Five documents. One thesis.

The current raise deck: twelve visual-first slides making the case for CRBRL's $5M seed at a $27M post-money cap — SAFE, targeted Q4 2026 close.

The arc runs from the perfect storm now repricing AI memory — server DRAM contract prices compounding roughly 2.9× in nine months while agent and RAG workloads write memory continuously — to the RAM-tax problem: a 35–70× cost gap between DRAM residence and NVMe residence for the same byte. CRBRL's answer is the 768-byte pipeline: store 8× more on disk, retrieve at ≈0.98 cosine fidelity, pay a fraction.

From there: the native-format versus bolt-on byte ledger, a six-strand compound moat, modelled TCO advantage widening from ~10× to ~40× with scale, the market and SOM funnel, the two-doorway go-to-market flywheel, and the ARR and raise ladder. One idea per slide; deep detail is deliberately deferred to the business plan, whitepaper and financial plan in the adjoining rooms.

The flagship investor business plan: the commercial case for compression-native AI memory infrastructure, end to end.

It traces the evidence chain from the memory wall to the 2025–26 DRAM repricing, triangulates the vector-database market at $2.5–3.2B (2025/26) growing to $8–12B by 2030, and sets out CRBRL's shipped product, competitive position, whitespace, segmentation, dual-cohort go-to-market, unit economics and five-year financial plan.

Its distinguishing discipline is an explicit claims audit, run as a feature: retired claims are retired on the page, whitespace claims are narrowed to exactly what the evidence supports, challengers are disclosed by name, and the moat is argued as a six-strand compound rather than any single silver bullet.

The scientific spine of the thesis. The paper states a single falsifiable hypothesis: at production scale, the binding economic constraint on AI retrieval is the cost of keeping full-precision vectors resident in DRAM — and a compression-native, disk-first architecture recovers it.

Three independently testable sub-claims are evaluated against peer-reviewed literature, public price series and measured system behaviour: H1 cost divergence (DRAM versus NVMe residence), H2 fidelity retention (≈8× compression at ≈0.98 cosine fidelity, established at codec-class level — TurboQuant, ICLR 2026; RaBitQ, SIGMOD 2024), and H3 architectural compounding (native compressed format versus feature-flag quantization).

The economic model quantifies the joint effect — roughly 10× at 10M vectors, 13× at 100M, 40× at 1B for the retrieval tier — and, in keeping with the falsificationist framing, the paper specifies exactly what evidence would refute each sub-claim and documents the current codec-literature disputes rather than eliding them.

The consolidated market report: a synthesis of every segmentation and market session from May to July 2026, refreshed against the live competitive landscape.

It maps the field — the incumbent vector databases, the memory-layer startups, and the codec-commoditisation wave — and states CRBRL's positioning and unique selling features honestly against it: where the whitespace is real, where it is contested, and by whom.

From the landscape it derives the target segments and ideal customer profiles, the compound-moat framing, the dual-cohort go-to-market, and a concrete 90-day launch plan.

The finance truth source: forecast, operating budget, capital-raise ladder and reverse-engineered milestones across three scenarios, with a live companion workbook (523 formulas; every assumption editable) in which the entire plan can be flexed.

The base ladder prices every round off an ARR gate — $5M seed at $27M post (Q4 2026), $20M Series A at $100M post on ~$4M ARR, $80M Series B at $1B post on ~$20M growing to $34M — with valuation treated as an output, never an input.

Five-year operating spend of ~$72.6M is capex-light by architecture (under 3%), the cash floor and raise triggers are modelled explicitly, and the plan's honest posture is stated in the document itself: run commitments on the conservative column, manage to base, and let the optimistic column be funded only by results that have already happened.