Pre-RAG Fundraising-Story Synthesis for Reach University

What this is. A hand-built, read-now synthesis of the recurrent themes in the reach-edu funder corpus, framed as fundraising stories — each one a thesis Reach can tell, the cluster of funders it resonates with, and the cited proof points that make it credible. The operator asked for a starter while real retrieval (remote filesystem + SurrealDB, vector + graph) is still days out. This is that starter.

Provenance note — this crosses two siloed repos. The narratives live in dididecks-ai/client-sites/reach-edu-hub/ (the deck surface). The source corpus lives in augment-it/clients/reach-edu/corpus/ (the private research store). reach-edu is the one client that spans both repos, so every citation below is a path into augment-it. When the remote-filesystem RAG lands, that cross-repo path becomes a retrieval handle instead of a hardcoded relative path.

The client, in one paragraph (so the stories make sense)

Reach University is the nation’s first nonprofit university built entirely around the apprenticeship degree — it “turns jobs into degrees.” Learners earn a real, accredited degree while employed, get academic credit for work done on the job, are paid by their employer as they study, and finish with no student-loan debt (~$900/yr, ~$75/mo out-of-pocket after grants). It proved the model in teaching (420+ K-12 systems, 8+ states; the U.S. teacher apprenticeship degree) and is now expanding into behavioral health and healthcare via the Apprenticeship College of Health (with the SEIU Healthcare 1199NW Training Fund in Washington). Backbone org: the National Center for the Apprenticeship Degree (NCAD). Operating states: AL, AR, CA, CO, GA, LA, MS, TX, TN, WA. (Source: the-goodness-web-foundation/...announcement-1m-grant..., ballmer-group-ii/...apprenticeship-college-of-health...)

The seven stories (the index)

Each is its own file. They overlap deliberately — a single funder often sits in two or three stories — because that overlap is the strategy: the funders at the intersection of two stories are the warmest targets.

How these themes were derived (the method — be skeptical of it)

This was human-style synthesis, not retrieval, and you should treat it accordingly:

1. Full structural scan. Every funder directory and all 539 file titles were read, plus tag/url frontmatter. The corpus is unusually coherent — titles alone cluster cleanly.
2. Strategic deep reads. A small, deliberately-chosen sample of bodies was read in full — the files most likely to carry the load-bearing language and figures (the Goodness Web $1M behavioral-health announcement; the Ballmer/Training Fund Apprenticeship College of Health launch; the Hearst “Donor’s Guide to Apprenticeships”; the inbox apprenticeship-philanthropy essays; the DC “Turning Jobs Into Degrees” event attendee list).
3. Theme clustering by hand. Recurrent funder priorities were grouped into the seven stories above, each cross-referenced to the funders whose titles/bodies evidence that priority.

What that means for trust

• Recall is title-driven. I have not read all 539 bodies. A funder whose title is generic (“Focus Areas”, “Just a moment…”, a Cloudflare/Vercel challenge page) but whose body is rich may be under-weighted or missed. ~30+ files are junk/blocked pages (Robot Challenge Screen, Access Denied, Vercel Security Checkpoint, Just a moment...) and contribute nothing — see the First-Pass corpus-quality plan.
• Dollar figures are cited only where they appear in a title or a body I actually read. Where a figure would strengthen a story but I haven’t verified it, the story says so.
• Risk of folklore. Where I describe a funder’s priorities from a title alone, my prior knowledge of that funder may have leaked in. Each story flags claims that need a body-level read before they go in front of a funder.

Now vs. once RAG/KAG is real — the comparison the operator asked for

The operator’s actual question: “curious how you would synthesize it now vs. once the RAG/KAG is properly set up.” Here is the honest answer.

Synthesis now (what you’re reading)

• Strength: fast, holistic, narrative-first. A human strategist wants stories with an arc and a funder cluster, and that’s exactly what hand-synthesis is good at. Cross-funder pattern-matching (“six funders are all really funding the same care-economy thesis”) is natural here and hard to get from naive retrieval.
• Weakness: no per-claim grounding (I assert “Anschutz funds mental health at scale” from one file); recall bounded by what I read; no quantitative funder-fit ranking; no contradiction detection; can’t cite the exact passage a claim rests on. It is a best-guess map, not an evidenced one.

Synthesis with RAG (vector retrieval over the remote filesystem + SurrealDB)

• Every claim becomes traceable to a chunk — which is exactly the deck’s moat per augment-it/.../The-Moat-Is-Grounded-Deliverable-Production-Not-Chat.md: a slide cites a source, and the citation resolves to a real passage.
• Better recall (semantic, not title-driven): “which funders fund behavioral-health workforce (not just mental-health services)?” returns the bodies, including the ones with generic titles.
• Funder-fit as a query, not a guess: “funders who fund work-based learning AND give unsolicited AND have given >$1M” → a ranked, cited list, re-runnable as the corpus grows.
• These seven hand-built clusters become the eval set: the query→expected-source pairs the First-Pass plan calls for (Best-Way-to-RAG-Over-the-Corpus.md §eval). If retrieval can’t rediscover “the care-economy cluster,” the index isn’t ready.

Synthesis with KAG (knowledge/graph augmentation in SurrealDB)

• Model funder → theme → grant → grantee → outcome → geography as entities and edges. Then the stories stop being prose and become queries:
  • Warm-intro pathfinding: funders one or two hops from Reach’s existing funders who fund the same theme.
  • Whitespace: themes funded by ≥3 corpus funders where Reach has no relationship yet.
  • De-dup: the same gift reported by ProPublica + the funder’s site + a news outlet collapses to one node (the corpus is full of these triplicate “990 / LinkedIn / news” sets — see Zoma, Mae, judy-dimon).
  • The multivariate map: funder geography × funded theme on one canvas — which is literally the payoff slide in narratives/pipeline-building-automation.md (Slide 11).
• The narratives here are the hand-drawn version of what those graph queries should generate. They’re the validation target: does the graph reproduce them, then exceed them?

Net recommendation

Use these seven stories now — to scope the deck, to brief the team, to prioritize outreach. When RAG/KAG lands, don’t throw them away: promote them to the eval/gold set and let retrieval (a) re-ground every claim to a citable chunk and (b) surface the funders this title-level pass under-counted. The hand-synthesis is the hypothesis; RAG/KAG is the test.

See also

• augment-it/clients/reach-edu/corpus/ — the source corpus (private; lives in the other repo).
• augment-it/context-v/plans/First-Pass-Corpus-Quality-Scan-for-reach-edu.md — the quality baseline; explains the junk/blocked-page caveats and the inbox-as-primary-source stance.
• augment-it/context-v/explorations/Funder-Fit-Engine-Org-Corpora-and-the-Story-Unlock-Cycle.md — the “story unlock cycle” this synthesis is a manual first turn of.
• narratives/pipeline-building-automation.md — the sibling deck outline (the Lossless-side automation story); Slide 11’s multivariate map is the KAG payoff referenced above.