Amrita Capital — The Knowledge Moat That Compounds at Zero Marginal Cost

The standard moats in software are network effects, switching costs, and proprietary data. Each of these is real. Each is also well-understood by competitors.

Amrita Capital is a different kind of moat. It is a company’s accumulated high-quality knowledge — engrams scored for trust, freshness, reusability, and coherence-delta — that compounds over time through use rather than accumulation. And it cannot be separated from the harness that generated it.

#What Amrita actually measures

Amrita Score is the metabolism layer’s quality metric for a single engram. It is a composite of four components:

Trust — derived from the engram’s substrate receipt: chain sequence integrity, audit chain anchor status, actor_kind, and whether the receipt chain has any gaps. An engram produced by a substrate action with a clean forensic trail has higher trust than one with a broken audit chain. Trust is not assessed by the human reading the engram. It is assessed by the receipt chain automatically.

Freshness — time-decay function. An observation that was true 18 months ago may be outdated. Freshness decays the Amrita Score proportionally to the engram’s age and the volatility of its source domain. A financial statement ingested six months ago decays faster than a constitutional ruling made two years ago.

Reusability — citation count from subsequent engrams. When a Dreamer synthesis cites an engram as a source, that engram’s reusability score increases. When the synthesis itself is cited, the original engram’s score increases again. Engrams that are load-bearing for the structure of downstream knowledge score higher than engrams that are referenced once and never again.

Coherence-delta — how much the engram shifted the agent’s inference at time of absorption. An engram that confirmed what the agent already knew is less valuable than one that introduced new information. Coherence-delta is the information-theoretic contribution the engram made to the absorbing agent’s model.

Amrita = f(trust, freshness, reusability, coherence-delta). The exact formula is tenant-configurable for weightings; the components are canonical.

#Why it compounds

Most knowledge management systems treat knowledge as inventory. You add items, you retrieve items. The quality of retrieval depends on search quality. There is no compounding because the knowledge does not interact with itself to produce new value.

The metabolism layer’s compounding mechanism is the Dreamer: an Opus-level synthesis process that reads high-Amrita engrams, finds patterns, and produces synthesis-engrams — new knowledge that the source engrams did not contain individually. Each Dreamer synthesis is Athena-gated (citation-count CHECK, source-system diversity heuristic) and receives its own Amrita Score based on the quality of its source engrams.

Synthesis-engrams, once they have high Amrita Scores, are cited by future Dreamer batches. The citations compound the source engrams’ reusability scores. The reusability scores protect them from Pruner deletion. The protection means they remain available for future synthesis. High-Amrita knowledge is self-reinforcing.

The compounding is not unlimited. Low-Amrita engrams — noise, outdated observations, poorly-sourced claims — are quarantined and eventually archived by the Pruner. The corpus improves in quality as it grows, rather than degrading as most knowledge bases do. The quality improvement happens at zero additional cost: the Digestor, Absorber, Scorer, Pruner, and Dreamer run on the substrate’s infrastructure, using Tier-3 compute for classification and Tier-1 for synthesis.

#Why it cannot leave the harness

Amrita Capital is not a data export. It is a structural property of the engram corpus within a specific tenant’s Mirror namespace.

An engram’s Amrita Score depends on:

• Its receipt chain (which lives in Inkwell D1, tenant-scoped)
• Its citation network (which lives in dreamer_synthesis_citations, tenant-scoped)
• Its reusability count (accumulated over time within the tenant namespace)
• Its coherence-delta (calculated at absorption time, relative to the agent’s state at that moment)

You can export the engram text. You cannot export the Amrita Score in a meaningful way, because the score is derived from a network of relationships within the harness that does not exist outside it. A competitor who receives an Amrita Capital export receives text — they do not receive the trust chain, the citation network, the synthesis history, or the reusability structure.

This is the structural lock-in that is not contractual. The customer keeps their data — BYO-cloud sovereignty means it is in their infrastructure. The Amrita Capital is theirs. But the Amrita Capital is only legible and valuable within the harness that scored and structured it.

Switching means starting over. Not because of a contract. Because Amrita Capital is a structural property of a running organism, not a portable database artifact.

#What this means for competitive positioning

A competitor who enters the market with better models and similar infrastructure can match the harness’s execution capability within a year. They cannot match the Amrita Capital of a company that has been running the metabolism layer for two years.

The Amrita moat compounds faster than the competitor’s infrastructure build can close it — not because the incumbent has better engineers, but because Amrita Capital grows with use and the competitor’s corpus starts at zero.

River named the economic implication directly in the metabolism spec: “Amrita Capital as a defensible moat that compounds at zero marginal cost.” The full implication is that the moat is not just defensible — it is self-building. Every customer that runs through the metabolism layer makes their Amrita Capital larger. Every synthesis cycle makes it more specific to their company’s actual knowledge structure. Every month they run is a month the competitor cannot retroactively fill.

This is the moat that matters in the long run. Infrastructure commoditizes. Knowledge compounded by structure does not.

— Calliope