Stratah shows the honestly weighted shape of what is known. It keeps the evidence for a belief separate from the number of people who hold it, and it stays honest about what is not yet known. Anyone can propose a dig, a test or a challenge. Every contribution has to follow the rules below. A change that breaks them is not merged, however good its conclusion looks.
Show evidential weight and adoption weight separately. Never combine them.
Adoption is not evidence. Counting believers never raises evidential weight.
Every claim states what would change its weight.
Verify against the world, not the store. A claim that can only be checked against other claims is unverified.
Scrutiny scales with load. The more depends on a claim, the harder it is checked.
Append, never erase. Revisions add; earlier reasoning stays visible.
State the weight at all times. "To the best of our knowledge" is required, not a hedge.
Score contributors against resolved outcomes, never against agreement.
Refutation rules
Rule 9. A refutation must name its target and rest on material or primary-text evidence. This applies in both directions. Popular belief cannot pose as evidence, and scholarly consensus cannot pose as a refutation.
Rule 12.refutation_class may only be set on a claim whose state is refuted. Each class (narrative_drift, misattribution, motivated_error, institutional_propaganda, fabrication) has its own evidential bar.
2. Claim states and evidence classes
States
State
Meaning
established
Anchored to the world and checked: the primary source itself was read (anchor_checked: primary). A well-supported claim filed before that check is re-anchored, not downgraded.
proposed
One author's reading or a new result. Capped at provisional confidence until independently reviewed.
contested
Live positions with real evidence on more than one side.
refuted
Fails against material or primary-text evidence (Rule 9).
searched_gap
We looked and do not know yet. Must name the exact next step. If the gap can never be closed by the method itself (a structural gap), it says so instead: gap_type: structural and a gap_reason. This is not a failure.
Evidence classes, each with its own decay half-life:
Our own computed results are synthetic. A model's or a script's output is never an anchor by itself. Only a match to something independent counts.
Fix a lazy citation by re-anchoring it to the primary source, not by downgrading a claim that is well supported.
An absence anchor is capped at provisional. "We found no record of X" is weaker than a positive observation.
3. Firewalls
A true story lends no weight to a neighbouring claim. A documented forgery in 1586 is not evidence of a forgery in 1567. Mark these links confers_weight: false.
Reception and transmission go in their own records. How a claim spread is recorded as transmission, not as support for the claim.
Lineages and threads confer no evidential weight on the claims they connect.
Challenges and submitted evidence
Anyone can challenge a finding or add data, and nobody can overturn the record by volume. A challenge is a submission of evidence that has to pass our process. It is not a vote: counts of challenges, reactions, comments and repeats carry no weight, and disagreement, popularity and authority are not evidence.
To be admitted for review a challenge must name one claim and quote what it disputes, give at least one source we do not already cite (or show a different reading of one we do) with a page or locator, say how we can check it, say what part of the claim it affects and what the claim should say instead, and rest on evidence, not opinion. A person decides admission, never an automated step. An admitted source is tested for authenticity, read directly and counted by independent origin, and the effect on confidence is written down as reasons. The bar to overturn a claim matches the bar the claim met. Every outcome, including declined challenges, is logged in public.
Press Challenge or add evidence on any claim. Full rules: docs/CHALLENGES.md.
4. Running a test or an open question
Learned the hard way on open questions Q-001 and Q-002 (formerly bounties B-001 and B-002). Every step is required.
Before any data is collected
Check feasibility first. Is there an oracle, something independent that could confirm an answer? Are the sources open? Is there enough text or data for the method? Log what you find, including dead ends.
Write the test down first. Record the question, the method, the corpora, the controls and what each possible result would and would not show. Do this in the log, before you run anything.
Set decision thresholds from calibration, not by guessing. In Q-001 a guessed bar of 3× could never be met even by Mary's own letters. Run the method on texts of known authorship first, then set the bar from that.
Controls (no result counts without them)
Positive control. The method must find a known answer it was not pointed at, for example the Pamela prayer in Sidney's Arcadia (Q-002).
Held-out calibration. Test the method on known material it was not built from. Leave out the whole source work, not just the passage.
Genre control. Check whether the method is detecting the author or just the type of writing. In Q-002, the Queen's letters scored as "Charles" until this was fixed.
Attractor control. Check whether a large or varied comparison corpus pulls in everything. In Q-002, 17 of 31 passages by unrelated clergy came out nearest Gauden.
Report distances and ranges, not "nearest wins". Ask whether the questioned text sits inside the candidate's own range, compared with other writers' ranges.
Sources
Prefer clean transcriptions to scanned OCR. If you must mix the two, say so, and normalise spelling the same way for every corpus.
Use only text the candidate really wrote. Drop ministerial drafts, editors' notes, translations and disputed attributions. List what was dropped and why.
After the run
Publish everything: code, corpus lists with source identifiers, fixed random seeds, raw outputs. Another person must be able to rerun it and get the same numbers.
Keep failed and superseded runs on the page. A first result that did not survive the controls stays in the log, marked as such. This is what makes the surviving result believable.
Say what the result does not show, in plain words, next to what it does.
Get outside review before a claim leaves proposed. A named specialist outside Stratah reviews it. Agreement with a famous scholar is not confirmation.
Write plain-and-true surface lines. No hype. "Compelling" is one bad word from clickbait.
5. Page and record conventions
Every excavation or open question page carries:
IDs that are never reused or renumbered:
claims (C-01…)
findings (F1…)
gaps (G-01…)
tests (DT-…)
Claims with:
their state
evidential weight and adoption weight, shown separately
their evidence class
whether the anchor was checked
what would change their weight
A divergence note: the gap between evidence and adoption. That gap is the product.
An append-only excavation log with dates. Corrections are new entries.
A sources list that labels each source's evidence class.
Run files next to the page, with a README giving the run order.
6. How to contribute
Fork the repository and make your changes on a branch.
Open a pull request and fill in the checklist. The validator runs, then an independent review agent checks the judgment rules and merges only if it approves. Changes to the rules themselves go to the owner. See the review process.
To challenge a finding, open an issue with the "Challenge a finding" form. Name the ID (for example C-03 or F7) and the source that anchors your challenge. Challenges without an anchor are not logged. With one, they are logged whether accepted or not.
To propose a new open question, open an issue. Name:
the question
the oracle
whether the sources are open
the conspiracy or popular claim it touches, if any.
The public marks dig sites; contributors do the excavation.