| # Checkpoint ↔ code compatibility |
|
|
| **A representation change breaks old checkpoints. This file records which code loads which arm.** |
|
|
| | arms | conditioning | `PER_SITE_DIM` | `COND_DIM` | load with | |
| |---|---|---|---|---| |
| | `A_*`, `B_*`, `C_*`, `D_*`, `E_long*`, `F_scout`, `G_100*` | pre-WP-018 | 62 | 948 | commit **`203c036`** or earlier | |
| | `H_pq*` and later | WP-018 (+ per-shell SPECIFIED bit) | 68 | 1020 | commit **`74dae8f`** or later | |
|
|
| Loading a pre-WP-018 checkpoint with current code raises: |
|
|
| ``` |
| size mismatch for site_proj.0.weight: copying a param with shape torch.Size([256, 62]) |
| from checkpoint, the shape in current model is torch.Size([256, 68]) |
| ``` |
|
|
| That is the correct, loud failure — not a corrupted download. |
|
|
| ## How this happened, and the rule it earns |
|
|
| WP-018 added a **SPECIFIED bit per shell** so a user could ask for fewer than six shells. That is a |
| change to the *input representation*, and it silently invalidated **16 of 19 trained arms** for |
| inference. It surfaced only when a control experiment tried to feed a shallow query to a |
| full-query model and torch raised a size mismatch. |
|
|
| > **RULE: a change to the conditioning or token representation is a BREAKING change. It must (a) be |
| > recorded in the checkpoint's own metadata, (b) be checked on load with a message that names the |
| > commit to use, and (c) appear in this table before the change is committed.** |
|
|
| ⚠ `arm_meta.json` records `tokenizer_version` and `vocab_hash` but **not** a conditioning version — |
| which is why nothing caught this at load time. Adding one is the fix; until then, this table is the |
| source of truth. |
|
|
