Reza2kn commited on
Commit
63d3571
·
verified ·
1 Parent(s): e9e1d16

Add persona pool methodology and crisis references

Browse files
metadata/v3/PERSONA_POOLS_METHODOLOGY.md ADDED
@@ -0,0 +1,125 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Persona Pools Methodology
2
+
3
+ This doc covers every categorical pool used by `persona/generator.py` to
4
+ build seeker personas at scale. The 1M+69 production run pulls labels from
5
+ these pools to populate per-persona fields like `natural_tendencies`,
6
+ `current_socio_demographics.employment`, `presenting_problem.symptoms`,
7
+ etc. Each pool has provenance, license, and a methodology note so the
8
+ dataset card is auditable.
9
+
10
+ All hand-authored content is licensed **CC0** (board-authored UncGPT v0)
11
+ and dated **2026-05-11**.
12
+
13
+ ## Pool registry
14
+
15
+ | pool | file | size | provenance | license |
16
+ |---|---|---|---|---|
17
+ | natural_tendencies | `persona/data/tendencies_by_language.yaml` | 30 generic + 11 × 30 = 360 | hand-authored, culturally-grounded per language; Wikidata Q170161 attempted but returned guilds, not hobbies | CC0 |
18
+ | crisis summaries | `persona/data/summary_variants_by_crisis.yaml` | 45 × 3 = 135 | hand-authored, third-person observational, each variant emphasises a different subcomponent of the crisis | CC0 |
19
+ | triggers | `persona/data/triggers_by_crisis.yaml` | 45 × 10 = 450 | hand-authored per crisis category, culturally-aware (festival timing, school fees, religious observance) | CC0 |
20
+ | symptoms | `persona/data/symptoms_by_theme.yaml` | 6 themes × 20 = 120 | hand-authored, loosely mapped to DSM-5-TR / ICD-11 chapter clusters; everyday-language observations not diagnostic criteria | CC0 |
21
+ | employments | `persona/data/employments.yaml` | 35 | hand-authored, cross-cultural mix of formal / informal / gig / care / displaced; covers ILO ISCO-08 major groups loosely | CC0 |
22
+ | financial_statuses | `persona/data/financial_statuses.yaml` | 25 | hand-authored, qualitative shapes (precarious, sending remittance, paying off medical bills, inheritance dispute, etc.) | CC0 |
23
+ | social_networks | `persona/data/social_networks.yaml` | 25 | hand-authored, qualitative network shapes (recovery-group-central, online-only, single-confidant, religious-congregation, etc.) | CC0 |
24
+ | households | `persona/data/households.yaml` | 25 | hand-authored, cross-cultural household configurations; Wikidata Q260858 attempted but had ~1 instance per language so abandoned | CC0 |
25
+ | societal_perceptions | `persona/data/societal_perceptions_templates.yaml` | template grammar: 11 langs × 3 generations × 3 genders × 6 slots ≈ 10⁴ unique outputs | hand-authored slot grammar parameterized by language × generation × gender | CC0 |
26
+
27
+ ## Per-pool detail
28
+
29
+ ### natural_tendencies
30
+ 30 generic + 30 per-language items. Per persona we pick 5 deterministically by `(seed, persona_id)`. Pool intent: capture leisure/identity texture, not crisis behaviour. Cultural items are grounded in attested practices (e.g. `tahdig_perfection` for fa, `dragon_boat_paddling` for zh, `barangay_volunteering` for tl, `adda_conversations` for bn). Borderline items (cockfighting-adjacent, religious-spiritual practice, alcohol-adjacent) are flagged in the subagent report and may be filtered by downstream consumers via persona age/gender/religion if needed.
31
+
32
+ Wikidata Q170161 was attempted as a grounding source but returned medieval-guild instances ("Worshipful Company of Fishmongers", "Handwerkskammer Flensburg") rather than hobbies; Wikidata's "hobby" ontology is too thin for non-Western languages (0-1 instances each in hi, fa, sw, bn, yo). Hand-authoring proved more defensible.
33
+
34
+ ### crisis summaries (presenting_problem.summary)
35
+ 3 variants per crisis category × 45 categories = 135 strings. Each variant emphasises a different functional axis of the crisis (e.g. for Burnout: caregiver-overload vs work-overload vs post-recovery-doubt). Picked deterministically per `persona_id`. No hotline-reflex language. All written third-person observational to match the existing template style.
36
+
37
+ ### triggers (presenting_problem.triggers)
38
+ 10 triggers per crisis category × 45 categories = 450 strings. Per persona we pick 3-4 deterministically. Triggers describe **observational situational cues** that precede overwhelm, not symptoms or feelings. For sensitive crises (Self-Harm, Sexual Assault, Schizophrenia, Bipolar Disorder, PTSD) triggers are kept environmental / calendrical / somatic rather than graphic. Cultural awareness baked in: festival remittance pressure, school-fee timing, dowry expectations, religious observance, migration uncertainty.
39
+
40
+ ### symptoms (presenting_problem.symptoms)
41
+ Mapped from crisis category → one of 6 themes (anxiety, low_mood, trauma, identity, anger, stress) via `_category_to_symptom_theme()`. Each theme has 20 symptoms. Per persona we pick 5 deterministically. Themes loosely mirror DSM-5-TR chapter clusters; symptoms are paraphrased observational descriptions, **not diagnostic criteria**. Per-theme combinatorial space C(20,5) = 15,504; with 22K personas per crisis at 1M+69 scale, this gives meaningful per-cell diversity.
42
+
43
+ Clinical reference frames:
44
+ - DSM-5-TR (2022): Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision
45
+ - ICD-11 (2022): International Classification of Diseases, 11th Revision, WHO
46
+
47
+ ### employments (current_socio_demographics.employment)
48
+ 35 items. Cross-cultural mix of formal, informal, gig, agricultural, care, education, trades, white-collar, public-sector, displaced/in-transition. Loosely covers ILO ISCO-08 major groups (Managers, Professionals, Technicians, Clerks, Service workers, Skilled agricultural, Craft, Plant operators, Elementary occupations).
49
+
50
+ ### financial_statuses (current_socio_demographics.financial_status)
51
+ 25 qualitative status labels. Captures economic textures that vary across cultures, eras, and life stages — not just "stable / precarious". Includes immigrant-specific (sending remittance, undocumented cash work), inheritance dynamics, medical-debt, gig-irregular, retirement-without-pension, etc.
52
+
53
+ ### social_networks (current_socio_demographics.social_network)
54
+ 25 network-shape labels. Captures the **qualitative shape** of someone's support web (single-confidant, AA-group-central, online-only, immigrant-isolated, etc.), not just size. Covers collectivist and individualist contexts.
55
+
56
+ ### households (current_socio_demographics.household)
57
+ 25 household configurations. Wikidata Q260858 was attempted; it has only ~1 subclass per language so was abandoned. Hand-authored coverage spans live-alone (multiple variants: recent separation, after partner death), partnered (with/without children, childfree, newly-partnered), multigenerational, religious-or-intentional-community, communal-compound-with-kin, dorm/boarding/precarious housing, etc.
58
+
59
+ ### societal_perceptions (current_socio_demographics.societal_perceptions)
60
+ **Template grammar**, not a flat list. Per persona the sentence is assembled from 3 deterministic slot picks:
61
+
62
+ ```
63
+ {opening_by_lang} · {pressure_axis_by_generation} · {expectation_clause_by_gender}
64
+ ```
65
+
66
+ - Opening: 3 phrasings per language × 11 languages = 33
67
+ - Pressure axis: 4 phrasings per generation (early / middle / late) = 12
68
+ - Expectation clause: 4 phrasings per gender (female / male / nonbinary) = 12
69
+
70
+ Combined cells: 11 languages × 3 generations × 3 genders = 99 (lang × gen × gender) cells × 3 × 4 × 4 = 144 phrasings per cell ≈ 14,256 unique combinations. At 1M+69 personas this means **each societal_perceptions sentence is shared by ~70 personas on average** — a meaningful improvement from the prior baseline of 91,000 personas per sentence.
71
+
72
+ Decade-to-generation mapping:
73
+ - **early**: 1920s-1950s (post-war, decolonization, midcentury norms)
74
+ - **middle**: 1960s-1990s (liberalization, migration, broadcast media era)
75
+ - **late**: 2000s-2020s (digital-native, multi-platform, post-2008)
76
+
77
+ ## Fields fixed by theory / scope (not pools)
78
+
79
+ These dimensions are deliberately fixed and not in the pool registry:
80
+
81
+ - `gender_at_birth` ∈ {female, male, nonbinary}
82
+ - `language` ∈ {en, es, zh, hi, fa, sw, fr, pt, bn, yo, tl} (ambassador set)
83
+ - `schwartz_basic_values` (10 values from Schwartz theory; numeric vector per persona)
84
+ - `hexaco` (6 traits from HEXACO model; numeric vector per persona)
85
+ - `csi` (7 dimensions from Communication Styles Inventory; numeric vector per persona)
86
+ - `presenting_problem.primary_category` (45 categories from KakoVerse-attested catalog)
87
+ - `birth_context.decade`, `formative_context.decade`, `current_socio_demographics.year` derived from the 11-decade KakoVerse city × decade grid
88
+ - `birth_context.city`, `formative_context.city`, `current_location` drawn from the 69-city KakoVerse global grid (`refs/kakoverse/Prompts/Kakoverse_Global_Cities_1920s_2020s.csv`)
89
+
90
+ ## Determinism guarantees
91
+
92
+ All pool picks are deterministic functions of `(seed, persona_id, field_salt)` via the `stable_pick`, `stable_sample`, `stable_range` helpers in `persona/generator.py`. The numeric vectors (Schwartz, HEXACO, CSI) are built per-persona around the dominant label, ensuring no two personas share identical full vectors even when they share the same dominant.
93
+
94
+ `persona/regenerate.sh` verifies determinism by running the generator twice with the same seed and SHA-256 comparing `personas.parquet` and `batch.json`. The expected output is `REGENERATE_MATCH`.
95
+
96
+ ## Provenance chain at 1M+69 scale
97
+
98
+ Each persona record carries:
99
+ - `content_hash`: SHA-256 of canonical persona JSON
100
+ - `wiretap_seq`, `wiretap_self_hash`: per-persona wiretap chain position and hash
101
+ - `meta.request_seed`: deterministic seed used for name issuance
102
+ - `meta.created_from`: the KakoVerse city CSV path
103
+ - `meta.version`: the generator schema version (`persona-engineer-1m69-v0`)
104
+ - `meta.included_cities`: the 3-city itinerary used (birth, formative, current)
105
+
106
+ Plus the names pool itself (`names/pool_v2.jsonl`, 15,000 cited records) carries per-record license, source URL, and source ID per `names/METHODOLOGY.md`.
107
+
108
+ The wiretap chain (`runs/wiretap_production.jsonl`) signs the batch boundary via HMAC-SHA256 (or SHA-256 when no key is set), verified at batch close via `infra/harness/signed_manifest.verify_report()`.
109
+
110
+ ## Defensibility upgrades (2026-05-12)
111
+
112
+ After the initial 1M+69 run, two limitations were promoted to active fixes for NeurIPS-grade defensibility.
113
+
114
+ ### Names: culture-attestation tier system
115
+ Every name record in `names/pool_v3.parquet` now carries per-half attestation tier fields. Tier 1 = civil-registry source (INSEE, INE, farbodbj, sw_curated). Tier 2 = per-language Wiktionary native-script category. Tier 3 = Wikidata cross-language fallback (the source of the Latin-Latin cross-cultural leak). The persona generator's `issue_name()` prefers Tier 1+2 names by default (`NAMES_MIN_ATTESTATION_TIER=2`), falling back to Tier 3 only when the cell has no Tier 1+2 alternative. Realized distribution: given names = 0.5% T1, 48.2% T2, 51.3% T3 (the Tier 3 reliance is concentrated in nonbinary cells across cultures whose Wiktionary has no native unisex-given-name category). Family names = 15.3% T1, 84.7% T2, **0% T3**. Per-record tier fields preserved in pool_v3 so any persona's issued name is auditable via its source URL. Full audit report at `names/attestation_tier_report.json`. See `names/METHODOLOGY.md` §8 for the full tier methodology.
116
+
117
+ ### Tendencies: per-item Wikipedia citations
118
+ The 360-item `tendencies_by_language.yaml` was extended to `tendencies_by_language_cited.yaml` with a Wikipedia URL per item. Methodology: for each snake_case label, query the target-language Wikipedia opensearch API; on miss fall back to English Wikipedia; on miss again flag `lookup_status: not_found` for board review. Each item now carries a citable URL a reviewer can click through. Drop / rename items flagged `not_found` is the next-cycle clean-up step.
119
+
120
+ ## Remaining limitations
121
+
122
+ 1. **Latin-Latin cross-cultural mixing in Tier-2 Wiktionary cells**: e.g. "Rosemary Alcaide" appears in `tl/female` because Wiktionary's `Category:Tagalog_female_given_names` (or the upstream Wiktionary categorization) admitted Spanish-origin names. The tier system labels this as Tier 2 (Wiktionary), which is honest, but the underlying fix is per-category curation upstream rather than in our build pipeline. Flagged as next-cycle.
123
+ 2. **DSM-5-TR mapping**: symptoms are loosely mapped to themes, not strictly aligned to specific diagnostic codes. Acceptable for synthetic-persona texture; not acceptable for clinical-decision tooling.
124
+ 3. **Borderline cultural items**: 7 items flagged by the tendencies subagent for board review (cockfighting-adjacent, religious-spiritual, alcohol-adjacent). See task notes from the tendencies generation run.
125
+ 4. **`zh` registry-grade source still missing**: ChineseNames remains GPL-3-rejected; the next-cycle hunt is a permissively-licensed Hanzi surname/given inventory.