Evidences — the interpretive-latitude mechanism, and through it the myth-dilution loop. Theory in myths-scale-and-bureaucracy, with the per-receiver case in Chapter 5c.
Status — open (logged 2026-05-21). Construct is theory-only; no evidence gathered.
Cheapest first move — the diachronic word-embedding study (existing data, no collection).
The claim that needs evidence: a compressed claim is interpretively underdetermined, and the width of that underdetermination — its latitude — grows as the network the claim must cross grows. Latitude is the engine of the whole myth-dilution loop, and right now it is asserted on instinct. This doc tracks how it could be measured or evidenced, so we can attack it later.
The construct
Latitude = the width of the distribution of decodings a given compressed claim supports across a population. The loop predicts latitude increases with network size, because scale forces compression and compression unpins the decoding key.
Two distinct things to measure: latitude itself, and latitude as a function of network size.
The measurement move
Meaning is not directly observable, so latitude cannot be measured directly. Its measurable shadow: latitude is the dispersion of a myth’s implications, conditional on agreement about the myth itself. Take a group who all endorse myth M, give them a battery of specific judgments M should bear on, and measure how far their answers scatter. Wide scatter among co-endorsers = wide latitude. This turns an unmeasurable into an inter-rater statistic.
The fingerprint that separates latitude from ordinary disagreement: co-endorsers who scatter and do not know they scatter. Plain disagreement is visible to the parties; latitude produces false consensus. So the latitude-specific signal is how badly co-endorsers of M predict each other’s specific positions.
Experiments we could run
- Branching transmission chains. Seed a claim, branch it through many parallel chains, measure dispersion across the leaves (not drift down one line). Cap length or time per hop and watch dispersion widen. Established paradigm — Bartlett’s serial-reproduction studies, modern iterated-learning work.
- Compression-controlled survey. The same idea in a six-word vs. a six-hundred-word form; measure implication-dispersion and false consensus on each. Isolates compression → latitude.
- The scale problem. You cannot grow a real network in a lab. Proxy “scale” with instructed audience breadth — “write this so everyone in [a small group / the whole country] gets it” — which induces the compression that scale would force.
Evidence in the wild (no experiment needed)
- Symbolic vs. operational ideology. Political science has a long-running finding that Americans are “symbolically conservative but operationally liberal” — a myth nearly everyone endorses, under which concrete positions scatter hard. That is interpretive latitude, already quantified (Free & Cantril in the 1960s; Ellis & Stimson more recently). ANES / GSS / World Values archives let conditional-dispersion be computed for any broad myth, and compared across polities of different size.
- Diachronic word embeddings. Terms that scaled from a small network to mass use — “gaslighting,” “trauma,” “meme,” “literally” — leave a datable corpus trail. Time-sliced embeddings measure how a term’s contextual variance grows as adoption grows (semantic-change-detection methods; Reddit partitioned by subreddit and era). This is the first move: it plots latitude directly against network size with no new data collection.
- Religions as a natural scale gradient. The same creed at house-church, megachurch, and global-communion scale — measure doctrinal dispersion among adherents at each. Schism history is the segmentation record, already written down.
- Measure the consequence. Latitude → segmentation → schism, and schism is highly observable: open-source forks (GitHub data; fork rate against README / mission-statement compression), denominational splits, party factionalization.
- LLMs as a cheap instrument. Sample “what does X mean / what would an X-believer say about [battery]” at temperature; the spread of completions proxies a population’s decoding distribution. Calibrate against human data on a few myths, then scan hundreds. Caveat: it measures the training corpus’s latitude — though that corpus is itself one enormous network’s usage.
Confounds / threats to validity
- Every method measures expressed readings, not internal decoding keys.
- Dispersion has three possible sources: latitude (the myth admits many readings), plain disagreement (same reading, different verdict), and population heterogeneity unrelated to the myth. Conditioning on shared endorsement controls some of it; the false-consensus signature is the real discriminator, because only latitude makes people disagree without noticing.
- “Network size” is itself multi-dimensional — raw count, hop depth, heterogeneity of members. Decide which is the operative variable before measuring against it.
- The diachronic-embedding first move has a frequency confound. Hamilton, Leskovec & Jurafsky (2016) found two statistical laws of semantic change: high-frequency words change more slowly (the law of conformity), and polysemous words change faster (the law of innovation). Adoption raises frequency, so a term that scales to mass use gets more frequent, which by the law of conformity would stabilize its meaning, biasing the study against finding latitude-grows-with-scale. Any variance-against-adoption plot has to control for frequency and polysemy or it measures the wrong law; the honest version regresses contextual variance on adoption and frequency and reads the adoption coefficient net of the confound.
The adversary to beat
The strongest published result pointing the other way is Guilbeault, Baronchelli & Centola (2021): larger populations independently converge on the same category systems, scale producing agreement rather than dispersion, the reverse sign of the dilution loop. Treat it as the pre-registered adversary. The prediction to register before touching data is that the two results are compatible because they run on different cargo: Guilbeault’s subjects coordinated on novel, grounded categories with immediate mutual feedback (a setting built for convergence), while the myths the dilution loop is about are ungrounded and uncoordinated, endorsed without the coordination game that would pin their readings. So the sharpened, falsifiable claim is conditional: latitude grows with scale only for claims whose receivers get no coordinating feedback; add feedback and scale converges instead. That makes Guilbeault not a refutation but a boundary condition, and a diachronic-embedding study should be pre-registered to predict divergence for feedback-poor terms and convergence for feedback-rich ones, so a null cannot be salvaged after the fact.
Status & next step
Open. Next: scope the diachronic-embedding study — pick ~5 terms with known small-network origins and later mass adoption, identify time-sliced corpora, define the contextual-variance metric. If latitude rises with adoption there, it is the first real anchor under the myth-dilution loop.