 / Transformers 5.5.4 | combined artifact change (configs differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 5 | HuggingFaceTB/SmolLM2-1.7B@ effd688a1292 → HuggingFaceTB/SmolLM2-1.7B-Instruct@ 31b70e2e869a | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configs differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 6 | HuggingFaceTB/SmolLM2-1.7B@ effd688a1292 → HuggingFaceTB/SmolLM2-1.7B-Instruct@ 31b70e2e869a | RTX 4060PyTorch 2.9.1+cu128 / Transformers 5.5.4 | combined artifact change (configs differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 7 | TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T@ 59f6f375b26b → TinyLlama/TinyLlama-1.1B-Chat-v1.0@ fe8a4ea1ffed | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configuration identical in the compared fields; tokenizer semantics differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 8 | TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T@ 59f6f375b26b → TinyLlama/TinyLlama-1.1B-Chat-v1.0@ fe8a4ea1ffed | RTX 4060PyTorch 2.9.1+cu128 / Transformers 5.5.4 | combined artifact change (configuration identical in the compared fields; tokenizer semantics differ) | yes | withheld | already present at the first evaluable position | 24 | Open report → |
| 9 | Qwen/Qwen2.5-0.5B-Instruct@ 7ae557604adf → KingNish/Qwen2.5-0.5b-Test-ft@ 7917c64eeab3 | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (no difference recorded in the compared configuration fields) | yes | yes | already present at the first evaluable position | 24 | Open report → |
| 10 | HuggingFaceTB/SmolLM2-1.7B-Instruct@ 1dfd32118ef1 → motexture/SmolLCoder-1.7B-Instruct@ d9b7eb7b380d | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (no difference recorded in the compared configuration fields) | yes | yes | already present at the first evaluable position | 24 | Open report → |
| 11 | Qwen/Qwen2.5-0.5B-Instruct@ 7ae557604adf → tetracta/qwen2.5-0.5b-instruct-lora-blocks-11-14local checkpoint SHA-256 3bef74e68554 | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configuration identical in the compared fields) | yes | yes | first observable at decoder block 12 of 24 | 24 | Open report → |
| 12 | Qwen/Qwen2.5-0.5B-Instruct@ 7ae557604adf → tetracta/qwen2.5-0.5b-instruct-lora-blocks-5-8local checkpoint SHA-256 b3100da75e1f | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configuration identical in the compared fields) | yes | yes | first observable at decoder block 6 of 24 | 24 | Open report → |
| 13 | Qwen/Qwen2.5-0.5B-Instruct@ 7ae557604adf → tetracta/qwen2.5-0.5b-instruct-lora-blocks-18-21local checkpoint SHA-256 8cc3c44c6b45 | RTX 5070 TiPyTorch 2.11.0+cu128 / Transformers 5.10.1 | combined artifact change (configuration identical in the compared fields) | yes | yes | first observable at decoder block 19 of 24 | 24 | Open report → |
Rows 9–10: numeric precision (torch_dtype) differs; recorded in the artifact identity.
Rows 11–13 are controlled fine-tunes produced for calibration; not for use. Their B identifiers are local checkpoint SHA-256 values, not uploaded model revisions. No checkpoint files are included.
Rows 1–8 have different output contracts, so output text is withheld. Rows 9–10 share a comparable output contract: the recorded text was compared and differs. The depth maps record a difference at every evaluable position. No weights-only or pure fine-tuning attribution is made. Training claims for rows 9–10 are the authors’ descriptions and were not independently checked.
Rows 9–10 were also run on RTX 4060 with the same difference, output-text status and onset status; their card class and runtime are recorded in the full-revision index. The sample-pair agreement of 6 of 6 pairs is separate from the controlled experiment’s 4 of 4 cells.
TinyLlama configuration comparison is identical in the compared fields; tokenizer semantics differ. Commercial service admission rights review is still pending, separately from this publication. Row 10 uses the SmolLM2-1.7B-Instruct revision of 4 December 2024 to match the derivative’s tokenizer length and chat contract.
Read Scope & Limitations and the correction record. Numerical profiles remain private. No numerical equivalence across card classes or runtimes, quality verdict, or fitness-for-use assessment is claimed.