onasafari/sf-promises

SF Promises Dataset

Overview

Civic accountability tracker for public commitments made across San Francisco — from City Hall, SF Board of Supervisors, and SFO Airport to Chase Center and neighborhood commercial corridors. Each record is linked to its saved source document and SHA-256 hash, evaluated through deterministic code checks, and published with explicit verification exceptions, date precision levels, missingness codes, and geographic scope metadata.

Available Configurations & Split Files

Every Hugging Face Dataset Viewer configuration reads from an isolated split path under data/ to prevent mixed-schema auto-loader collisions:

Config Name	Split Path	Rows	Columns	Description
default / commitments	data/commitments/train.csv	54	72	Primary enriched tabular ledger with precision flags, missingness semantics, geo scope/centroids, evidence counts, and check exception flags
standardized_16col	data/standardized_16col/train.csv	54	16	Standardized master schema (sf_promises_master.csv) for cross-city comparisons
quarantined	data/quarantined/train.csv	13	10	Registry of planted tests and rejected claims stopped before publishing
intake_ledger	data/intake_ledger/train.csv	63	23	Outcome-blind stratified sourcing register across 10 pillars and all 11 supervisor districts

Root convenience files (promises.csv, sf_promises_master.csv, quarantined.csv, intake_ledger.csv, promises.sqlite, SCHEMA.json, data_summary.json, evidence_review_queue.json, intake_candidate_register.json, promises.json) are kept alongside the data/ split directories for direct script access.

---

Methodology & Provenance

1. Outcome-Blind Sourcing & Stratified Discovery

Commitments are discovered via a 3-stage pipeline (src/sourcing/):

1. Stratified Search Planning (SearchPlanner): Enforces outcome-blind query construction (rejecting outcome terms such as broken, failed, kept, scandal, delayed) across 10 San Francisco Life & Culture pillars and all 11 supervisor districts.
2. Retrieval (DiscoveryHarvester): Stores verbatim source bytes and SHA-256 hashes from municipal portals, allowlisted local journalism, and institutional filings.
3. Qualification (QualificationIntakeAgent): Strips any outcome labels before evaluating commitment specificity and date/scope readiness, recording audit status in intake_ledger.

2. Verbatim Quoting & Cryptographic Integrity

Every statement is checked byte-for-byte against its saved disk snapshot via SHA-256 hash and substring matching. Submissions with altered quotes or unapproved domains are stopped in quarantined.

3. Deterministic Arbitration & Visible Exceptions

Rather than hiding partial verification failures, every record exposes all_checks_passed, failed_check_count, and the specific check results:

• 47 of 54 published commitments pass all deterministic checks (all_checks_passed = true).
• 7 commitments (P82, P83, P17, P61, P37, P57, P58) remain published with explicit exception flags (all_checks_passed = false, totaling 22 failed or unrecorded checks).

4. Evidence Reconciliation Model

promises.db stores 124 total evidence rows across the project:

• 99 direct verdict-supporting/contradicting claims attached to the 54 published commitments
• 11 background context records
• 8 intake/archival discovery records
• 6 duplicate/superseded claim rows retained for lineage

Every published record (including all 17 broken commitments) carries verdict_supporting_evidence_count >= 1, distinguishing affirming vs. refuting direction from whether the evidence supports the final verdict.

5. Confidence & Impact Metrics

• Verdict Confidence ($C \in [0.0, 1.0]$): $$C = W_{\text{tier}} \times \left(1 - \frac{N_{\text{contradicting}}}{N_{\text{total}} + 1}\right)$$ where Tier 1 (Official Government) has weight $1.0$, Tier 2 (Press) has $0.85$, and Tier 3 has $0.70$.
• Impact Score ($I \in [1, 10]$): Standardized civic priority scale combining geographic scope, budgetary magnitude ($>\$10\text{M}$ capital investments score $\ge 7$), and affected population.

---

Known Biases & Research Limitations

• Sample Size & Single-Record Entities ($n=54$, 31 canonical entities): 24 of 31 canonical entities currently have only $n=1$ tracked commitment (flagged via entity_sample_warning = true). Aggregate keep/break percentages per entity should not be extrapolated as broad track records.
• Topic & District Skew: Published records skew toward Civic Life and Housing/Neighborhoods and omit Districts 2, 4, 8, and 11; the intake_ledger split documents ongoing outcome-blind collection closing those gaps.
• Coordinate Display Centroids: Only 12 records represent exact street-level pins (is_centroid = false). The remaining 42 records carry is_centroid = true (citywide, regional, or neighborhood display centroids) and must not be interpreted as physical project locations in spatial joins.
• Date Precision & Missingness: Imprecise statements (e.g., "2024") preserve precision in promise_date_precision / due_date_precision (year | month | day). Missing dates/addresses carry explicit reason codes (not_applicable | not_yet_known | not_captured).

---

Citation & BibTeX

If you use this dataset in civic-tech research, public policy analysis, or AI governance evaluation, please cite:

@dataset{sf_promises_2026,
  author = {Anders, Ariel and Contributors},
  title = {SF Promises: A Sourced, Cryptographically Verified Ledger of San Francisco Civic Commitments},
  year = {2026},
  month = {November},
  publisher = {Hugging Face},
  howpublished = {\\url{https://huggingface.co/datasets/onasafari/sf-promises}},
  note = {Version 2.1.0. Winner of Best Use of Keenable, Archil Hackathon}
}

---

Update Cadence & Versioning

• Ingest Cadence: Ingest runs periodically with automated source checks.
• Durability Audits: Automated checks re-validate source availability, archive URLs, and hash consistency.
• Version: 2.1.0 (November 2026).

---

Fields & Data Dictionary

Core Promise Data

• promise_id (string): Unique identifier (e.g., P01, P02, SFO-01)
• impact_rank (integer): Civic impact & public debate rank (#1 to #54)
• promise_text (string): Clean promise statement
• verbatim_quote (string): Exact quotation from official primary source document
• entity (string): Raw entity string
• entity_canonical_name (string): Standardized canonical entity name (e.g., Daniel Lurie, Golden State Valkyries, SFMTA)
• entity_category (enum): politician, city_department, sports_team, corporation, cultural_organization
• entity_party (string/null): Political party affiliation if politician (D, R, I)
• entity_term_start (string/null): Date when entity took office or leadership role
• entity_term_end (string/null): Date when entity term ended or ends
• role (string): Official title or position at time of commitment

Temporal Data

• promise_made_date (string): Date commitment was first uttered or published (YYYY-MM-DD)
• deadline_date (string/null): Target deadline or promised completion date (YYYY-MM-DD)
• completion_date (string/null): Date milestone was verified complete or operational
• days_remaining (integer/null): Days remaining until deadline (relative to export date, for active commitments)
• days_overdue (integer/null): Days overdue past deadline (for broken or stalled commitments)
• promise_age_days (integer): Age of promise in days since announcement
• time_to_verdict_days (integer): Days elapsed between promise announcement and formal audit verdict

Geographic & Spatial Data

• neighborhood_canonical (string): Standardized neighborhood name aligned with SF Planning Department boundaries (e.g., Tenderloin, Mission Bay, Golden Gate Park, SFO Airport, Citywide)
• address_full (string/null): Full street address or intersection
• latitude (float): WGS84 latitude coordinate
• longitude (float): WGS84 longitude coordinate
• supervisor_district (integer/null): San Francisco Supervisorial District (1 through 11)
• is_citywide (boolean): true if policy/program affects the entire city; false if site-specific

Categorization

• category_l1 (enum): Primary broad sector: Public Safety, Housing, Transportation, Infrastructure, Culture & Recreation, Economy & Commercial, Environment & Energy, Civic Governance
• category_l2 (string): Specific subcategory (e.g., Crisis & Mental Health Services, Airport Infrastructure, Bicycle & Pedestrian Infrastructure)
• secondary_categories (array): Additional related domain classifications
• tags (array): Free-form tags for filtering (e.g., homelessness, mental_health, stabilization_center)

Verdict Metadata

• verdict (enum): Outcome status (kept, in_progress, broken, undecided)
• verdict_reason (string): Plain-English factual rationale from arbitrator checks
• verdict_reason_code (string): Structured enum (on_time, delayed_but_kept, numeric_target_missed, in_progress_on_track, goal_abandoned, etc.)
• verdict_confidence (float): Confidence certainty score (0.00 to 1.00)
• verdict_evidence_count (integer): Number of supporting evidence claims filed
• verdict_contradiction_count (integer): Number of contradicting claims filed
• impact_score (integer): Standardized civic importance score (1 to 10)
• completion_percentage (integer): Completion milestone percentage (0 to 100)

Source Provenance

• source_url (string): URL of original document
• source_tier (enum): Official, News, Reference
• source_credibility_score (integer): Credibility rating (1 to 10)
• source_type_standardized (enum): official_government, news_mainstream, news_local, corporate_release, legal_record, academic_report
• source_domain_authority (integer): Domain authority rating (1 to 10)
• fact_check_status (enum): verified, disputed, unverified
• source_archived_url (string): Wayback Machine permanent archive URL
• source_hash (string): SHA-256 cryptographic digest of saved source bytes

---

Quickstart

Python (pandas)

import pandas as pd

# Load primary commitments (51 columns)
df = pd.read_csv("https://huggingface.co/datasets/onasafari/sf-promises/raw/main/promises.csv")
print(f"Loaded {len(df)} commitments")
print(df.groupby('category_l1')['verdict'].value_counts())

# Load standardized master table (16 columns)
df_master = pd.read_csv("https://huggingface.co/datasets/onasafari/sf-promises/raw/main/sf_promises_master.csv")
print(df_master[['promise_id', 'entity', 'verdict', 'confidence']].head())

Python (sqlite3)

import sqlite3
import urllib.request
import pandas as pd

urllib.request.urlretrieve(
    "https://huggingface.co/datasets/onasafari/sf-promises/resolve/main/promises.sqlite",
    "promises.sqlite"
)

con = sqlite3.connect("promises.sqlite")
df = pd.read_sql_query(\"\"\"
    SELECT entity_canonical_name, neighborhood_canonical, verdict, impact_score
    FROM promises
    WHERE category_l1 = 'Public Safety'
    ORDER BY impact_score DESC
\"\"\", con)
print(df)

Hugging Face Datasets

from datasets import load_dataset

# Load commitments configuration
dataset = load_dataset("onasafari/sf-promises", "commitments")
print(dataset['train'][0])

# Load quarantined audit test cases
quarantined = load_dataset("onasafari/sf-promises", "quarantined")
print(f"Quarantined count: {len(quarantined['train'])}")

Project Links

• Live Web Explorer: https://arii.github.io/sf-promises/
• GitHub Repository: https://github.com/arii/sf-promises