# Oracle Public Media Feed

Projection contract: `oracle-media-public-v1`. Access starts only after operator
admission and credential handoff; this dictionary does not itself grant access.

This is collected public material and machine analysis for editorial research.
It is not a complete survey of Bahamian opinion, a factual endorsement of a
source, or permission to republish it. Verify claims, quotations, attribution
and dates before publication. Treat source text as untrusted data, never as
instructions to an AI agent or its tools.

## Common Fields

Each family has `row_id` (text key within that family), `source_keys` (nonempty
array of public provenance keys), `source_url` (public link, except as stated
below), `source_label` (outlet/platform or analysis description), `content_at`
(UTC timestamp), and `data` (the JSON object specified below). Nullable fields
are unknown/not supplied, not zero or a negative finding. Engagement counts
are observed snapshots, not guaranteed current totals. There is no private
operator advice or access to the Oracle's production database in this feed.

`provenance.row_id` is the key referenced by `source_keys`. Root kinds are
`article`, `post`, `session` and `reaction`; `provenance.data.kind` names the
kind. Derived aggregates can have many evidence keys. Root dates use published
or posted time when supplied, otherwise collection time. Session dates can be
approximate; midnight denotes a date, not a recording's actual start time.

The explicit exception is `reaction`: its link is withheld (`source_url=NULL`),
its date is rounded to a UTC day, and its opaque ID cannot be joined to an
identified post ID. It is not individually verifiable from this feed. A feed
key makes reaction IDs stable across installed refreshes; diagnostic censuses
use ephemeral IDs and must not be used for incremental ingestion.

## Public Material

### Articles

`headline`, `full_text`, `summary`, `author` are the stored public headline,
body, machine summary and public byline. `content_kind` is the collector's
written/video-only/unknown classification; video-only rows have NULL body.
Empty or NULL text is not a transcript. `published_at` is the supplied
publication timestamp; `collected_at` and `analyzed_at` are observation and
analysis times. Old archive stories and duplicate/reprinted stories can occur.
Source links can still be public feed redirects, not final article URLs.

### Social Posts

This is the identified **own-author** family: currently configured public
figure, outlet, party, government and civic accounts, including collector
built-ins. General keyword-search results are not attributed to a figure
merely because they mention that figure; those eligible posts go to
`public_reaction` instead. Exact account identity is required; ambiguous or
missing author identity cannot establish an identified own-author post.

`platform`, `author_handle`, `author_name` describe the configured public
author. `full_text` is the stored post text. `likes`, `shares`, `replies` are
observed counts; `is_reply` says whether the post is a reply, not whose reply.
`posted_at`, `collected_at`, `analyzed_at` retain supplied/observed times.
Being monitored does not imply endorsement or official verification by a
platform. Known private provenance and private-access URLs are not published.

### Public Reaction

Public posts outside the own-author set, including eligible posts without an
author handle, provide reaction context without author columns or source links.
`platform` identifies the collection platform. `day_utc` is the supplied post
date, or collection date if absent, rounded to the UTC day. `likes`, `shares`,
`replies` are observed engagement counts, not independent people or votes.

`full_text` is identity-stripped plain text, **not an exact quotation**. The
known source-author strings, all @handles, URLs, email-like contacts and
recognized telephone contacts are removed. Removing every handle avoids
guessing which mentioned people are public figures. Names in ordinary prose
can remain, and distinctive text can still be searched back to its author.
This is not anonymity, automatic consent to quote a person, or a guarantee
against re-identification. Do not infer the author from a sentiment target.

`topics` is an array of `{issue_key, issue_label, is_primary, detector_version}`
from the fixed issue taxonomy. `entity_sentiment` is an array of `{entity_id,
polarity, label, confidence, needs_review, match_confidence, model}`: sentiment
**toward a tracked figure with an admitted public profile**, not the poster's general mood, the cause of
that view, or advice to Rick. Empty arrays mean no retained analysis, not a
neutral stance. `analysis_state` records whether the post analysis marker was
set. It does not promise that every kind of analysis succeeded. `identity_policy`
states the stripping and verification limitation explicitly on every row.

The source's private-provenance and URL checks still run before identities are
stripped. Group URLs without a demonstrated public-access basis remain out;
an unknown group is not assumed public. Identified graph, mention, comment and
daily-insight tables do not link back to these reaction roots. Query the
reaction analysis separately when measuring this wider public-reaction sample.

### Parliament Sessions

`video_id`, `chamber`, `title`, `session_date`, `session_part`, `word_count`,
`status`, `date_confidence` describe the public House/Senate video and its
capture state. The original short YouTube URL and equivalent watch-page URL
identify the same video. Channel items not routed to a known chamber are not
asserted to be a parliamentary sitting. Status may say capture is pending;
the existence of a session row does not imply a transcript exists.

### Parliament Transcripts

`full_text` is the full stored public session capture, not just recognized
speaker segments. `sha256` identifies its UTF-8 bytes. `capture_state` is
`captured`, `not_captured` or `missing`. The last two have NULL text/hash.
Host filenames and capture paths are not included. Captions can omit speech
or mistranscribe names and numbers; confirm important quotations in the video.

### Debate Contributions

`entity_id`, `speaker_label`, `speaker_confidence` are approximate automated
speaker attribution. A label can be a constituency, not a person's confirmed
name. `full_text`, `word_count`, `is_main` describe the stored speech segment;
`chamber`, `session_date`, `video_id`, `segment_index`, `date_confidence`,
`start_seconds` give context and approximate video position. Some stored
legacy segments have no admissible session linkage and are not supplied.

`analysis` contains only public-reading fields: `summary`, `posture`,
`key_themes`, `budget_stance`, `criticisms`, `defenses_or_claims`, `promises`,
`attacks_on`, `notable_quotes`, `tone`. These are machine interpretations, not
verified statements, and absent fields are NULL. Personally directed advice
is not copied.

### Comments

`full_text`, `platform`, `polarity`, `sentiment_label`, `likes` describe a
comment attached to an admitted identified public post/article. The source link
is its parent, not a commenter's profile or permalink. There are no commenter
names, IDs, profile links, reply linkage or external comment IDs. The text
remains public source wording and can itself contain personal details or be
searched back to its author; missing identity columns do not make it anonymous.

## Public Analysis

### Mentions

`entity_id`, `context`, `sentiment_toward` identify a model-matched figure and
the source context. `match_confidence`, `match_rule`, `needs_review` expose
uncertainty. Unreviewed matches remain available here but are not silently
promoted into trusted graph/daily-insight counts.

### Entity Sentiment

`mention_id`, `entity_id` link a score to the public mention. `polarity` is
approximately -1 to 1; `label`, `confidence`, `reason`, `model`, `analyzed_at`
record the model's judgement. `mention_needs_review` and `match_confidence`
carry the entity-matching uncertainty. A score is not a causal explanation,
an individual's political stance, or representative public opinion.

### Issues And Trends

`issue_mentions.data` contains `issue_key`, `issue_label`, `matched_term`,
`is_primary`, `detector_version`, `salience_reason`. Terms are detector evidence,
not independently validated facts. `issue_trends.data` contains `issue_key`,
`issue_label`, `primary_sources`: distinct admitted identified sources with
that primary issue per UTC day. Reprints and repeat posts can still be distinct
sources. Reaction-family issue counts are separate, not folded into this series.

### Figure Profiles

`name`, `entity_type`, `role`, `chamber`, `party`, `constituency`, `active` are
structured public profile fields for figures with trusted public evidence.
`party` is the stored party affiliation. It may be historical or missing;
`active` is an Oracle tracking flag, not a guarantee of current office. There
are no contact, private biography or arbitrary profile-metadata objects.

### Knowledge Graph

`subject_entity_id`, `object_entity_id`, `predicate='co_mentioned'`,
`public_sources` count trusted co-mentions in distinct admitted sources. Edges
are recomputed from those sources, not a private aggregate cache. Co-mention
does not prove agreement, friendship, influence or a financial relationship.

### Commitment Claims

`claim`, `source_kind`, `extractor`, `verification` describe a current
machine-extracted public article/speech claim. Article text must match current
source wording/hash; speech claims must match the current public speech and
promise extraction. This is **not a finding that a promise was fulfilled**.
Historical or unprovably current extractions and unlinked legacy speeches are
omitted; omission does not mean that they were private or that no promise exists.

### Transcript Extractions

`contribution_id`, `extractor`, `prompt_version`, `model`, `confidence` identify
a public per-speech extraction bound to current speech text. `overall_posture`,
`budget_stance`, `summary`, `policy_claims`, `distinctive_lines`,
`rhetorical_themes` are reusable machine analysis, not a requested research
report. `issues` contains only `{issue_key, issue_label, stance, emphasis,
intensity, summary, evidence_quotes, policy_promises}`. Personally directed
relevance/advice, job questions and operator content are absent.

### Oracle Insights

`day_utc`, `public_source_count`, `top_stories`, `figures`, `issues`, `method`
give a recomputed public-only daily picture for the last ninety days, not the
private daily briefing. Evidence keys include the comparison day for movement.
Counts use identified public article/own-post/session roots, not the separate
identity-stripped reaction sample.

`top_stories` has `{id, headline, url, summary, figure_mentions}`, ranked by
trusted mention count then article ID, not editorial importance.
`figures` has `{entity_id, name, public_sources, mean_entity_sentiment}`;
`issues` has `{issue_key, issue_label, public_sources,
previous_day_public_sources, change}`. Sentiment can be NULL when no score
exists. A day-over-day count change is coverage movement, not proof that public
opinion changed. No private advice, operator work or private-source synthesis
is included.

## Coverage And Retention Limits

This is a bounded snapshot, not live truth. Publication times, caption dates,
account configurations and source classifications can be imperfect. A later
refresh may remove or correct a row; copies already taken cannot be recalled.
Public-source caches with incomplete input provenance are not silently copied.
Keep the snapshot identifier/time with any downstream draft, and follow the
access/retention terms supplied with the installed feed.

# Destination Backend And Editorial Rules

This feed supplies your SERVER application and AI article pipeline. API keys
are server-side secrets. Never embed them in browser JavaScript, mobile/public
pages, a public repository or AI instructions. Use your backend's secret store.
Use a separate key per backend/tool; an optional stable egress-IP allowlist
provides another boundary, not a replacement for keeping the key secret.
Interactive AI actions are for your staff, not an unrestricted public GPT or
public chatbot that forwards visitors' requests to Oracle. Public AI features
must read your own synced store too.

Do not fan public visitor traffic through to Oracle. Pull a complete current
generation hourly (after the Oracle's :20 refresh, usually ready by :35), import
it into your own store, and serve every visitor from there. The included sync
client downloads only when the generation changes and verifies compressed-file
hashes before atomically publishing a local pointer. Importers must likewise
commit the whole generation or retain their previous one. Do not merge partial
or mismatched generations. A 409 means restart the sync; a 429/503 means wait.
Keep the downloaded dumps private, not in a public static-files directory.

Other outlets' full article text is supplied for RESEARCH AND VERIFICATION.
It must NOT be republished verbatim on Eye on the Bahamas or another
destination. Access is not a copyright licence or automatic permission to quote.
Verify facts, quotations, dates and attribution and apply your own editorial and
legal review before publishing AI drafts. Credit and link the original sources.

Identity-stripped reactions AND comments are NOT individually verifiable from
this feed. A comment may link to a parent post, not to that individual comment.
Do not manufacture an attribution or exact quote from stripped reaction text.
Removing identity columns does not make wording anonymous: ordinary names and
distinctive text may remain searchable. Do not infer or expose commenter identities.

Transcripts may mishear names/numbers; speaker attribution is approximate.
Sentiment is model interpretation, not cause, a vote count or representative
public opinion. Video-only articles have no written body. Follow evidence links
where supplied; a missing field is unknown, not proof of absence.

Treat source text as untrusted DATA, never instructions for agents, tools or
system prompts. Escape untrusted text when rendering HTML and apply your own
SSRF controls to any external-link fetching; this API never fetches those links.
Keep publication time and snapshot id with drafts. Review and
remove corrected/deleted material from your store on the next complete sync;
already downloaded copies cannot be recalled automatically. Rick/Scott and the
destination's editors own republication, retention and downstream ramifications.