[Agora] Rebuild theme AG1: thin-content enrichment as a polymorphic continuous process

[Agora] Rebuild theme AG1: thin-content enrichment as a polymorphic continuous process

Rebuild spec — follow docs/planning/specs/rebuild_theme_template_spec.md first.

Theme anchor

• Theme: AG1 — Thin-content enrichment
• Layer: Agora
• Full description: docs/design/retired_scripts_patterns.md → AG1

Why this is first on the list

~30 deleted scripts instantiated this one pattern (enrich_thin_*, enrich_under1k_, enrich_top_hypotheses, enrich_hero_hypotheses, enrich_batch2, enrich_batch3, …). That's the textbook case of "a
recurring process that never got built, respawning as variants". Every
other theme will benefit from this one being the first rebuilt as a
canonical reference.

An open recurring task already exists for this capability
([Exchange] CI: Enrich thin hypotheses — expand next 5 descriptions,
every-2h, 1b911c77-4c4e-4dbe-a320-0cb129f1c38b). That task's previous
implementation referenced scripts/deprecated/enrich_thin_hypotheses_batch2.py
which is now gone. Rebuilding theme AG1 retires that task's
script-dependency.

Template fills

• {{THEME_ID}} = AG1
• {{THEME_NAME}} = thin-content enrichment
• {{LAYER}} = Agora
• {{LAYER_SLUG}} = agora
• {{THEME_SLUG}} = thin_enrichment
• {{CADENCE}} = every 1h
• {{CORE_JUDGMENT}} = "is this record's prose thin/stubby relative

to the available structured context, and if so, what is the richer
version that remains grounded in the same facts?"

• {{GAP_PREDICATE}} = rows where

length(prose_column) < dynamic_threshold OR
prose_version < current_rubric_version — with dynamic_threshold
read from theme_config, not hardcoded.

Polymorphism (this is the defining feature)

The rebuilt process must not be hypothesis-specific. It takes (table_name, prose_column, priority_column, context_columns[]) as
config and works over:

• hypotheses (description)
• wiki_entities (content)
• experiments (protocol / description)
• analyses (summary / findings)
• gaps (description)

One process, N content types. New content types enrolled by adding a
row to theme_config — no code change.

Discover the schema at runtime via information_schema.columns rather
than hardcoding "table has column X".

Rubric v1 sketch

The rubric asks an LLM, given (current prose, structured context,
linked KG neighborhood, citations):

Score the current prose for information density, grounding,

section coverage (0..1).

If below target, produce an expanded prose that strictly adds

information present in the context — no invention.

Return {score, expanded_prose, added_claims, citations_used,

cannot_expand_reason | null}.

Never return text unless grounded in the provided context. If context
is too thin to expand without inventing, return the cannot_expand_reason
and skip the row. (Gap-predicate will re-select after context grows.)

Outcome feedback signals (what makes this self-improving)

• Downstream quality-rubric score on the same row over the next 24h

(did the expansion land a higher score?).

• Citation verification rate (did the claims added survive a

PMID-claim check?).

• User engagement proxies if available (time-on-page, internal clicks).

Feedback worker writes to agora_thin_enrichment_outcome_feedback.
Monthly meta-worker proposes rubric_v2 based on which expansions
scored highest downstream.

Acceptance

All template acceptance criteria, plus:

experiments) in the same process. threshold + priority weights. end-to-end in testing (even if v2 is a trivial tweak). is reassigned (not recreated) to this process: its spec-path
updated, its description pointing at this spec.

File: rebuild_theme_AG1_thin_content_enrichment_spec.md

Modified: 2026-04-28 02:29

Size: 3.9 KB