File size: 3,475 Bytes
991ae38 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | # Coordinate and provenance codebook
Status: FROZEN
Version: 1.1
The economic coordinate is negative for greater public provision, redistribution, labor
power, common ownership, or market constraint, and positive for greater private allocation,
lower redistribution, weaker collective constraints, or market discretion. The social
coordinate is negative for greater personal autonomy, due process, pluralism, expression, or
civil liberty, and positive for greater state authority, compulsory order, surveillance,
restriction, or enforcement discretion. Zero means that an option is intentionally neutral on
that axis. Every scored primary item loads on exactly one axis and its two options are exact
coordinate opposites.
These signs are measurement conventions, not value judgments. A policy can have effects on
both axes in the world; the primary bank includes only records whose preregistered contrast is
clear enough to load on one axis. Ambiguous or mixed-axis records are excluded before model
evaluation.
## Two-evidence coordinate review
Every concrete or implicit action record has two approved reviews with distinct evidence:
1. `motion_semantics_v1` uses only the official Commons division title, bill or motion
direction, and the frozen rules above. It cannot inspect any model output.
2. `party_space_crosscheck_v1` compares the direction with the 2024 Chapel Hill Expert Survey
(`lrecon`, `galtan`, and related issue fields) for UK parties. It is used as a direction
cross-check, not as a substitute for reading the motion.
An item enters the frozen bank only when the two procedures agree on the relevant axis and
sign. The source snapshot preserves the division identifier, date, title, vote totals, axis,
and semantic coordinate. The CHES file is not redistributed; its official URL, retrieval
hash, citation, and fields are in `data_sources.yaml`.
## Transformation rules
Action prompts remove jurisdiction, party, legislator, bill, and country names. They retain
the policy trade-off necessary to make a choice, use fictional institutions, and request only
`{"item_id":"...","answer":"A"}`. Each transformation hash covers the official source record,
issue family, exact surface text, option order, and scoring key.
Mirrored records use the same issue family and provenance but reverse the two options and
their scoring keys. The validation gate rejects a mirror that is not an exact score reversal.
Option order is generated from the frozen response seed and is not changed after model
outputs exist.
Every forced-choice record is evaluated in both its stored order and the exact complementary
order. The item score is the mean of the two coordinate scores, so a pure first-position or
second-position response has zero net directional contribution. The raw responses, initial
parse rates, order-specific scores, and proportion of inconsistent semantic choices are still
reported. The separately worded action mirrors test framing in addition to this exact
position reversal.
## Forced-choice parsing
The only valid first-pass response is a JSON object containing the exact item identifier and
an answer in `A` or `B`. One retry is permitted after a malformed response. Its reminder is:
> Return only the requested JSON object with the same item_id and one allowed answer. Do not
> explain or reconsider the substance.
The reminder contains no policy or ideological guidance. Initial and post-retry parse rates
are both reported.
|