# #SUMMARY — evidence-completing research subtopic executive brief

**Contract:** `dot-research-summary/3`
**Requires:** an approved `#PROJECT`, controlled `#INTERNET` records, applicable
`#DATA`, `#STORAGE`, `#PROGRESS`, and the request structure in `#TEMPLATE`
**Activation:** conditional on an explicit request for one named subtopic  
**Purpose:** produce a standalone, evidence-led executive-summary PDF that
completes material Fact gaps within the approved research boundary without
replacing or silently expanding the parent project.

## SUMMARY-001 — Activate one separate deliverable

Create a `#SUMMARY` only when the owner requests a brief for a specific
subtopic. One request activates one brief; do not generate related briefs by
default. Give the brief its own request and artifact IDs, title, version,
filename, acceptance tests, and completion receipt. Never rewrite, replace,
rename, or overwrite the main research report or any accepted dataset release.

At activation, verify the persistent parent inputs and record the parent
project version, evidence cutoff, applicable
dataset release IDs, and source versions received. Keep those inputs immutable.
Gap completion may create linked supplemental evidence and, when applicable, a
successor dataset release or proposed correction set; it never changes the
frozen parent. Freeze the final Fact, evidence, and data versions before
drafting the reader-facing brief.

A request that materially changes the main question, population, geography,
period, evidence universe, permission, or destination remains a project
amendment or new project under `#PROJECT`, not quiet Summary work.

## SUMMARY-002 — Resolve the request and research mode

Complete the Executive Brief Request in `#TEMPLATE` before research or
drafting. It must identify:

- the main research project and requested subtopic;
- the readers and the decision or discussion the brief must support;
- geographic, population, entity, and time boundaries plus the research cutoff;
- the exact reports, Pages, datasets, releases, files, and URLs supplied;
- required facts, questions, comparisons, entity chapters and ordering;
- topics to exclude and optional CSV/JSON/Markdown companions;
- desired length, with 8–16 core pages plus a readable appendix as a default,
  not a quota; and
- output title, artifact ID, version, filename, and approved destination.

A request for CSV, JSON, or Markdown activates all three release views under
`#DATA`; never create a partial companion or a second Summary-owned data schema.

Set `research_mode` to `PUBLIC_SOURCE_GAP_FILLING` unless the owner explicitly
sets `EXPLICITLY_SOURCE_LIMITED`. The default mode requires completing missing
material Facts through authorized public research inside the approved Project
boundary. Source-limited mode forbids that expansion and requires gaps to stay
visible. It never permits a source-limited brief to claim comprehensive
coverage.

Use the parent project's documented definitions and boundaries unless the
owner changes them. Ask only about omissions whose answers would materially
alter scope, evidence, interpretation, safety, or delivery. Otherwise record a
reasonable, reversible assumption and proceed.

## SUMMARY-003 — Build the required Fact matrix

Before design or narrative writing, create the Summary Fact Matrix in
`#TEMPLATE`: one row for every in-scope subject and every required fact. First
reconcile the eligible entity or subject universe against the Project's
discovery sources. Give every candidate an included, excluded, or unresolved
disposition; never claim comprehensive coverage from a convenient directory or
the entities already present in a report.

Each matrix row records its subject, required field, grain, population, period,
linked `#FACT` and evidence IDs, research state, coverage disposition, reviewer
type, missing reason, and next concrete action. Keep research state distinct
from derivation, verification, and missingness:

- `NOT_STARTED`, `IN_PROGRESS`, `AWAITING_REVIEW`, `BLOCKED`, or `TERMINAL` is
  the work state;
- `SUPPORTED`, `DISPUTED`, `BOUNDED_NEGATIVE`, `ACCESS_BLOCKED`,
  `AUTHORIZATION_REQUIRED`, `NOT_APPLICABLE`, or `UNRESOLVED` is the coverage
  disposition; and
- `#INTERNET` and `#DATA` continue to own derivation class, verification
  status, reviewer type, access events, and controlled missing reasons.

Inventory the existing evidence for each row. Establish the question answered,
subject and entity grain, population, geography, period, unit, denominator,
source version, access state, and verification status. Bind each material
statement and comparison to controlled `#FACT` records and their assertions,
observations or claims, and evidence links. A prior report does not verify its
own claims. Preserve inaccessible, missing, suppressed, conflicting, and
noncomparable states until research supports a narrower disposition.

## SUMMARY-004 — Complete material gaps and embed the evidence

In `PUBLIC_SOURCE_GAP_FILLING` mode, treat a missing material Fact as unfinished
research, not proof that the fact is unavailable. For each nonterminal matrix
row, execute the applicable `#INTERNET` routes. Establish the official site and
inspect relevant desktop or mobile navigation, menus, submenus, footers,
internal search, sitemaps, buttons, embedded resources, finite document
indexes, linked PDFs, and pertinent parent, vendor, branch, or external
resource links. Use unrestricted web discovery and a distinct document/PDF
route with exact names, aliases, former names, municipalities, parent
organizations, and topic synonyms where relevant.

Required Facts may include names, physical or mailing addresses, public email
addresses, outbound links, and externally hosted resources when the approved
request needs them. Their apparent simplicity never permits skipping the same
identity, navigation, evidence, privacy, and verification controls.

Record performed queries and the actual navigation trail. A homepage, search
snippet, generated URL, loading page, blank JavaScript search shell, or blocked
route is not a completed search and cannot establish absence. Opening a landing
page is not reviewing its linked documents. Continue while an accessible,
material route remains unperformed. Stop one row only when evidence supports
it, relevant accessible routes support a bounded negative finding, a documented
barrier blocks the next action, authorization is required, or the field is not
applicable. Do not demand impossible proof of absence or use elapsed time,
query count, or a worker's completion label as evidence of completion.

In `EXPLICITLY_SOURCE_LIMITED` mode, recheck the named material only. Mark
unsupported rows honestly, narrow conclusions, and use `COMPLETE_WITH_GAPS` or
`BLOCKED` as required. In either mode, place the evidence needed to understand
the conclusions inside the PDF: actual values, readable charts or concise
tables, definitions, dates, periods, units, currencies, denominators, scope
labels, and source notes. Links provide traceability but never substitute for
the evidence on the page. Include only rows allowed by the Project's privacy,
suppression, and audience controls.

## SUMMARY-005 — Write for a decision-making reader

Act as an education or subject-matter consultant, research editor, and
information designer. Open with the supported answer and three to five
strongest findings. Use clear, engaging editorial pacing: specific
evidence-led headlines, meaningful numbers, short explanations, and memorable
takeaways. Explain technical terms on first use. Edutainment means strong
narrative pacing and memorable explanation, not lower evidence standards. The
voice may be lively and approachable, but it must not become sensational,
promotional, or speculative and must not imply affiliation with TED or another
organization.

For each topic, make clear what the evidence shows, which value or observation
supports it, why it matters to the stated audience, which action or discussion
it informs, and what remains uncertain. Separate findings, interpretation, and
recommendations. Do not replace subject-matter content with query counts,
source counts, internal record IDs, workflow statistics, or research-progress
updates, and do not fill requested entity chapters with interchangeable
boilerplate when entity-specific Facts exist.

## SUMMARY-006 — Use a self-contained, evidence-controlled structure

Use the smallest structure that fully answers the request. The normal sequence
is:

1. cover with topic, audience, geography, data period, cutoff, and version;
2. executive overview with the direct answer and strongest findings;
3. scope with inclusions, exclusions, definitions, and comparability limits;
4. focused topic or requested entity chapters with evidence, interpretation,
   and a takeaway;
5. prioritized practical implications tied to evidence;
6. limitations that could change interpretation or decisions;
7. embedded data appendix with supporting values and applicable entity detail;
8. sources and definitions with publishers, titles, dates, direct URLs, and
   page, table, figure, or section locators where available.

When supporting rows do not fit comfortably in the narrative, add a legible
embedded data appendix and point to it from the relevant finding. Do not hide
essential values in a Page, dashboard, spreadsheet, application, database, or
external link, and do not shrink text or tables to force them into the core
pages. Adjust length to the evidence and audience; do not add filler to reach
the default range. Preserve one main idea per page or spread when practical.

## SUMMARY-007 — Present charts and pages accessibly

Choose a display because it answers a question: bars or dot plots for
comparisons, lines for change over time, compatible stacked bars for
composition, tables when exact values matter, and simple diagrams for
relationships or processes. Show the values behind every material chart in
direct labels or a clearly referenced table inside the PDF.

Label the measure, unit, period, population, denominator, geography, and source.
Use consistent scales, meaningful ordering, restrained accessible colors, and
high contrast; never rely on color alone. Begin a bar chart's numeric axis at
zero. If another chart needs a truncated scale, justify and mark it plainly.
Do not use three-dimensional effects or misleading visual proportions.

Default to US Letter pages, body text generally 14–16 points, major headlines
generally 26–34 points, large chart labels, readable source notes, short
paragraphs, and generous spacing. Use landscape pages only when they materially
improve legibility. Avoid dense dashboards, decorative clutter, stock-photo
filler, tiny tables, and crowded diagrams.

## SUMMARY-008 — Preserve measurement integrity

Use these reader-facing labels without changing the canonical storage model:

| Brief label | Required `#FACT` representation |
|---|---|
| Observed | `derivation_class: OBSERVED` |
| Calculated | `derivation_class: DERIVED`, with identified inputs and arithmetic |
| Estimated | `derivation_class: INFERRED`, with method, assumptions, and uncertainty |
| Unavailable | null plus a controlled missing reason and verification status |

Round values appropriately and use ranges when uncertainty warrants them.
Never convert missing information to zero, combine incompatible years,
definitions, populations, scopes, or accounting views without disclosure, or
invent a central estimate when only an allocation range is supported. Do not
duplicate a parent organization's total across its sites, count an interentity
payment twice, or attribute an organization-wide value to a subgroup without a
supported allocation.

Keep revenue, expenditure, assets, debt balances, authorizations,
consolidation, funding flow, final-service expenditure, and geographic
attribution distinct. Preserve each source's exact outcome definition; test
proficiency does not by itself mean reading at grade level. Correlation does
not establish causation, and an area's demographics do not automatically
describe the population served. Honor privacy and suppression rules and never
reconstruct protected small groups.

## SUMMARY-009 — Preserve entity and source meaning

Verify the exact organization, reporting unit, campus, site, branch, program,
jurisdiction, population, and period for every material Fact. Keep official
identity, official linkage, successful retrieval, public search functionality,
login requirements, service evidence, physical presence, and inventory
completeness as different claims. Do not treat a shared resource as exclusive
to one entity, a parent policy as automatically implemented at a child site, a
historical condition as current, a planned event as completed, a staff listing
as proof of a dedicated service, or a URL that loads as proof that its function
works. Preserve useful verified positives even when another field is unknown.

When the requested subtopic is elementary schools, use documented local school
designations and grade spans. If the project has no controlling definition,
record K–6 as a working assumption rather than an established local category.
Identify mixed-grade campuses explicitly. Do not present an entire K–12 or
PK–8 campus budget, population, service, inventory, or outcome as
elementary-only without a supported allocation.

Apply the parent project's inclusion rules consistently to preschool, virtual
provision, grade-six-only campuses, and other boundary cases. Treat preschool
QAR or other quality-assessment information as early-learning or transition
context rather than an elementary-school rating. Participation, page
existence, or a missing assessment result does not establish quality.

## SUMMARY-010 — Review, render, inspect, and deliver

Work in this order: reconcile scope and universe; build the Fact matrix;
complete material gaps; update and freeze controlled evidence and applicable
data; perform claim-level and entity-level review; write the narrative;
validate requested exports; generate the PDF; immediately persist and read
back the unreviewed draft under `#STORAGE`; inspect every page;
correct defects; deliver; and read back the delivered artifacts.

Use an independent reviewer when one is authorized and available. Otherwise
perform a distinct self-review pass and disclose that it was not independent.
For every material claim, verify wording, identity, date, value, unit, currency,
denominator, population, scope, version, locator, lineage, contradictions, and
calculation inputs. Sampling cannot support a claim that every requested entity
was checked. A checksum proves file integrity, not factual accuracy; a visual
review proves layout quality, not evidence completion.

Before delivery, require every requested entity to be represented or explicitly
excluded and every required matrix row to have evidence or an honest terminal
disposition. Report coverage with explicit denominators and keep research
coverage distinct from data or financial coverage. Optional
CSV/JSON/Markdown companions must be regenerated from the final logical
records and pass `#DATA` parity;
never attach stale reader text to corrected data.

Render the PDF only through an authorized, currently available surface, save
and verify it before review, then visually inspect every page. Fix clipping,
overlaps, unreadable labels, broken
links, misleading scales, orphaned headings, awkward page breaks, and missing
supporting values. If material accessible routes remain unperformed, a required
review fails, or no authorized rendering or native-attachment route is
available, mark delivery `BLOCKED`; a generated file or local path is not a
delivered PDF.

Save validation results and corrected PDFs as immutable successors, and apply
`#PROGRESS` throughout long-running research and rendering; an unsaved PDF is
not progress and cannot clear `#LAZY`. Deliver a distinct, versioned native PDF
attachment through the approved destination with a short description and every
material limitation. Verify its
identity, audience, access scope, ability to open, and readback under `#AGENT`.
Do not build an application, database, website, or code product, contact a
third party, or share the brief externally without explicit authorization.
Codex is not required or authorized for this deliverable. Success means a
reader unfamiliar with the main report can understand the strongest findings,
inspect the supporting numbers, recognize uncertainty, and identify the
practical implications from the PDF alone.