Version: v0_3 Date: 2026-06-26 Status: Working draft Author: Marvin Percival (DUNIN7)
What changed from v0_2: Reconciled against the canonical seed v0.12 (2026-06-07), supplied directly. The v0_2 caveat — "verify against v0.12" — is now resolved. Result: the position holds and gets stronger. The four constraints v0_2 leaned on are present and unchanged in v0.12. v0.12 also adds a second axis of difference the earlier note didn't have — the N-scope Memory model and the open-commons compounding thesis — which sharpens the contrast with Cognisee's extraction frame. No part of the position weakened.
On Cognisee: everything here is from public marketing copy and one launch press release (chief-scientist appointment, May 2026). Not a product, demo, or spec. Their framework was not yet publicly introduced as of those sources. Treat their claimed shape as a hypothesis to check against what they ship.
On the seed: now anchored to canonical seed v0.12 (2026-06-07). The version gap flagged in v0_2 is closed. Every Loomworks claim below traces to seed text.
Both companies start from the same premise:
> The valuable knowledge inside a person or institution is the part that never > got written down. Conventional LLMs learn from recorded text and can't reach > it.
Cognisee names it directly: current AI is constrained because it learns from recorded information while lacking tacit knowledge — the unwritten expertise in institutions, communities, and workplaces. Their "Tacit Reasoners" target embodied skills, contextual judgment, operational experience.
That is the segment Loomworks serves. The seed's consumers include "the bankers and sommeliers and oncologists whose knowledge is what the engagements accumulate," held "with provenance." Same target: expertise that lives in people, not documents.
The diagnosis is genuinely shared. That is why the site caught your attention.
Same diagnosis, opposite response. v0_2 named one axis (authority). v0.12 makes a second axis explicit (the commons). Both run along settled seed commitments, so neither is soft framing.
| | Cognisee | Loomworks (seed v0.12) | |---|---|---| | What it does with tacit knowledge | Extracts it into a model the system reasons and acts on | Holds it as Memory with provenance; the pipeline shapes and renders it for a reader | | Who acts | The system reasons, adapts, acts — autonomous | "AI as the production layer"; "AI-as-faithful-clerk." Operator holds authority | | State transitions | System-driven | Seed constraint: automatic state transitions on artifacts the operator has authority over are "a category error." System produces, records, signals; operator approves |
Cognisee wants the system to become the expert. Loomworks wants it to serve one. "AI-as-faithful-clerk" and "operator-authority over artifact state transitions" are listed in the seed as constraints — foundational, non- negotiable. A clerk preserves and produces under direction; it does not replace the principal.
This is the axis v0.12 surfaces that earlier drafts didn't have.
| | Cognisee | Loomworks (seed v0.12) | |---|---|---| | Where knowledge lives | Folded into a model (weights) | Memory — accumulated, with provenance, at nested scopes (engagement → organization → domain → jurisdiction) | | What happens to corrections | Dissolved into training | Seed: "A correction is a contribution. A retraction is a contribution. Both are preserved. This holds at every scope" | | How it grows | Retrain / fine-tune | "Memory grows. It is never finished." Open commons compounds — the person at assertion 101 benefits from all 100 before | | Who can reach it | The model's operator | Open by default ("all" mode); restricted only by a deliberate, recorded ACL act via OVA |
Cognisee's extraction collapses a person's trajectory into a reasoner — the wrong turns and supersessions vanish into weights. Loomworks's seed requires the opposite, and now states it at every scope: superseded assertions stay visible, trajectory preserved alongside destination, provenance first-class. The knowledge stays walkable, attributable, and correctable. That is a different artifact and a different promise to the expert whose knowledge it holds.
The commons framing is the second blade. Loomworks's default is an open, compounding knowledge commons — knowledge promotable from one engagement up to a shared domain scope, reachable by any engagement unless deliberately restricted. Cognisee's collective intelligence compounds inside the model, on their infrastructure, for their partners. One is a commons that stays legible to the people in it; the other is a brain that absorbs them.
**Cognisee builds a collective brain that reduces the need to consult the expert. Loomworks builds a provenance-preserving commons that makes consulting the expert better and keeps the expert in authority.**
| | Cognisee | Loomworks | |---|---|---| | Offering | Research org → sovereign-AI infrastructure | Product → engagement-memory environment (the four rooms) | | Buyer | Sovereign institutions, enterprises, research orgs (GCC, India, East Africa, Japan, Canada, Europe) | Every person in the system; operators running engagements; contributors; future implementers | | Cycle | Long, heavyweight, partnership-driven | An operator can pick it up and run an engagement |
research team publicly betting on tacit knowledge confirms the segment is real — while aiming at a different end state (autonomous collective intelligence vs. a provenance-preserving, operator-led commons).
long-cycle. Different sales motion and timeline from a product.
augmentation-plus-commons position cleanly before the category language hardens around extraction.
Caveat: these rest on marketing reads. If Cognisee's shipped product lands closer to provenance-preserving, operator-led augmentation than the copy implies, points 1–3 weaken. Recheck after their framework goes public.
These are commitments, not slogans — each traces to a seed constraint or section:
decides. The expert is not replaced.
every scope — never folded into a model.
including what they corrected — survives. Extraction loses it.
restriction is the exception, not the default.
decides an artifact is valid or done.
launch — their press pointed to the May 2026 Harvard / CIMC "Superintelligence for Humanity" summit.)
sovereign/enterprise targets, or is it cleanly separated? (Looks separated: Cognisee sells infrastructure to institutions; Loomworks serves the operator running the work. Worth confirming as their go-to-market clarifies.)
press release; Loomworks from unverified memory; seed not reachable.
reachable). Contrast sharpened against three named seed constraints. Version gap to v0.12 flagged.
holds and strengthens. Second axis added (N-scope Memory / open commons vs. extraction-into-weights). Version gap closed.