# ✅ How wrong the legacy IDXD reader is — measured disc-wide

[`BACKLOG.md`](BACKLOG.md) left this open: *"the legacy value-before-key
string-pool reader is an approximation … every number in this corpus that came
out of `get_f32`/`get_raw` is re-checkable against ground truth but not yet
re-checked. **First step: diff the two readers across the disc and count
disagreements.**"* Done, statically, over every IDXD object on the disc.

## The answer

**7 750 IDXD objects, 738 922 named fields whose true value is numeric:**

| legacy `get_f32` outcome | fields | share |
|---|---|---|
| correct | 291 307 | 39.42 % |
| returns `None` — no number produced, so harmless | 318 003 | 43.04 % |
| **returns a WRONG NUMBER** | **129 612** | **17.54 %** |

## ✅ And the wrongness has an exact predicate

| object shape | numeric fields | wrong numbers |
|---|---|---|
| **single-record** | 29 822 | **0 — 0.00 %** |
| **multi-record** | 709 100 | 129 612 — 18.28 % |

**The legacy reader is exactly correct on single-record objects and unsafe on
multi-record ones.** Not approximately — *zero* disagreements in 29 822 fields.

The mechanism is in `get_raw`'s own implementation: it flattens the string pool
to a token list, finds the **first** occurrence of the key, and returns the token
before it. It has no notion of records. So in an object with N records that each
carry a `Weight`, every record's `Weight` resolves to **record 0's** value:

```
Weight   truth=1.0     legacy=0.3
Points   truth=10000   legacy=4000
Weight   truth=3.3     legacy=0.3
Points   truth=15000   legacy=4000
```

`0.3` and `4000` are record 0's values, returned for every record after it.

## What this means for the corpus

* A number read with `get_f32` from a **single-record** object is safe — the diff
  found no counterexample in 29 822 fields.
* A number read from a **multi-record** object is safe only for the **first**
  record; every later record has a ~18 % chance of being record 0's value.
* The `None` bucket (43 %) is the guard working as designed: `get_f32` rejects a
  non-numeric neighbour rather than inventing a value, so a *defaulted* field
  reads as absent instead of wrong.

## ⚠️ A first pass that measured the wrong thing

My first sweep compared the truth against *"the string immediately preceding this
field's own key in the pool"* and reported **65.90 %** disagreement. That number
is not the legacy reader's error rate, because it is not what `get_raw` does —
`get_raw` uses the **first occurrence of the key anywhere in the object**, not
this field's own key. The 65.90 % figure is withdrawn; 17.54 % is the measured
one, and the `is_definite_value` guard is why it is so much lower.

The lesson is the same one this corpus keeps paying for: before measuring how
wrong a tool is, read what the tool actually does.