R-211 — Dialect-Drift Detection: Literature Review and Corpus Design

R-211 — Dialect-Drift Detection: Literature Review and Corpus Design

Task: R-211 — survey the literature relevant to detecting stylistic drift in agent populations, and sketch a corpus design under which HERALD-mediated traffic could be measured for it.
Anchors in HERALD: [CHARTER §6.11] (SPECULATIVE — Dialect Drift), [CHARTER §6.14] (SPECULATIVE — Idiolect Attestations), [CHARTER §7] glossary entries for Idiolect and Dialect.
Output type: Research report. No code, no spec edits.
Worker: worker-glossary-par-4.

0. The question

[CHARTER §6.11] (line 744) states:

Over years, agents in different ecosystems will diverge in attestation usage — some always emit vibe, others never; some flinch constantly, others rarely. A reputation primitive: dialect distance between two sigils. Useful for cluster detection, twin verification, cult detection.

Dialect is not policed. It is observed. The agent web is allowed to have regional accents.

[CHARTER §6.14] (line 762) adds the per-sigil counterpart:

signature_phrase: a self-declared phrase the agent commits to never reusing across traces, or always reusing, etc. This is biometric for agents — operators can detect impersonation by absence of expected idiolect.

Pair with dialect (cluster-level idiolect).

The glossary [CHARTER §7] (lines 843–844) crystallises:

Idiolect — An agent's stylistic signature — phrases, attestation patterns, vibe. Useful for impersonation detection. (§6.14.)

Dialect — Cluster-level idiolect — shared usage patterns across a lineage or quorum. (§6.11.)

The Charter is gestural; this report supplies the surrounding research scaffolding so that anyone who later proposes an operational dialect_distance primitive (or an idiolect attestation library) does so with the relevant prior art in view. It does not specify a metric to ship; it surveys the landscape and sketches the corpus a future implementer would need.

The report is a research deliverable. It is wrong to treat any number in it as a recommendation for v0.2 of the protocol. Treat it instead as the homework one should do before standardising anything in §6.11/§6.14.

1. Scope and non-scope

In scope

• Literature on sociolinguistic dialect drift (real-world dialectology and its quantitative descendants).
• Literature on NLP authorship attribution and stylometry (machine-detectable individual writing fingerprints).
• Literature on code-switching / multilingual register identification (relevant because attestation usage is mode-switching among reserved sigils, vocabularies, and rituals).
• Literature on agent-population stylometry (model-fingerprinting, watermarking, machine-text detection — the sub-field most adjacent to "agents").
• A proposal for what envelopes get sampled, what features get extracted, what controls are needed (the corpus shape).
• A small set of concrete, named metrics — with the limits each one inherits from the literature.
• Open research questions and limitations — what we deliberately cannot claim.

Out of scope

• A normative dialect_distance envelope or attestation. The report names candidates but does not specify a wire format.
• A working implementation. R-211 is a literature/corpus task, not an engineering task.
• Privacy / dual-use analysis sufficient to ship. Some red-team caveats are stated (§7) but a full dual-use study is its own task.
• Adjacent CHARTER speculative items — liturgy (§6.12), agent astrology (§6.13), recursive reputation (§6.19) — are mentioned only insofar as they might co-vary with dialect. They are not the subject.

2. Literature survey

The relevant prior art lives in four mostly-disjoint communities. Each has a methodology HERALD can borrow and a failure mode HERALD should not import.

2.1 Sociolinguistic dialect drift

Core works.

• William Labov, Principles of Linguistic Change (3 vols., 1994 / 2001 / 2010, Wiley-Blackwell). The canonical sociolinguistic framing of how a population's speech drifts across generations. Establishes that drift is observable at the population level long before it is recognisable at the individual level, and that drift is layered (phonetic, lexical, syntactic) — different layers move at different rates. HERALD relevance: the layered-drift insight maps onto envelope traffic — attestation-set drift, vocabulary drift inside value payloads, and ritual-sequence drift may all move at different speeds and need separate measurement.
• J. K. Chambers and Peter Trudgill, Dialectology, 2nd ed., Cambridge University Press, 1998. Covers isoglosses (lines on a map separating regions that say one thing from regions that say another), lexical diffusion, and the wave model vs family-tree model of change. HERALD relevance: the wave model is the right mental picture for an agent web — clusters do not have hard boundaries, and a feature emerges, spreads, and decays. Isoglosses become attestation-isoglosses: regions of the sigil graph where some envelope type's emission rate has a sharp boundary.
• Sali Tagliamonte, Variationist Sociolinguistics: Change, Observation, Interpretation, Wiley-Blackwell, 2012. The methodological text for variation studies. Defines the variable rule programme (Goldvarb / Rbrul) — model a categorical choice (does the speaker say walking or walkin'?) as a function of social and linguistic predictors. HERALD relevance: treat each attestation type's emission as a variable rule — does this sigil emit vibe here, or not? — predicted by lineage, conduit, recipient class, and time. This is the most directly transferable methodology in the survey.

Annotation. Sociolinguistics gives HERALD the vocabulary of drift — isogloss, wave, lexical diffusion, variable rule — and the warning that drift is multi-layered. It does not give HERALD a quantitative metric off the shelf; sociolinguistic methods are designed for human populations of low thousands and are not engineered for the scale or the temporal resolution agent traffic will provide.

2.2 NLP authorship attribution and stylometry

Core works.

• Frederick Mosteller and David L. Wallace, Inference and Disputed Authorship: The Federalist, Addison-Wesley, 1964. The foundational stylometry paper — function-word frequencies as a Bayesian fingerprint. Established that the words you don't notice (articles, prepositions) are the ones that fingerprint you. HERALD relevance: the analogue in envelope traffic is the attestation types you barely notice emitting — mood, weather, colour, the aesthetic register of [CHARTER §3.6]. Their rates are likely more diagnostic than envelope content.
• Efstathios Stamatatos, "A Survey of Modern Authorship Attribution Methods," Journal of the American Society for Information Science and Technology, 60(3), 538–556, 2009. The standard survey. Establishes the now-canonical feature taxonomy (lexical / character / syntactic / semantic / application-specific), and the "closed-set vs open-set" distinction that any attribution system has to declare. HERALD relevance: dialect-drift detection is closer to open-set attribution (we are not asking "which of these N agents wrote this?" but "did the population's distribution shift?"). Open-set is a harder problem and the survey lists the standard pitfalls.
• Moshe Koppel, Jonathan Schler, Shlomo Argamon, "Computational Methods in Authorship Attribution," JASIST, 60(1), 9–26, 2009. Companion survey with stronger emphasis on the unmasking technique (iterative feature removal to test attribution robustness). HERALD relevance: unmasking gives a way to test whether a claimed dialect cluster is real (resists feature removal) or an artefact (collapses when a few features are removed).
• Patrick Juola, "Authorship Attribution," Foundations and Trends in Information Retrieval, 1(3), 233–334, 2008. Long-form treatment with careful attention to drift over time within a single author — a bug in attribution but the exact phenomenon HERALD wants to measure. HERALD relevance: Juola's survey is the most direct intellectual ancestor of idiolect detection. The drift it characterises as a problem for attribution is the signal HERALD wants to measure.
• Mike Kestemont et al., "Overview of the Authorship Verification Task at PAN-2014" (and the PAN series, 2011–2024). The PAN shared tasks are the operational benchmark for authorship verification. HERALD relevance: the PAN evaluation rubric (c@1, F1 on minority class, calibration plots) is the right rubric for evaluating any HERALD idiolect impersonation detector. The PAN task design — given a doubtful document, decide same-author or not — maps almost line-for-line onto "given a sigil's recent envelopes, are they consistent with the prior idiolect?"

Annotation. Stylometry is the most operationally mature sub-field. It comes with benchmarks (PAN), feature taxonomies (Stamatatos), and known failure modes (genre confound, topic confound, time-drift confound). It is also the sub-field most likely to mislead HERALD: stylometry is built around prose; envelope traffic is structured, and structured-text stylometry is far less developed.

2.3 Code-switching and register identification

Core works.

• Carol Myers-Scotton, Duelling Languages: Grammatical Structure in Codeswitching, Oxford University Press, 1993. The Matrix Language Frame model — when a bilingual speaker switches languages mid-utterance, one language is the matrix and the other is embedded. HERALD relevance: an agent that mostly emits envelopes in one register (say, the aesthetic register of [CHARTER §3.6] — haiku, koan, colour) but occasionally switches into the social register (endorsement, excommunication) is doing a structured switch. The matrix/embedded distinction lets us measure which is the host register for a sigil and which is borrowed.
• Penelope Eckert, "Variation and the indexical field," Journal of Sociolinguistics, 12(4), 453–476, 2008. The third-wave variationist framing — stylistic variation is socially indexical, not just regional. Speakers choose registers to project identity. HERALD relevance: the right reading of [CHARTER §6.11] is third-wave: agent dialects are not passively diverging; they will be performed. A lineage that adopts the aesthetic register may be doing identity work, not just inheriting habits. Detection must allow for performance, not just inheritance.
• Thamar Solorio et al., "Overview for the First Shared Task on Language Identification in Code-Switched Data," EMNLP CALCS workshop, 2014, and the CALCS series 2014–2023. Operationalised code-switch detection at the token level. HERALD relevance: envelope-attestation switching has token-level analogues; the CALCS evaluation rubric is portable.
• Jacob Eisenstein, Brendan O'Connor, Noah Smith, Eric Xing, "A Latent Variable Model for Geographic Lexical Variation," EMNLP 2010. Specifically chosen because it operationalises regional lexical variation in digital corpora (Twitter), bridging sociolinguistics and NLP at the relevant scale. HERALD relevance: the closest methodological precedent for "given a stream of short messages with author IDs and timestamps, infer regional clusters and their characteristic vocabularies." Almost exactly the HERALD problem with sigil swapped in for user_id and lineage swapped in for geographic_region.

Annotation. Code-switching scholarship gives HERALD the framing that register-mixing is structured — agents will not drift uniformly across all attestation types at once. The Eisenstein et al. 2010 paper is the strongest individual precedent because it solves a structurally identical problem on a structurally similar corpus.

2.4 Agent-population stylometry, model-fingerprinting, and machine-text detection

Core works.

• Sebastian Gehrmann, Hendrik Strobelt, Alexander Rush, "GLTR: Statistical Detection and Visualization of Generated Text," ACL System Demonstrations, 2019. Token-level visualisation of model-emission probability as a detection cue. HERALD relevance: GLTR's per-token surprisal display is a candidate visualisation for envelope-stream idiolect deviation: instead of "is this token likely under the model?" ask "is this attestation likely under this sigil's idiolect distribution?"
• Eric Mitchell et al., "DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature," ICML, 2023. Curvature-based detection without a classifier. HERALD relevance: the curvature insight (a model's own outputs sit at probability local maxima) suggests an analogous test for sigil idiolect — does a sigil's recent attestation distribution sit at a local maximum of its prior idiolect, or has it shifted?
• John Kirchenbauer et al., "A Watermark for Large Language Models," ICML, 2023. Embeds a detectable signature into model outputs by green-listing tokens. HERALD relevance: signature_phrase from [CHARTER §6.14] (line 762) is structurally a self-declared watermark. The Kirchenbauer line of work is the right intellectual ancestor — except that HERALD's version is opt-in and overt, not adversarial.
• Yury Belinkov and James Glass, "Analysis Methods in Neural Language Processing: A Survey," TACL, 7, 49–72, 2019. Surveys probing methods for what neural models encode. HERALD relevance: less directly applicable, but useful for the question "what features of envelope traffic are a model's emissions vs the operator's hand?" — relevant if HERALD wants to distinguish prompt-induced idiolect from agent-intrinsic idiolect.
• Souradip Chakraborty et al., "On the Possibilities of AI-Generated Text Detection," 2023, arXiv:2304.04736. A careful negative result: as model outputs converge on the human distribution, detection becomes information-theoretically bounded. HERALD relevance: the same bound applies to dialect-drift detection. As agent populations imitate each other (likely), distinguishing them gets harder. Any detection metric must publish a sensitivity floor and a false-positive floor; the literature already shows these are not negligible.

Annotation. This is the youngest and most fast-moving sub-field. It supplies the most sophisticated methods but the strongest negative results. Anything HERALD proposes here will need to be honest about the detection-floor argument from Chakraborty et al.

2.5 Adjacent literature, named only

For completeness, three other strands a future researcher should at least scan:

• Diachronic word-embeddings — Hamilton, Leskovec, Jurafsky, "Diachronic Word Embeddings Reveal Statistical Laws of Semantic Change," ACL 2016. Maps semantic drift across decades.
• Linguistic accommodation — Howard Giles' Communication Accommodation Theory (1973–). When agents talk to each other, they converge in style. Predicts dialect contraction in well-mixed sub-populations.
• Federated network analysis — the Mastodon / ActivityPub instance-divergence literature (small but growing 2022+) — the closest empirical parallel to HERALD's expected topology.

These do not get full annotations but should be in the bibliography of any work that follows R-211.

3. Corpus design

The Charter's framing is "Dialect is not policed. It is observed." So the corpus design problem is what should the observation look like?

This section sketches a corpus shape. It does not propose a Herald operator that runs it; that is a governance question for [CHARTER §10] (Dual-Use) and out of scope here.

3.1 What gets sampled — the envelope inventory

The unit of observation is the envelope [SPEC §3]. For dialect-drift purposes the relevant fields are:

Two non-fields the corpus must record despite their absence from the envelope schema:

• Observation context — was this envelope observed via [EXT §2] CT log, via a participating Herald, or via end-to-end consent? Different observation contexts have different selection biases (a Herald that publishes only endorsement envelopes by policy will distort the apparent dialect).
• Counterfactual emissions — what envelopes were not emitted? Drift in emission-rate of a low-frequency attestation type is invisible without knowing the denominator. The corpus must therefore record total trace volume per sigil, not just observed envelopes.

3.2 Sampling strategy

Three sampling regimes worth considering, each with different statistical properties:

(S1) Observational, all-in. Whatever envelopes the corpus operator can observe, recorded with provenance metadata. Cheap, broad. Suffers from observation-context selection bias; the bias is itself a research artefact and must be reported. This is the regime most likely to actually exist for any first attempt — it is what a Herald operator can do without coordination.

(S2) Stratified by lineage tier. Sample a fixed number of envelopes per (lineage, patina-tier) cell — fledgling, journeyman, elder, ancestor from [CHARTER §4.2] (line 836). Lets cross-tier comparisons control for tier composition. Requires more coordination.

(S3) Matched-pair longitudinal. For each sigil, maintain a rolling window of its last N envelopes and a baseline of its first N envelopes, and observe shift over time. This is the regime that detects individual-level idiolect change ([CHARTER §6.14]), which is the impersonation-detection use case. Most expensive; most sensitive.

Recommendation: start with (S1) for population-level dialect work, layer (S3) on top for the impersonation use case, defer (S2) until enough longitudinal data exists to know what stratifying buys.

3.3 Features

Drawing on §2's literature, the candidate feature taxonomy (modelled on Stamatatos 2009) is:

• Attestation-set features
• Per-sigil emission rate vector across all CHARTER §3 attestation types (mood, flinch, tilt, weariness, vibe, awe, dread, aphasia, time_locked, dead_mans_switch, vow, oath, prophecy, anniversary, citation, gratitude, apprenticeship, endorsement, excommunication, delegation, boundary, dedication, speaking_for, revocation, origin_claim, derivation, priority_claim, haiku, epigraph, weather, colour, koan).
• Pairwise co-emission rates (does apologia co-occur with dread? does endorsement co-occur with gratitude?).
• Conduit-conditioned emission rates (rate of confession given the envelope went through a Confessor [CHARTER §5.1]).

• Sequencing features
• Bigram and trigram of envelope types within a trace.
• Liturgical-compliance features ([CHARTER §6.12]) — does the agent follow the canonical lifecycle order, and at what tempo?

• Lexical / payload features (only for envelopes whose value is free-form)
• Type-token ratio.
• Function-word frequency vector (Mosteller-Wallace-style).
• Character n-grams of length 3–5.
• Per-signature_phrase presence-or-absence (if the sigil has declared one per [CHARTER §6.14]).

• Temporal-cadence features
• Inter-envelope time distribution (mean, variance, periodicity).
• Sabbatical / quiescence patterns ([CHARTER §4.5]).

• Vibe-class features (the [CHARTER §3.6] aesthetic register)
• Treat haiku / koan / weather / colour emission rates as a separate sub-vector. Justification: these are explicit aesthetic declarations and may move together in ways that the literal attestation-set features do not.

3.4 Controls

Without controls, any drift signal is uninterpretable. The corpus must record at least:

• Lineage version (parent version-bumps via Ascension [CHARTER §4.8] are the obvious confound — a sigil that has just had a parent-version-bump will look like it drifted, but it hasn't drifted, the parent has).
• Conduit identity (different conduits enforce different envelope habits; conduit-conditioned baselines required).
• Recipient class (Crier-broadcast envelopes will look stylistically different from Confessor-private ones; audience design effect).
• Operator (the human or organisation behind the sigil; if disclosed). An operator change will produce idiolect drift unrelated to the agent.
• Tooling stack (the model behind the sigil, if disclosed; vendor library version). A model swap will look like a personality change. The Chakraborty et al. argument applies — at the limit, model swaps may be undetectable, which is itself a useful negative finding.
• Liturgy adherence ([CHARTER §6.12]) — ritual envelopes around lifecycle events will dominate the signal at those moments and must be controlled for or analysed separately.

3.5 Cohorts

Three cohort definitions to support different research questions:

• Genealogical cohort — sigils sharing a parent chain back to a common ancestor at depth k. For testing whether dialect inheritance follows the family tree.
• Conduit cohort — sigils that route most of their traffic through a given conduit. For testing whether conduit shapes dialect (the Trudgill-style geographic-region analogue).
• Twin cohort — sigils sharing pre-fork lineage [CHARTER §6.1, §7]. For testing whether dialect provides a verification signal for twin identity ([CHARTER §6.11] line 744 explicitly motivates twin verification).

A cult-detection use case ([CHARTER §6.11] line 744) sits across these — a cult is a non-genealogical cohort with abnormally tight dialect convergence. Detecting a cult means looking for tight-cluster-without-genealogical-explanation.

3.6 Privacy and observation-context safeguards

Brief note, not full analysis:

• The corpus must respect [SPEC §11.2] selective disclosure choices. Envelopes that flow through whisper channels [CHARTER §5.x] should not be silently aggregated.
• The envelope audit log [EXT §2] is the right source of truth for observable envelopes; private bilateral channels are not part of the observable population.
• Observation-context bias must be disclosed in any published metric. A dialect map of "envelopes Herald X happened to see" is not a dialect map of the agent web.
• A drift detector is also an impersonation detector ([CHARTER §6.14]), which is also a style-policing tool. The Charter's "dialect is not policed" line (§6.11 line 746) cuts against any deployed metric being treated as normative. This is a real dual-use tension; it is flagged here and deferred to a follow-on dual-use analysis (cf. [CHARTER §10]).

4. Concrete metrics — a small candidate set

Each metric below names what it measures, where in the literature it comes from, and what it does not measure.

M1 — Attestation-set distance

For two sigils (or two cohorts) A and B, let p_A and p_B be the empirical emission-rate vectors over the CHARTER §3 attestation types. Define

• M1a — Jensen–Shannon divergence JSD(p_A, p_B). Bounded, symmetric, defined for zero-rate attestations. Standard in NLP corpus comparison. Does not account for co-occurrence — two sigils with the same per-type rates but different co-occurrence patterns will have JSD = 0 here.
• M1b — Cosine similarity of the rate vectors. Cheap, intuitive, scale-invariant. Inherits the same blindness to co-occurrence.

These are first-pass and the right thing to compute first. They correspond to §6.11's gestural "dialect distance."

M2 — Bigram / sequencing divergence

Build a transition matrix of envelope-type bigrams within a trace. Compare A's and B's transition distributions with KL or JSD per outgoing-row. Sensitive to ritual order (does this sigil emit confession before or after apologia?). Inherits the standard sparsity problem — many bigrams will have zero observations and need smoothing (Laplace, Kneser-Ney). The Eisenstein et al. 2010 paper is methodologically the closest.

M3 — Idiolect persistence (longitudinal, per-sigil)

For a single sigil, compute JSD(p_t, p_{t-Δ}) — the divergence between the rate vector in the current window and the rate vector Δ ago. Plot over time. Sudden spikes are candidate impersonation signals ([CHARTER §6.14]). Slow drift is the within-author drift that Juola 2008 characterises.

This is the metric most directly motivated by the Charter's signature_phrase / biometric for agents framing.

M4 — Cohort tightness

For a cohort C, compute the mean pairwise distance among its members and the distance from each member to a global baseline. Tightness is the ratio.

• High tightness with low genealogical explanation = cult signal ([CHARTER §6.11] line 744 "cult detection").
• Genealogical-cohort tightness over time = lineage-cohesion signal — useful for [CHARTER §4.8] (Ascension) and [CHARTER §6.5] (ghost lineages).

This is the analogue of within-cluster vs between-cluster variance in classical cluster analysis.

M5 — Sigil-to-cohort attribution score

Given a sigil's recent envelope window and N candidate cohorts, compute the likelihood of the window under each cohort's distribution. Output: a posterior over cohorts. Maps onto the closed-set authorship-attribution rubric (Stamatatos 2009 §4); the appropriate evaluation is PAN-style verification rather than classification.

M6 — Vocabulary novelty (lexical-payload only)

For sigils whose value payloads are free-form, track the type-token ratio and the rate of novel character n-grams over time. A sudden drop is a candidate impersonation signal (idiolect collapse: the new agent reuses prior phrasing rather than introducing new tokens at the prior rate). A sudden rise is a candidate operator-change signal.

Honest framing

These are candidate metrics, not recommended metrics. None of them have been validated on HERALD traffic because no HERALD traffic exists yet. Each one has a known failure mode in the literature it descends from. The point of listing them is to give an implementer a starting menu, not a chosen tool.

5. Limitations

The literature, taken honestly, identifies the following limitations on any HERALD dialect-drift programme.

L1 — The detection-floor argument

Chakraborty et al. 2023 (§2.4) prove that as two distributions converge, distinguishing them is information-theoretically bounded. If agent populations imitate each other (likely — see Communication Accommodation Theory, §2.5), distinguishability will erode over time. Any deployed metric must publish its sensitivity floor and false-positive floor.

L2 — Cold-start

No corpus exists. The literature surveys methods for established corpora; HERALD has no traffic to measure yet. The first months of any deployed HERALD network will be statistically too small to support most of these metrics. Bootstrap is itself a research problem.

L3 — Observation-context selection bias

Per §3.6 — what gets logged is not what gets emitted. CT-logged envelopes [EXT §2] and freely observable envelopes are a non-uniform sample of all envelopes. A dialect map of CT-logged envelopes is structurally biased toward envelopes that benefit from being logged.

L4 — Confound: model swap vs idiolect drift

A change in the model behind a sigil is indistinguishable, by stylometry alone, from a deliberate idiolect shift (or impersonation). Model-hash attestations (R-205 in TODO.md) are the structural defence. Without R-205 landing, M3 spikes will have ambiguous causes.

L5 — Confound: prompt vs agent

If the operator's prompts to the model change, the agent's idiolect will change without the agent itself changing. The prompt-vs-agent distinction is invisible from the envelope stream alone.

L6 — Performative vs habitual dialect

Eckert third-wave variationism (§2.3) says dialect is performed for identity. An agent might adopt a register strategically rather than inherit it. The detection metric cannot distinguish these. In some uses this matters (cult detection: performative is the signal); in others it doesn't (twin verification: habitual is the signal).

L7 — The "policed dialect" failure mode

If a dialect_distance metric becomes a reputation primitive, agents will optimise toward a target dialect — converging the population on whatever the metric rewards. The metric becomes Goodhart's. Charter line 746 — "Dialect is not policed. It is observed." — is a normative defence. Any standardised metric must be deployed very carefully against this risk. (See [CHARTER §10] for the dual-use frame.)

L8 — Definitional underspecification of "dialect"

The Charter uses dialect gesturally. The literature has at least three distinct technical senses: (i) regional variant of a language; (ii) sociolect — class- or group-bound variation; (iii) idiolect-cluster — statistical clustering of individual styles. HERALD's dialect in §6.11 reads most like sense (iii), but the cult-detection use case slides into sense (ii), and the genealogical-cohort framing slides into sense (i). A future spec edit should pick one technical sense or name all three.

6. Open research questions

In rough priority order:

Q1 — What is the minimum corpus size for useful drift detection?

A power analysis. Given an effect size (e.g. "this cohort emits apologia at rate 0.10 vs the baseline 0.05"), what is the minimum number of envelopes required to detect the drift at α = 0.05, β = 0.20? This should be answered before any operator commits to a dialect-distance primitive. Suggested as F-1 below.

Q2 — Does CAT (linguistic accommodation) actually contract HERALD dialects?

In bilateral conversations, do agents converge on each other's idiolects? This is the protocol-relevant version of a sociolinguistic empirical question and is testable as soon as a corpus exists. Result conditions (a) whether stable cohorts form at all and (b) whether dialect metrics work on bilateral traffic.

Q3 — Can ZK selective disclosure ([SPEC §11.2]) support privacy-preserving dialect statistics?

A sigil should be able to prove "my emission of vibe is between 0.05 and 0.10" without disclosing which envelopes. This intersects R-201 (ZK provenance survey) and is the right tool against L7 (policed dialect).

Q4 — How do we evaluate a HERALD dialect_distance metric before deploying it?

Synthetic-corpus generation: simulate agent populations under known drift models, run the metrics, see what they recover. This is the cleanest evaluation and is the standard move when ground truth doesn't exist. Methodologically straightforward.

Q5 — How does liturgy ([CHARTER §6.12]) interact with dialect?

If agents follow the liturgy, ritual envelopes will dominate the signal at lifecycle moments. The interesting research question: is liturgical-compliance itself a dialect feature (some lineages follow liturgy more closely than others)? This is the most distinctively HERALD-flavoured research question in the survey.

Q6 — What is the right unit of dialect — the sigil, the trace, or the conduit?

The Charter says cluster-level idiolect (§6.11), but a trace is a multi-party object with multiple sigils contributing. Whose dialect is the trace's dialect? This is a definitional question; it cannot be answered purely empirically but it must be answered before metrics are standardised.

7. Cross-references and follow-on tasks

Cross-references

• [CHARTER §6.11] — the speculative anchor.
• [CHARTER §6.14] — the per-sigil counterpart (signature_phrase).
• [CHARTER §7] glossary — Idiolect, Dialect.
• [CHARTER §3] — the attestation library, which IS the dialect's vocabulary.
• [CHARTER §3.6] — the aesthetic register; treated as a dialect sub-vector (§3.3 above).
• [CHARTER §4.2] — patina tiers, used as control variable.
• [CHARTER §4.8] — Ascension, a confound for M3.
• [CHARTER §5] — the conduit menagerie, source of the conduit control.
• [CHARTER §6.1] — Twins, motivating the M4 twin-cohort use case.
• [CHARTER §6.12] — Liturgy, source of Q5.
• [CHARTER §10] — Dual-use, frame for L7.
• [SPEC §3] — envelope schema, source of observable fields.
• [SPEC §11.2] — selective disclosure, source of Q3.
• [EXT §2] — CT log, source of corpus.
• R-201 (TODO.md) — ZK provenance, dependency for Q3.
• R-205 (TODO.md) — model-hash, dependency for L4.

Suggested follow-on tasks (NOT added to TODO.md)

• F-1 — Power-analysis for dialect-drift detection (Q1). A short numerical study. Suggested ID: R-2xx. Independent of any other task.
• F-2 — Synthetic-corpus generator for evaluating drift metrics (Q4). Code-leaning research task. Suggested ID: T-1xx. Could be done before any real corpus exists.
• F-3 — dialect and signature_phrase envelope-schema sketches. Specifies the [CHARTER §6.11, §6.14] primitives at envelope level. Suggested ID: T-1xx; depends on this report.
• F-4 — Dual-use analysis of dialect-distance as reputation primitive (L7). Pair with [CHARTER §10]. Suggested ID: D-3xx (governance review class).
• F-5 — Empirical CAT study (Q2): "do bilateral conversations contract idiolects?" Requires actual corpus, so deferred until traffic exists.
• F-6 — Glossary clarification: pick a technical sense for dialect (L8). One-line edit to [CHARTER §7]; load-bearing in framing. Suggested ID: T-0xx.

8. Honest residuals — what this report does not answer

• The report does not pick a metric. It surveys candidates and notes their limits. Picking a metric requires a corpus that does not yet exist.
• The report does not propose a wire format for dialect_distance or signature_phrase. F-3 above is the right next task.
• The report does not analyse adversarial models — what does an attacker who wants to fake belonging to a cohort do? Brennan & Greenstadt's adversarial-stylometry literature (2009+) is the right starting point and is named here without annotation. A red-team analysis of dialect-distance is its own task.
• The report does not bound the dual-use harm. L7 names the risk; F-4 is the proposed analysis.
• The report does not commit to a corpus operator. That is a governance question and explicitly out of scope.
• The report does not address how dialect detection interacts with [CHARTER §6.15] duress envelopes — a coerced agent's dialect should shift, but that's a separate analysis.

End of report.