← All docs

Project BOSC — Research Course

Status: living draft. This charts what we’re investigating and, in sequence, what we built to investigate it. Inquiry drives engineering.

Where the inquiry stands. The deconstruction phase answered the two first questions: what it is (a hardened ~275 MW data-center campus on the N. Cole corridor) and who it is for (Google — now on three independent corpus sources: the AEDG release, the PAAC minutes, and the LACRPC minutes; see DOSSIER §1). The course now turns from identification to analysis and forecasting: the beneficiaries and how they correlate to the broader data-center boom, the compute capacity and workloads the plans imply (the GovCloud question is substantially answered — Liz Schwab’s “classification levels”), the Maumee watershed comparison (does Fort Wayne mirror the dilution thesis?), and localized economics beyond utility consumption. See §1.5 for the forward tracks.

Evidence discipline: claims below are tagged [verified] (read directly from a source document or committed extraction, with a citation), [filename] (suggested by a file/folder name, not yet read), or [open] (a question, not a finding). Never promote a [filename]/[open] to a fact without extracting and citing the source.

0. Where we are

A working three-stage platform (ingest → extract → analyze) with a Claude-driven agent layer. Extraction is generalized within one document kind — OPC cost estimates (contractor-agnostic Estimate + format profiles) — and validated live against the Tetra Tech roundabout sheets.

The corpus is wider than the extractor: 56 raw documents across 7 collections (aedg, oepa, recorder, permits, plans, regulatory, sanitary), of which only the 2 OPC sheets are extracted. The other ~54 are deeds, NPDES permits, building permits, and plans — genres we cannot yet parse.

1. The line of inquiry

1.1 The landscape

Activity clusters on the North Cole Street corridor in Allen County (Lima), Ohio — a single geography touched by four threads that are normally separate:

1. Privately-funded public roadwork. Six Tetra Tech OPC estimates (~$14.2M) for roundabouts and corridor work at Cole/Diller, Cole/Bluelick, the Primary Access Entrance (Beery Rd & N. Cole), Cole/West (SR 115), and the Cole & Bluelick corridors — privately funded via a BOSC Team deposit to the port authority (PAAC), dedicated to the County on completion. [verified: data/extracted/aedg/roundabouts.summary.opc.yaml]
2. County utility capacity. County-operated wastewater plants on the same roads — American II WWTP (3230 N. Cole St; NPDES 2PH00006; applicant Allen County Board of Commissioners; discharge nr. 4140 Diller Rd), American Bath WWTP (3226 N. Cole St; NPDES 2PH00007; receiving water Pike Run), and Shawnee II WWTP (NPDES 2PK00002; Ottawa→Auglaize→Maumee→Lake Erie). [verified: data/documents/oepa/*-fact-sheet.pdf]
3. Land assembly. A port-authority deed recorded 2025-11-18 referencing Amazon, plus parcel deeds tied to a developer entity. [filename: data/documents/recorder/port-authority/202511180011830-amazon-deed.pdf, recorder/bistrozzi-deeds/]
4. Developer entities. Building-permit sets and Secretary-of-State filings for codenamed entities (bistrozzi, dazzler) including LLCs such as Magenta Capital, Tilted Gate, and a “Bistrozzi Addition.” [filename: data/documents/permits/]

The thesis, now established by the corpus: a Google data-center campus is driving a coordinated build-out of public road and utility capacity on the Cole Street corridor, with the port authority as the financing/land vehicle. [verified: DOSSIER §1, §6–§7] The live questions have moved downstream — who benefits and at what public cost, what the campus consumes and runs, and how this fits the national pattern — the forward tracks in §1.5.

1.2 Core research questions

• Roadwork. What is the full scope and cost, and does the OPC reconcile to the funding instrument? Who bears cost overruns? (We have the estimates; we lack the agreement/funding docs.)
• Utilities. What discharge capacity do the WWTP permits authorize, who is the served customer base, and is new capacity tied to the development?
• Land. Grantor→grantee chain for the port-authority/Amazon parcels and the developer deeds; consideration; dates; parcel IDs.
• Entities. Who controls the codenamed LLCs (officers, agents, addresses), and how do they relate to the port authority and to Amazon?
• Environment/history. What do the 1996 CWA consent decree and sanitary CNA establish about prior obligations on this system?
• Timeline. A single chronology linking permits, deeds, SoS filings, and the roadwork agreement.

1.3 What each collection can answer

1.4 Corridor context — open leads [open]

Regional context for the cloud-consumer-economics axis. These are questions and leads, not platform facts — none are asserted in the dossier or entered into the entity graph until a document backs them. Defense-ecosystem actors appear only where the public record already places them, as [open] context.

• Mineral rights / oil-gas preemption. Recorded mineral severances or pipeline easements in the corridor, and ORC §1509.02 (ODNR preemption of local oil/gas regulation) — applicability [open]; needs the Recorder instruments.
• Rail. The CSX Toledo Subdivision ROW and any spur, with 49 U.S.C. §10501 federal preemption — existence of the ROW is documented; project use [open].
• Parallel consumers (other counties). Thor Equities / Thor Van Wert / Highland55 (Urbana, Van Wert); CyrusOne; Platon Investments / Dynamo Ventures (TX, shared-organizer overlap with Montfort) — corridor [open]; no in-corpus document yet (Platon/Dynamo rest on a third-party aggregator profile).
• Roshel / International Motors (Springfield APA, 2026-03-30). Logged strictly as corridor context, not a connection — the evidence does not link it to BOSC, and it must not enter the entity graph.

1.5 Forward tracks — from identification to forecasting

With what and who settled, five tracks carry the inquiry forward. Each is an analysis, not a missing document.

1. Beneficiaries & relation classes. Move the entity graph past mechanical edges (who conveyed to whom) to the nature of each party’s tie to the project: direct approval, direct management, direct beneficiary, possible end-user, environmental beneficiary, and relations to other government bodies. The graph stays corpus-verified — this is a classifying overlay, not new parties (Google stays an annotation, not a node). [in progress]
2. Compute capacity & workloads. Translate the disclosed plant — ~313 MW backup / ~275 MW IT, 36 cooling towers — into a capacity reading and the workload classes it fits. The GovCloud question is substantially answered: Liz Schwab addressed it in indirect language (“classification levels”). Associate the capacity with concrete workload profiles and price the GovCloud premium. [analysis]
3. The Maumee watershed comparison. Lima’s discharges screen as effectively undiluted at low flow. Does Fort Wayne — and the other large Maumee Basin dischargers — mirror that dilution thesis, or is Lima distinct? A basin-scale read using the ECHO NPDES inventory already in the corpus. [analysis]
4. Hydrology, expanded to hypotheses. Reflect that BOSC output routes to Lima (FM-2) and American II (FM-1) only; Shawnee II has no known routing (the FM-3 lead is theorized, not confirmed). Then run hypotheses at three levels — macro (Maumee basin), local (Lima loop), site (per-campus / per-WWTP). [in progress]
5. Localized economics. Move past utility consumption to quantitative local baselines — population over time, employment by industry, export orientation — so the economic argument is grounded in the place, not only in qualitative entity research. [in progress]

These feed the bigger picture, Economics, and Hydrology.

2. The engineering, sequenced to the inquiry

The dispatch seam already exists (extract_page(kind=…)); today only opc is registered. Each new genre is a new kind + structured model + profile, reusing the OPC pattern (hybrid OCR+vision read → forced-tool-use → Pydantic validation → provenance). Order is driven by inquiry leverage.

Phase A — unlock the highest-leverage genres ✅ done

1. Deeds extractor (kind=deed). Grantor/grantee/parcel/consideration/ instrument/date. [done] — validated live and swept across recorder/; reproduces Periplus’s hand-curated parcel ledger 11/11 (tests/test_periplus_crosscheck.py).
2. NPDES permit extractor (kind=npdes). Facility, permit no, applicant, outfalls, effluent limits, receiving water, public-notice dates. [done] — text-first read, swept across all 9 oepa/ docs.

Phase B — entities and breadth

3a. SoS business-filing extractor (kind=sos). [done] — vision-primary read of Secretary-of-State LLC filings (watermark.models.BusinessFiling): entity name, filing id, formation jurisdiction, registered agent + address, organizer. Swept the three permits/bistrozzi-permits/sos-* filings (all Delaware foreign LLCs); feeds the entity graph with organized_by / registered_agent edges and a shared_agent shell signal. Surfaced: Magenta Capital + Tilted Gate share a registered agent (Corporation Service Company) and organizer (Michael Montfort); Bistrozzi Addition uses CT Corporation / Scott Ziance. 3b. EPA permit-action extractor (kind=epa). [done] — the permits/ collection is not building permits. It is a stream of Ohio EPA Division of Surface Water actions on the project: Permits-to-Install (sanitary sewer), 401 Water Quality Certifications / Isolated Wetland Permits, USACE Section 404, plus dated agency correspondence (incomplete notices, comment letters). watermark.models.EpaPermitAction captures the letter header (agency, program, permit no, action, dates, applicant + address, contact + firm, project, affected resource). Text-first read; feeds the timeline (regulatory milestones) and entity graph (represented_by / affiliated_with; the EPA applicant resolves onto the same Bistrozzi node as the deeds). Sweep complete: all 30 EPA Section-401 / DSW / USACE permit-action PDFs in permits/ are extracted (data/extracted/permits/*.epa.yaml) — 26 high-confidence, 4 medium (the medium are delineation reports / cover pages where the visible number is an internal project no., not an agency-issued permit — flagged, not errors). Findings: Bistrozzi LLC and Tilted Gate LLC share a Wilmington DE mailing address (2801 Centerville Rd, PMB); a second codename “Project Dazzler” (Tilted Gate, USACE 404, principal Timothy Chadwick); counsel Vorys (Tangeman/Ziance), engineer EMH&T. 3c. Wetland-determination extractor (kind=wetland). [done] — the 2 USACE Wetland Determination Data Forms in permits/ are a different shape than the permit-action letters (a field-botany point-sample worksheet), so they get their own kind. watermark.models.WetlandDetermination captures the sampling point, location/coords, applicant, and the three regulatory criteria (hydrophytic vegetation / hydric soil / wetland hydrology → is_wetland). Both points (WD-1, WE-1, *.wetland.yaml) tie to Project BOSC on the Bistrozzi parcels (Sugar Creek Twp/Allen, former soybean field, hydric Westland-Rensselaer soils): each is formally not a wetland despite hydric soil + wetland hydrology, because hydrophytic vegetation is absent due to farming disturbance. 4. Plan read (kind=plan). [done] — an .odg is a vector drawing, so watermark.documents.read_odg extracts the titleblock/legend/callout TEXT (the authoritative content) plus the preview thumbnail, tolerating the file’s bad CRC via a raw zlib inflate. SitePlan captures project, sheet, discipline, phase, scale, the design team, and legend key_features. Validated live on LMA1A-95-SPS: “American Industrial Park Site,” Lima — Grading & Storm Plan, 95% SPS, by EMH&T (Civil) / CI Design (Architecture, Boston) / WSP USA Buildings (MEP, Troy NY); features include a substation, anti-ram barriers, security fence, containment areas, fiber duct bank — the signature of a hardened data-center campus.

Phase C — the cross-document layer (where the research lives)

5. Entity/parcel resolution. [done] — watermark.pipeline.entities: normalize party names to a canonical key, merge variants, classify conservatively (government / corporate / individual / trust / facility / water; flag Delaware as a signal, not a verdict), and link conveyances / utility operation / discharge. watermark entities + agent entities tool. Resolves Bistrozzi’s four acquisitions and the Port-Authority→Amazon edge. Facilities key on their base permit number; deeds-side trustee recitals are parsed (_parse_trustee_recital / _register_deed_party) into a trust node + its trustee persons linked trustee_of, with the conveyance running from the trust — so a deed person and an SoS organizer/principal of the same name reconcile to one node (the _split_principal de-fragmentation now also applies to deed parties).
6. Timeline assembly. [done] — watermark.pipeline.timeline: one sorted chronology across deeds/NPDES/OPC, deduped across corroborating artifacts. watermark timeline + agent timeline tool.
7. Corridor view. [done] — decided: build (the frozen corridor geometry was sitting unused and the join is the research question, not Periplus platform code). watermark.gis.corridor spatially joins every watch item (facilities + force mains) and recorded parcel onto the frozen Periplus corridor.geojson study area
   • corridor-centerline.geojson routes: in-study-area flag, distance to the nearest corridor route, the route, and station (chainage) along the roadwork road centerline (the roadway the roundabouts OPC prices). watermark corridor shows the join; watermark corridor --map adds the corridor (study area) + roadwork (road centerline) layers to gis-findings.geojson. Pure/hermetic (shapely+pyproj over committed GeoJSON, like watermark.hydrology.geo); the corridor geometry stays cited external corroboration, never edited in place.

Both built on watermark.pipeline.corpus — a loader that reads every committed extraction into one typed Corpus, classified by content shape.

Phase D — close the deferred carve-outs (as they block inquiry)

• Entity-graph polish [done] — contact-resolution noise cleaned (multi-value ; split; middle-initial merge; no self-affiliation).
• Unify the hand-authored detail YAML [done] — watermark.pipeline.corpus parses the bespoke roundabouts.detail.opc.yaml into the generic Estimate shape in memory at load (markers preserved on disk per data discipline), so it joins corpus.estimates and reconciles (7/10 — the 3 fails are the known pre-existing ROADWAY/PAVEMENT transcription gaps). No more corpus.unrecognized.
• Agent tools over the structured data [partly done] — timeline and entities tools added; program_overview / reconcile_estimate already exist.
• Section-subtotal accuracy [machinery done; live pass gated on a key] (#40) — analyze.reconcile_with_repair is the self-correcting loop (re-extract offending sections, reconcile again, up to max_rounds); watermark reconcile-repair characterizes the 3 pinned ROADWAY/PAVEMENT gaps offline (test_reconcile_repair.py), and a caller supplies the live higher-fidelity re-extractor (Opus A‑B). The reviewed artifact is characterized, not rewritten.
• extract-sweep sweep + assembled OPCSummary for the roadwork [machinery done; live sweep gated on a key] (#39) — extract.sweep_opc_pages + extract.assemble_opc_summary (watermark extract-sweep) regenerate the summary from a page range and reconcile it; tested offline on synthetic estimates. The legacy 25% / OPCSummary reconcile path is kept (the original “retire the 25%” goal was dropped as contrary to current conventions — the 25% lives in a Profile, not a hardcode).

Phase E — hydrological forecasting (water / stormwater / sewage)

The platform’s first move from deconstruction to forecasting. The Lima system is one closed flow loop on two rivers — Auglaize/Ottawa → Lima WTP → municipal + data-center demand → county/Lima WWTPs → Ottawa River — and the binding constraint is the Ottawa’s (and its tributaries’) low flow. We bring over Periplus’s Tier-0 design idea (SCS-CN + mass-balance) as document-grounded Python, not its solver/GIS stack. Every numeric input is a ProvenancedValue tagged document|connector| assumption|derived. (watermark.hydrology, see the plan.)

8. Water-balance spine + low-flow assimilative screen. [done] — Increment 1. watermark hydro (+ agent hydrology_balance tool) assembles the WWTP discharges (cited design flows from watch-items.geojson) routed to their receiving waters, grounds the abstraction reach with live USGS NWIS streamflow (Ottawa at Lima, gauge 04187100; offline-aware cache + committed fixtures), and screens each discharge against the stream’s cited 7Q10 (data/reference/hydrology/low-flow-7q10.yaml, read from the Ohio EPA fact sheets in our corpus). Headline finding, document-grounded: the two county plants on tiny tributaries discharge more than the stream’s entire 7Q10 — American Bath → Pike Run 0.01:1, American II → Dug Run 0.42:1 dilution (both violation; the American II fact sheet states the acute ratio itself as 1.3:1). Shawnee II → Ottawa mainstem has no cited 7Q10 and is skipped, not invented.

9. SCS-CN stormwater runoff. [done] — Increment 2. watermark storm (+ agent stormwater_runoff tool) runs the Tier-0 SCS chain (watermark.hydrology.solver: Type-II rainfall → curve-number excess → SCS unit-hydrograph convolution; plus a Muskingum-Cunge routing module) for a pre- vs post-development design storm over the campus footprint. Rainfall is live NOAA Atlas-14 (point query, offline cache + cited fallback); the footprint is document-sourced (recorded Bistrozzi parcels, ~340 ac); land cover (prior use “Neff Farms” → cropland; campus → impervious) and HSG (Allen County → C) are cited assumptions; curve numbers from the cited TR-55 table (cn-lookup.yaml). Finding: paving the footprint lifts the 25-yr 24-hr peak ~373 → 482 cfs (+109) and runoff volume ~75 → 100 ac-ft (+25 ac-ft of detention to hold post-development discharge to the pre-development rate). Scope note: this is the steady-state low-flow check’s complement, not a coupling — a design storm is a different flow regime than 7Q10, so the storm does not “collapse the 7Q10 dilution”; the stormwater node seam stays inert until a wet-weather scenario couples event runoff into the balance. The HSG is now SSURGO-sourced (connectors. ssurgo: the footprint’s grid-sampled dominant hydrologic soil group via USDA Soil Data Access), falling back to the cited “C” assumption offline — SSURGO actually shows the footprint is predominantly dual B/D (tile-drainable lake-plain lows) with upland B, not C.

10. Scenario diffing + dossier. [done] — Increment 3. watermark scenario (+ agent hydrology_scenario tool) evaluates baseline vs data-center buildout on the cooling consumptive-fraction knob (watermark.hydrology.scenario): the campus draws cooling water from the same Ottawa/Auglaize supply the WWTPs discharge to, and the evaporated fraction is a net basin loss. Results persist to committed, self-auditing data/scenarios/{baseline,buildout}.scenario.yaml. The new grounding that makes it land: the Ottawa mainstem 7Q10 is now cited at 0.2 cfs (Lima Refining fact sheet 2IG00001, USGS 04187100; 1Q10 = 0 cfs — the river nearly dries at design low flow, heavily abstracted upstream for Lima’s own supply), which also un-skips Shawnee II → Ottawa in the assimilative screen (now a violation, 0.04:1). watermark hydro-report renders the whole Tier-0 story as the evidence-tagged HYDROLOGY.md dossier (regenerable).

11. Sourced cooling design basis. [done] — replaces the bare “5 MGD, TBD” assumption with watermark.hydrology.cooling.derive_cooling_basis, a basis derived from disclosed campus data by two independent cited methods: top-down power × WUE (OEPA air permit P0138965: 114 gensets × 2.75 MW ≈ 313 MW backup → ~275 MW IT load × ~1.8 L/kWh evaporative WUE → ~3.1 MGD consumptive) and bottom-up blowdown × cycles (the documented 2.5 MGD FM-2 discharge at ~5 cycles → ~10 MGD upper bound). They disagree ~3× (FM-2 isn’t purely cooling blowdown), so the basis reports the 3.1–10 MGD range; the buildout scenario defaults to the conservative power-based central (overridable via --cooling-demand). Headline is now sourced and robust: even the low estimate = 4.85 cfs net basin loss ≈ 24× the Ottawa 7Q10; the upper bound ~77×. Inputs are document/assumption-tagged, demands derived.

12. Tier-1 escalation — EPA SWMM. [done] — watermark tier1 (+ agent tier1_swmm tool) runs the real EPA SWMM5 engine (watermark.hydrology.swmm, via pyswmm) for two questions Tier-0 only approximated. Detention sizing: bisect a basin’s bottom-orifice diameter until the released post-development peak matches the pre-development peak — for the 25-yr storm, SWMM finds post 579 vs pre 215 cfs, held by a ~42 ac-ft basin. Sanitary wet-weather surcharge: dry-weather base + RDII gives a ~16.9 MGD storm peak that exceeds both documented plant peak capacities (American II 3.6, Shawnee II 12.6 MGD) — i.e. SSO risk, tying to the 1996 consent decree. The engine is a native extension that may not load (it gets Killed: 9 under macOS hardened runtime on some wheels — ad-hoc codesign the swmm-toolkit dylibs to clear it); everything degrades gracefully via a subprocess availability probe (tests skip, CLI reports unavailable). Footprint/storm/plant-capacities stay document/connector-sourced; the network + hydraulic params (imperviousness, RDII R-T-K, basin geometry) are flagged assumptions, since we lack the as-built drainage network.

13. Ground the detention result in the real civil design. [done] — watermark storm-plan (+ agent storm_plan_inventory tool, watermark.hydrology.stormplan) transcribes the campus grading & stormwater plan (sheet 1A-C-3104, EMH&T 95% SPS, Not For Construction — the .odg under plans/bistrozzi-plans/) into a reviewed artifact (data/extracted/plans/lma1a.storm-inventory.yaml). The sheet’s pipe connectivity/inverts are vector geometry with no schedule table, so a routable SWMM network is deliberately not fabricated (omission over invention). What it does state we ground: the storm-structure rim population (207 labels, 820.5-828.75 ft, ~8 ft relief, document-cited) and the conveyance inventory (catch basins, curb/inlet, storm sewer, headwall outfalls, rock check dams, overland flood routing). The headline grounded fact — no detention, retention, or infiltration storage is shown (the negative is auditable: seven storage terms searched, all absent) — reframes item 12’s basin: it is the on-site control the as-drawn 95% design omits, not a modeled redesign. Wired into watermark tier1 (the detention finding now cites the sheet) and the dossier’s Tier-1 section.

14. Ground the sanitary surcharge in cited design flows + the SSO mandate. [done] — the surcharge had rested on a flat assumed base flow and an invented RDII rate. watermark.hydrology.sanitary loads a vendored cited table (data/reference/hydrology/sanitary-basis.yaml, the 7Q10-table pattern) of per-plant permitted average / peak hydraulic design flows — American II 1.2/3.6, Shawnee II 3.0/12.6 MGD (peaking factors 3.0x, 4.2x) — from the OEPA NPDES permits + watch-items; American Bath’s peak is omitted (uncited, not guessed). The surcharge now compares the campus’s wet-weather contribution against each plant’s documented wet-weather headroom (peak − average): 16.9 MGD vs American II’s 2.4 and Shawnee II’s 9.6 MGD. The campus dry base is the document-cited 2.5 MGD FM-2 discharge (RDII R stays a flagged assumption). The decisive context is regulatory and now surfaces as a finding: the collection system is already under a 2005 OEPA mandate to eliminate all SSO bypassing by 2015, with $11.8M of storm-water I/I remediation and a 21→48-inch trunk rebuilt purely to equalize wet-weather I/I (1996 federal CWA consent decree; Allen County CNA-2005) — so the headroom is documented as effectively already spent. watermark tier1 + agent sanitary_basis tool + the dossier’s Tier-1 section. (Indian Brook pump-station as-built is scan-only; the discipline-agnostic kind=sanitary engineering extractor now exists (#41) — the .sanitary.yaml is gated on a keyed vision pass, #124.)

3. Immediate next steps (proposed)

1. Confirm/adjust this course.
2. Build the deeds extractor and run it on the port-authority/Amazon deed — first real test of a non-OPC kind, and it directly advances the land thread.
3. In parallel, the NPDES extractor over oepa/ (cheap, text-first) to pin utility capacity and dates.
4. Stand up a minimal timeline from whatever the first extractions yield.

4. Decisions for you [open]

• Scope/priority: is the land-ownership thread the priority, or utilities, or roadwork-funding? That reorders Phase A/B.
• GIS: do we rebuild any corridor/parcel geography in BOSC, or stay document-only for now?
• Output: what’s the deliverable of the research itself — a briefing for the County Engineer, a reconciliation memo, an entity/timeline dossier? (First cut landed: DOSSIER.md — an evidence-disciplined synthesis of the graph + timeline, regenerable as the corpus grows.)