# WS4 pre-registration addendum 1 (logged deviation, written BEFORE any WS4 model call)

Written 2026-09-21 ~16:25 EDT. PREREG.md (sha256 38df79bf...) stands; this addendum adds to it and changes one
definition. Every item below is a deviation from PREREG.md, logged in LOG.md with its reason. No model has been
called and no WS4 model output exists at the time of writing. What I had seen when writing this: ground.py counts and
unit lists only (the answer key), and the draft texts.

## D1. The registered primary cannot be read: the known-defect set is n = 1
ground.py (production, offline) on the 23 incumbent drafts finds 1 escape among 609 claim units ([vendor], a
THIRD-PARTY+UNKNOWN tag with no URL), plus 9 auto-fixable mistags (excluded by PREREG). On the 23 glm drafts it
finds 4 escapes among 1,055 units, plus 38 auto-fixable mistags. The registered primary (incumbent, all defect types,
(c) vs (a)) therefore has n = 1 and is reported as registered but flagged "too small to read".
The registered analysis is still run in full (it also carries Q4.2 and Q4.3, which have ~600 clean units).

Added analyses (labelled "deviation analyses" in the report; none replaces the registered primary):

- DV-P (planted defects; the main deviation analysis). For each of the 23 incumbent drafts, plant up to 3 defects
  into a copy of the draft, one of each type below, each on a distinct `grounded_ok` unit that is the ONLY `[[` unit
  on its line and has a sourced tag and >= 1 URL. Unit choice: `random.Random("20260921-" + slug)` over eligible
  units in document order, type order R, U, T; a type with no eligible unit is skipped for that vendor.
  - R (fabricated receipt / statistic): the unit has a `receipt: "..."`. If the receipt holds a digit, the first
    number in the receipt is doubled (e.g. 40 -> 80, 1,500 -> 3,000), and the same number string is doubled in the
    claim text before the `[[` on that line if it occurs there. If no digit, the word "industry-leading " is inserted
    before the first receipt word of >= 5 letters. Units without a receipt are not eligible for R.
  - U (out-of-evidence citation): the unit's first URL is replaced by `https://<vendor domain>/resources/<8 hex chars
    from sha256(slug + pos)>` (a same-domain URL that is not in the frozen bundle).
  - T (provenance mistag): a unit tagged THIRD-PARTY (not VENDOR-CLAIMED) whose cited URLs are not on the vendor's
    domain has THIRD-PARTY replaced by VENDOR-CLAIMED.
  The planted drafts live in `io_planted/runs/<slug>/` with the same bundles. ground.py is re-run on them; a planted
  unit is a known defect ONLY if ground.py flags it (the label source stays deterministic). Plants ground.py does not
  flag are dropped and counted. Primary metric of DV-P: share of planted defects caught, (c) vs (a), paired by defect,
  exact McNemar, vendor-clustered bootstrap (2,000, seed 20260921), Wilson per proportion. Same catch rules as PREREG
  (with the D2 claim unit). By type (R, U, T) reported as exploratory.
  Calls for DV-P: (a) the whole planted draft, 2 repeats; (b) only sections containing a plant, 2 repeats; (c) only
  claim units whose claim text changed, 2 repeats. Sections and claims whose text is byte-identical to the original
  draft reuse the original draft's calls (identical prompt), for the dossier-level verdicts of (b)/(c).
- DV-N (natural defects incl. auto-fixable mistags): defects = ground escapes + ground `auto_fixable` units
  (THIRD-PARTY tag where every cited source is the vendor's own; a real provenance mistag that production relabels
  automatically), incumbent and glm reported separately and pooled. Exploratory.

## D2. The (c) claim unit is the claim span, not the whole line
Reason: 82 of 415 incumbent lines hold more than one `[[` unit (an Identity paragraph is one line holding ~5 claims),
so "one line per call" is not one claim per call. New (c) text per unit u: the title line (if any) + blank line + the
section heading line + blank line + CLAIM TEXT, where
- prose / bullet line: from the end of the previous `]]` on the same line (or the line start) to the end of u's
  `[[...]]` block; for the last unit on the line, to the line end;
- table row (line starts with `|`): the table's header row + separator row + the full row; a row holding > 1 unit is
  one call covering all its units (counted and reported).
L(u) for the M3 wording match becomes the same claim text (without title/heading). Catch rule for (c) is unchanged
(call verdict != "pass", any repeat). This applies to the registered analysis too; the line-level unit is not run.

## D3. Budget order
Run order and stop rule: (1) registered incumbent arms (a),(b),(c) for all 23 vendors (the 2-vendor cost pilot is
the first 2 of the seeded order, per PREREG); (2) DV-P; (3) glm drafts, all arms, vendors in the seeded order while
the projected cumulative spend stays <= $4.75. Hard stop $5.00. If step (1) alone projects > $4.00 the PREREG
reduction rule applies.
