Indelis is the Synthesis Layer.

You can own 10,000 notes, a full Readwise, and a year of chat history — and produce almost nothing from them. This is the Knowledge Graveyard: not just notes nobody retained, but works nobody made. The bottleneck is not access; it is the three preconditions of synthesis. GROUND: you cannot build on what you do not trust — AI answers arrive confident and unsourced, saved notes decay, contradictions hide. GRASP: you cannot connect what you do not hold — insight happens over knowledge internalised in a mind, never over files on a disk. ATTENTION: you cannot synthesise in fragments — the tools around knowledge work are paid to interrupt the deep work creation requires.

Why now: the AI-generated-content flood makes 'work a human can stand behind' the genuinely scarce thing. Declining trust in AI output is the highest-scoring timing driver in our PESTLE scan — and it rises exactly as reputational stakes rise for anyone who ships work with their name on it. The same wave is the threat we must clear: 'chat over your files' is now a weekend project, so retrieve-and-summarise is a vanishing differentiator. As frontier assistants externalise recall and first drafts, the question stops being 'did you remember?' and becomes 'can you stand behind this?' — a higher bar no external memory can clear for the user, because the receipts must be the human's.

Three layers already have owners. Capture belongs to storage (Notion, Obsidian). Access belongs to answer engines (ChatGPT, NotebookLM). Recall belongs to SRS (Anki, Readwise, Recall, RemNote). None owns the act that justifies all three. We are not claiming the broader category is unowned — that would be dishonest with NotebookLM shipping cited flashcards and ChatGPT shipping Study Mode. The precise, defensible claim is narrower: Indelis is the first to own the synthesis act over verified, held knowledge.

What makes this a category and not a feature is a new falsifiable unit no incumbent measures: grounded synthesis (GS) — work you shipped whose claims trace back to notes you verified and hold. Storage measures how much you saved; answer engines, how fast you got a reply; SRS, recall. GS measures whether knowledge became defensible work, with time kept in the denominator — same work in less time is a better score, the structural opposite of engagement apps. The signature move behind GS is the verified connection (the Weaver): it proposes a non-obvious link across your verified graph, with auditable sub-scores, and waits for your accept. The machine proposes and grounds; the human composes.

A word on method. TAM is top-down on the categories Indelis actually sells into; SAM and SOM are bottom-up. Consumer-PKM data is thin and dispersed — roughly a 2× spread across sources — so these are order-of-magnitude ranges, not point estimates. We would rather show an honest band with low confidence than a precise-looking number we cannot defend. We deliberately do NOT fold in enterprise KM (~$20B) or e-learning ($300B+) — those credit institutional buyers we do not sell to.

The honest caveat we will not bury: SAM rests on weak third-party install aggregators, and self-host is free by architecture — so the most sovereignty-motivated users, our best-fit beachhead, may take the free Docker image and never pay. Real SAM revenue may skew below the people-count. SOM is the most assumption-driven tier: there is zero pre-launch conversion data, and the no-streaks/no-feed stance suppresses the viral levers most solo SaaS rely on. Treat it as a planning band, not a forecast.

The near-rivals matter most because they ship part of the loop. NotebookLM is free, Google-backed, and the closest shipped analogue: grounded ask-your-sources with citations, plus cited flashcards — but it is session-scoped, not a compounding verified graph, with no per-claim trust bands and no sovereignty. Readwise (~3M MAU) resurfaces what you underlined, not verified atomic claims. Recall (~500K claimed users) ships nearly the same loop — but with no per-claim trust and no human gate. scite has offered per-claim evidence since 2019 and stays niche — an honest signal that demand for this rigour may be thinner than it feels from inside the problem.

The honest threat is free-and-already-open: 'just use ChatGPT/Claude' — instant, fluent, cheap, now carrying long-term memory and a Study Mode. Our answer is not to attack the models — Indelis routes to them under a six-class consent gate, on your terms. The difference is that answer engines give answers, not authorship: nothing is yours to hold, verify, or stand behind, and they never tell you when two of your own sources contradict. We claim exactly one thing: first to own the synthesis act over verified, held knowledge.

Self-host is the free Docker image: your data, your machine, full Markdown export, no account — privacy by physics. The hosted tier is the same product run for you under per-tenant isolation, encryption-at-rest, a non-training policy, and full export. We do not fork the two; they differ by configuration. Inference passes through a six-class consent gate on a model you choose — local Ollama or your own cloud key, passed through, never stored — so we are not reselling tokens at a markup.

We will not soften the central tension: the self-hosters who fit Indelis best are the most sovereignty-motivated users, and they can take the free image and never pay. We treat free self-host as funnel, not loss — it removes every adoption objection (no migration), builds trust in a market primed for data-sovereignty (GDPR, the EU AI Act), and a worked self-hosted vault is exactly the user most likely to convert or refer. The paid wedge is not a feature wall; it is the operational burden a serious maker would rather we carry: hosting, sync, backups, managed inference.

The near-term product is mundane and singular: a personal synthesis engine for one person. Everything we sell, size, and ship is that. The vision below is not on the roadmap and is not a feature — we state it only so you understand what the architecture is bending toward. A vault worked for years passes through three stages: graveyard (a pile of the un-internalised) → organism (living, verified understanding that produces new work) → substrate (the soil from which two public goods can grow).

Two public goods fall out of long use, as a by-product. First, a substrate of identity: a durable record of what a person became and made. Second, a learning substrate for the next generation of AI: a corpus of atomic, verified, human-internalised knowledge — with the human syntheses built on it, each traceable to its ground — an antidote to model collapse and the rarest training signal there is: not text, but judgement with receipts. We note these as structural possibilities, not a plan.

We are not asking you to bet on a finished product or on traction we do not have. The substrate is strong; what still runs on stubs is precisely where the headline value sits — the full Assayer pass and the Weaver's verified connections. This round exists to retire that build risk on a wired, authenticated path, not a demo mock. Concretely, the capital buys three things: runway to wire the loop end-to-end; the chance to validate the one bet we cannot assume away — that a project deadline sustains review where a daily habit fails; and the bandwidth to reach early-access design partners in the beachhead.

This is a private conversation, founder-led. We are not naming a stage or an amount on this page; if you are the right partner for a pre-launch synthesis-layer bet on a beachhead of researchers who must stand behind their work, reach out and we will share the detail directly. One design commitment we keep regardless of the round: composition stays human — the Weaver proposes connections and the Assayer grounds them, but the work that ships is yours.