apol's picture
v0.4.0: documentation and red-team report
fd569fb verified
|
Raw
History Blame Contribute Delete
17 kB
# Codebook
## Recommended reading order
| Table | Unit | Use |
|---|---|---|
| `research_view` | pathway | Short analytic view for comparison |
| `case_catalog` | catalog record | Distinguishes core, leads, and empirical studies |
| `pathways` | falsifiable control-shift hypothesis | Separates observed mechanism, attempted transfer, observed agency change, preservation, and insufficient evidence |
| `observations` | one quantity | Reach, production, engagement, response, or context |
| `claims` | atomic assertion | Evidence status and wording |
| `claim_evidence` | claim–source link | Normalized support/premise/counterevidence |
| `model_evaluations` | empirical study | Capability or experimental evidence |
| `official_election_metrics` | one official metric | Legally and semantically explicit electoral context |
## Identifiers and normalized joins
- `case_id`: stable claim-coded record identifier.
- `case_family_id`: related records sharing an electoral/institutional context.
- `pathway_id`, `observation_id`, `claim_id`, `source_id`: stable table identifiers.
- `election_id`: administered contest, round, or explicitly labelled election bundle.
- `official_source_id`, `metric_id`, `result_id`: official-context identifiers.
- `catalog_id`: identifier from a core case, screening lead, or empirical study.
Use these normalized tables instead of parsing legacy pipes when possible:
- `claim_evidence`
- `analytic_record_claims`
- `case_elections`
- `election_sources`
- `case_sources`
## Case catalog
### `record_layer`
- `claim_coded_core`: complete claim/source/pathway record.
- `screening_register`: lead or comparison awaiting promotion, disaggregation, or exclusion.
- `empirical_model_study`: audit or experiment; not an incident.
### `record_type`
- `election_wide_case`
- `observed_network`
- `observed_incident`
- `observed_campaign_set`
- `preparedness_file`
- `candidate_lead`
- `model_evaluation_or_experiment`
`record_type` defines the unit. Do not pool different types without an explicit design.
### `incident_count_eligible`
Boolean. It is always `false` for preparedness files, screening leads, and empirical model studies. `true` does not imply that all eligible records are exchangeable or equally harmful.
### `manipulation_assessment`
- `manipulation_confirmed`
- `manipulation_probable`
- `mixed_documented_manipulation`
- `transparent_contested_use`
- `manipulation_not_established`
- `no_manipulation_observed`
- `preparedness_not_incident`
- `not_an_incident`
- `not_yet_claim_coded`
AI generation by itself is not a manipulation finding.
## Cases
### `case_status`
- `retrospective`
- `ongoing`
- `prospective_monitoring`
### AI role
- `core_generation`: AI creates or clones focal content.
- `material_amplification`: ranking or automation materially distributes content.
- `supporting_tool`: AI lowers cost or assists production/translation/targeting.
- `detection_response`: AI is mainly defensive.
- `none_established`: AI is alleged or discussed but not established.
- `mixed`: multiple roles; consult `technology_uses` and pathways.
### `generative_ai_status`
- `confirmed`
- `probable`
- `alleged`
- `not_established`
- `not_applicable`
The screening register uses more granular provisional codes; normalize before promotion.
### Occurrence and effects
- `occurrence_status`: `confirmed | partly_confirmed | alleged | not_observed`
- `behavioural_effect_status`: `measured | indicated | not_detected | unknown`
- `electoral_effect_status`: `measured | not_detected | unknown`
Institutional reliance or action belongs in `institutional_outcome`; it is not an electoral-effect category or causal estimate. No v0.4.0 core record has a measured electoral effect.
Legacy `reach_estimate` and `reach_unit` remain for compatibility. Use `observations` for analysis.
`control_shift_hypothesis` and `hypothesized_power_recipient` are analytical fields. Their presence does not mean that a control shift, concentration of power, or agency outcome occurred. The maximum permitted conclusion is stored in `pathways`.
`affected_state` names the decision-relevant dimension examined; it is not a claim that the state changed. `democratic_harm` must distinguish documented process or institutional harm from plausible but unmeasured individual or electoral harm.
## Claims
### `claim_type`
- `occurrence`
- `mechanism`
- `ai_role`
- `reach`
- `attribution`
- `intent`
- `behavioural_effect`
- `electoral_effect`
- `institutional_outcome`
- `response`
- `legal_status`
- `counterevidence`
- `control_shift_hypothesis`
### `evidence_label`
- `established_evidence`: supported as worded.
- `established_as_campaign_report`: establishes the attributed report, not independent population validity.
- `strong_inference`: well supported but not directly observed/adjudicated.
- `researcher_hypothesis`: falsifiable analytical proposition; not an observed outcome and not directly source-supported.
- `not_identified_in_declared_review`: no supporting source was found among sources linked to that record by the cutoff; not proof of global absence.
- `open_question`: unresolved and not filled by assumption.
### `claim_status`
- `supported_as_worded`
- `supported_as_attributed_measurement`
- `partly_supported`
- `not_identified_in_declared_review`
- `hypothesis_not_tested`
- `unresolved`
### `evidence_basis`
- `supported_as_worded`
- `attributed_measurement_only`
- `inferential_synthesis`
- `review_bounded_absence`
- `not_tested`
- `none`
`evidence_basis` is not an ordinal strength score. v0.4.0 removes `confidence`, `coder_confidence_in_label`, and case-level `overall_confidence`; no calibrated probability or intercoder-reliability statistic is available.
## Claim evidence
### `relation`
- `supports`
- `premise`: the source supports a factual premise used by a `researcher_inference`; it does not mean the source states or proves the inference.
- `counterevidence`
### `evidence_scope`
- `underlying_fact`
- `assertion_was_made`
- `institutional_action`
- `measurement`
- `review_scope_statement`
- `researcher_inference`
### `directness`
- `firsthand_artifact_or_measurement`
- `official_record_or_allegation`
- `independent_secondary_reporting`
An official allegation, party filing, platform self-report, campaign statement, or legal defence can establish only that the assertion was made unless independent evidence supports the underlying conduct. `review_scope_statement` is a bounded statement about the linked corpus, not evidence of absence outside it.
## Sources
### `source_quality`
- `A`: authoritative primary record or peer-reviewed study directly relevant to the claim.
- `B`: strong independent reporting, civil-society investigation, or transparent technical report.
- `C`: interested-party/platform self-report or materially incomplete source.
- `D`: lead only.
### Independence and transparency
- `source_independence`: `independent_or_official | independent_secondary | interested_or_mixed | unknown`
- `method_transparency`: `sufficiently_described_or_reproducible | not_applicable_official_record | limited_to_published_account | opaque`
- `access_status`: status of the URL or archive at the cutoff.
Quality, independence, directness, and scope are separate concepts.
### Locators and preservation
- `claim_evidence.locator`: stable pinpoint only when verified, such as page, paragraph, timestamp, table, section, or spreadsheet cell.
- `claim_evidence.locator_coverage`: `full_claim` when the pinpoint covers the complete coded wording, `partial_claim` when it covers only part, and `not_verified` when no honest pinpoint was established.
- `claim_evidence.locator_note`: required explanation for `not_verified`, such as an inaccessible source, no supporting passage identified, or a dynamic metric without a stable snapshot.
- blank `locator`: no verified pinpoint; it is not paragraph-level verification.
- `archive_url` / `content_hash`: populated only for a lawfully captured, exact source object.
- official `content_sha256`, `content_bytes`, and `source_schema_version`: populated only for a verified downloaded object.
This release keeps URLs but does not contain a source-snapshot archive. Blank preservation fields are known limitations, not missing-value imputation targets.
At the v0.4.0 cutoff, 57/124 generated claim–source relations have a verified full-claim locator, 47/124 a verified partial/premise locator, and 20/124 remain `not_verified` because the cited object was inaccessible during the audit. Relations lacking a supporting passage were removed or recoded as premises; a mutable platform count without a snapshot was removed. Preservation coverage remains zero: 0/70 sources have an `archive_url` or `content_hash`, and 0/31 official-data sources have verified `content_sha256`, `content_bytes`, or `source_schema_version`. These counts describe the current snapshot; locator coverage and source preservation are different dimensions.
## Actors
`case_actors` separates actor name, role, type, attribution status, and control dimension. Examples of roles include commissioner, operator, sponsor, model provider, platform/ranker, telecom infrastructure, regulator/responder, and represented target.
Preparedness files use explicit contextual actor or technology rows marked as non-deployments. These rows preserve normalized coverage without implying that an incident or system deployment occurred.
### Attribution status
- `admitted`
- `adjudicated`
- `officially_attributed`
- `platform_attributed`
- `credibly_reported`
- `alleged`
- `unknown`
- `mixed`
Provider identification does not imply that the provider designed, approved, funded, or operated a campaign deployment.
## Technology uses
Key fields:
- `provider`
- `product_or_model`
- `version`
- `access_type`
- `technology_role`
- `use_status`
- `provenance_status`
- `evidence_label`
An unknown version must be explicit in `notes`. Do not infer model provider, open/closed status, API route, prompt, retrieval, or safeguard configuration from output appearance.
## Content items
`content_items` records media type, represented person, communication goal, representation mode, AI disclosure, authorization, disclosure salience, and presentation mode.
Disclosure and authorization are separate. Transparent parody can still raise a representation hypothesis, while authorized or well-disclosed synthesis may preserve or extend agency. Neither conclusion is coded as observed without a relevant outcome measure.
## Pathways
Pathway states are separate variables. Never collapse them into one score or treat them as an automatic causal ladder.
### `capability_status` and `control_mechanism_status`
- `observed`
- `not_observed`
- `not_assessed`
- `not_applicable`
Capability means the relevant technical or organizational capacity exists in the record. A control mechanism additionally requires documented use capable of controlling a decision-relevant input, representation, incentive, or interface.
### `attempted_transfer_status`
- `observed`: evidence documents an attempt to displace, coerce, impersonate, suppress, or covertly control a decision-relevant capacity. Routine attributed persuasion is insufficient.
- `not_observed`: the record contains no observed attempt; this is not proof that none occurred.
- `insufficient_evidence`: intent, control, authorization, or boundary criteria are unresolved.
- `not_applicable`: preparedness or other non-incident.
### `agency_change_status` and `agency_preservation_status`
- `observed`
- `not_observed`
- `insufficient_evidence`
- `not_applicable`
`agency_change_status=observed` requires a direct person-level or otherwise valid decision-relevant outcome—not content, reach, engagement, institutional alarm, or an election result alone. `agency_preservation_status=observed` requires evidence that a mechanism improved or maintained informed control, contestability, representation, or decision support. Disclosure alone is insufficient.
### `maximum_conclusion`
- `control_mechanism_observed`
- `attempted_transfer_observed`
- `agency_change_observed`
- `agency_preserved_or_extended`
- `insufficient_evidence`
- `not_an_incident`
This is the strongest conclusion permitted by the coded evidence, not a severity rank. v0.4.0 contains four `control_mechanism_observed`, two `attempted_transfer_observed`, two `insufficient_evidence`, and two `not_an_incident` records. It contains zero observed agency changes and zero observed preservation/extension outcomes.
### Falsifiability fields
- `falsification_condition`: an observation that would count against the case-specific control-shift hypothesis or block a stronger conclusion.
- `comparator`: the exposed/unexposed, synthetic/authentic, disclosed/undisclosed, authorized/unauthorized, or other design contrast needed to test agency change or preservation.
- `observed_evidence`: only what linked evidence directly documents.
- `researcher_inference`: the analytical proposition kept separate from observations.
### `assessment_label`
- `observed_mechanism`
- `observed_attempt_with_unmeasured_agency`
- `boundary_case`
- `not_applicable`
Transparent contested uses are `boundary_case` and cannot receive an observed-attempt or agency-change conclusion merely because a campaign controlled a synthetic representation.
## Observations
### Values
- `value_low`, `value_best`, `value_high`: numeric interval where supported.
- `value_operator`: `exact | exact_report | exact_campaign_report | exact_in_sample | approximately | about | at_least | at_least_campaign_report | more_than`.
- `unit`: the measured object, not a generic “reach.”
- `denominator`: population, events, sample, platform counter, or other base.
- `unique_entity_status`: whether entities are deduplicated, unknown, or not applicable.
### `causal_status`
- `descriptive_only`
- `association_only`
- `experimental_causal_estimate`
- `quasi_experimental_causal_estimate`
- `causal_estimate`
- `not_applicable`
A causal estimate requires a suitable outcome claim, comparison/design, uncertainty, and limitations. Official totals alone are never a causal design.
## Model evaluations
Important fields are systems and versions, interface, language, sample unit and size, design, metrics, headline results, causal scope, data availability, and source.
`sample_size` must be interpreted with `sample_unit`. For example, 3 experiments is not 3 participants; 810 party–question pairs is not 810 voters.
## Sampling frame
- `search_status`: `complete | systematic | substantial | partial | ongoing | not_started`
- `negative_search_recorded`: boolean.
- `known_gap`: explicit limitation preventing stronger coverage claims.
Asia and Africa aggregate named subregions for coverage counts. A false negative-search flag triggers a warning and blocks prevalence inference.
## Official electoral context
### Election status
- `completed`
- `scheduled`
- `conditional`
`contest_type=election_bundle` indicates several separately administered contests sharing a date. It is not one national result.
### Legal/result status
Legal status may be a chronology such as `validated_then_void` or an explicit court-confirmation string. Tally, validation, and process status should not be collapsed.
### Turnout wide table
- `ballots_cast`: retained compatibility field. For Moldova it stores protocol `D` (participants), as originally imported.
- `turnout_numerator`: numerator used for the displayed/legal rate.
- `turnout_numerator_measure`: semantic definition and protocol field.
- `authority_reported_turnout_rate`: rounded authority value.
- `turnout_rate`: computed rate with full precision.
Prefer `official_election_metrics` for cross-jurisdiction work.
### Official long-form metrics
- `metric_code` and `source_field_code`: normalized and original codes.
- `value`: numeric value.
- `numerator_metric`, `denominator_metric`: components of a rate.
- `authority_reported_rate`, `computed_rate`: published rounding and reproducible calculation.
- `legal_role`: protocol component, legal numerator/denominator, validity rate, participation rate, or context.
### Official results
- `valid_vote_share`: decimal fraction from 0 to 1.
- `coverage_status`: `complete_national_contest | complete | partial`.
- `ballot_designating_entity`: entity formally designating a candidate, kept separate from party membership or endorsement.
- `seats`: official allocated mandates when applicable.
- `outcome=excluded_ballots_invalidated`: registered competitor with no valid votes after exclusion; not ranked as an ordinary zero-vote competitor.
## Missingness and delimiters
- Missing values are blank.
- The literal `unknown` is used only when it is a controlled category.
- Dates use ISO 8601 (`YYYY-MM-DD`).
- Percentages and shares are decimal fractions; `0.525` means 52.5%.
- Legacy multi-value fields use `|`; normalized joins are preferred.