# Civilisers: a source-grounded comparison — protocol draft

Status: proposed, not run. This is not a preregistration or a result. The existing V1
lineups and V2 source audit informed this draft; their material is discovery evidence,
not unseen validation data. No new model judgments or significance estimates accompany it.

## The question

Do the proposed civiliser pairs share a more distinctive narrative structure and set of
details than other teaching or founding figures in a defined comparison collection?

The first useful output should show the **strongest alternative matches**, with the actual
passages, agreements and differences. It should not start with a verdict about a lost
civilisation. Resemblance, its frequency in a selected corpus, and its historical cause are
three separate questions.

## A small, feasible first run

Use Viracocha, Quetzalcoatl, Bochica and Oannes as the four named targets. Aim for roughly
20–30 other teaching/founding figures, subject to source feasibility. This is a planning
range, not a sample-size calculation or permission to stop when a desired pattern appears.
One person can check the decisive source passages; AI can retrieve and organise candidates,
extract assertions and propose comparisons. Independent specialist validation would remain
a separate limitation, not something replaced by agreement between models.

### 1. Define the comparison collection before scoring it

- Choose and name a bounded reference collection or index, and publish an eligibility rule
  for teaching/founding figures. Do not let an AI's favourite examples silently define the
  worldwide population. The precise collection and eligible list are still to be chosen.
- Include plausible close competitors: other teachers, founders and culture-bringers,
  including female figures. Avoid a pool dominated by unrelated trickster or war stories.
- Retain same-region and potentially related examples, but label them. They can be useful
  parallels without being independent negative controls. Report within-region and
  cross-region comparisons separately.
- Log every included/excluded candidate and its source availability. Set the stopping rule
  before scoring, based on coverage and feasibility, not on the target's rank.
- If the final set is purposive, report ranks within that set. Do not call its distribution
  a worldwide base rate, even if the pool is much better matched than V1's.

### 2. Build comparable source packets

- Start with the accessible accounts already identified in V2, then check earlier witnesses
  and alternative versions. An online translation is an inspected edition, not automatically
  the earliest source or an unmediated Indigenous account.
- Use complete relevant episodes with page/chapter/line references, not stitched biographies.
  Identify who speaks, who acts, who reports the tradition and who supplies an interpretation.
- Set a comparable search-effort and source-coverage rule for every figure. Record missing
  coverage. Equal word counts alone would not make unlike sources equivalent.
- Keep variants as separate packets and translations/copies in the same source family.
  A figure with more versions must not silently get more chances to produce a good match.
- Record arrival, instruction, crisis, opposition, departure and later expectation as
  separate events. Preserve the actors and order. Missing mention is not a negative trait.

### 3. Compare meaning at two levels

The desired resemblance is not merely word overlap, and a list of rigid yes/no details
can also miss it. Use two independently reported views:

1. **Narrative relationships:** who changes whose life, what knowledge is brought, the
   order of the change and departure, and the role of a catastrophe. Different vehicles
   can occupy a similar narrative role; record that broader match without erasing the
   literal difference.
2. **Specific details:** named skills, companions, objects, prohibitions, directions and
   distinctive event sequences. Generic teaching or travel should not by itself count as
   an unusually specific shared package.

Each proposed correspondence must cite both passages. Also require the strongest mismatch,
uncertain readings and any dependence between the sources. Keep a source quotation separate
from the model's inference about its significance.

An independent second AI reader should attempt to verify the passage support and challenge
the match. Publish disagreements; do not treat model consensus as historical validation.
Names can be hidden as a sensitivity check, but do not call recognisable myths truly blind.
Models may know accounts not supplied to them; unsupported outside knowledge must not enter
the recorded match.

### 4. Give the proposed pair real competition

For each target, compare every eligible candidate in the defined pool. Show at least the
five strongest candidates under each comparison view, including unexpected ones, rather
than only displaying the result for the proposed partner. Avoid presenting thousands of
reshuffled lineups of the same stories as thousands of independent observations.

Before the main scoring pass, fix the rubric, treatment of ties and missing evidence,
candidate/source units, pooling rule for variants and ordering of the packets. Randomise
presentation order and record exactly what each reader received. Rubric tuning on the V2
material is development; it cannot become a preregistered discovery retroactively.

Calibration should include:

- Two faithful paraphrases/translations of the same episode, labelled as a reading check,
  not independent historical confirmation or a universal ancestry threshold.
- Synthetic contrast pairs that swap arrival for departure or transfer a skill to a
  different actor. Label these as artificial checks, never ancient evidence.
- Other teachers with generic roles in common but different event structures.

If readers cannot consistently distinguish the controlled changes, do not publish a
rarity verdict from their rankings. Keep the source atlas and report that instrument limit.

### 5. Test what the result depends on

- Does the preferred match change between the broader structure and specific-detail views?
- Does it change when generic teaching/travel is set aside?
- Does it survive a different source version or a conservative translation reading?
- Does it depend on merging several figures or on treating a chronicler's gloss as the tale?
- Do additional source families, held aside from rubric development, support the proposed
  correspondence? Log prior exposure; AI pretraining prevents a claim of perfect blindness.

Report the source-family-level ranks, ties, coverage and disagreement. A rank is a
description of this corpus, not a probability of chance invention. Any later inferential
test needs a defensible sampling/null model and must account for shared sources and
multiple examined pairs. No p-value is specified in this draft.

## What a reader should get

A page for each target with its actual source packet, the proposed partner and the best
alternative matches. Each comparison shows shared structure, literal differences,
unresolved readings and the passages that would change the judgment. A separate source
history tracks where particularly striking details are attested and where the trail ends.

Before calling this an experiment in progress, resolve the collection, coverage threshold,
final corpus, comparison rubric and review procedure. Freeze the dated protocol before the
main ranking pass; preserve subsequent changes as amendments. Until then, V2 remains a
useful source-audit pilot and this remains a proposal.