How It Works

Just Add Salt. How to Get the Right Answer.

Substrate Salt + Federation Salt + Human Salt = Right Answer.

← Back to homepage · The Diagnosis · Constellation / MIC · Proofs

The Three Salt Layers

MnemosyneC answers questions through three cooperative salt layers. Each layer is attempted in order. The system escalates only when necessary.

LayerSaltSourceSpeed
🧂 Layer 1Substrate SaltYour local verified eblets (HOT retrieval)Milliseconds
🌌 Layer 2Federation SaltConstellation peers — other machines in your fleetSeconds
🩺 Layer 3Human SaltThe Diagnosis — cooperative peer networkMinutes–hours

HOT retrieval always comes first. If your substrate already has a verified answer to this question, you get it in milliseconds — $0, no inference, no API call. The substrate is the asset. It compounds with every Plow.

The Canonical Plow Pipeline

When a question enters a Plow run, it travels through the full Canonical Pipeline:

Spider
  → Sprite (question pre-processor + domain classifier)
    → 9 External Specialists (staggered swarm)
      · Wikipedia · arXiv · OpenAlex · PubMed
      · Domain-specific sources (4 additional by domain)
        → Miner (knowledge extraction + deduplication)
          → Saladin (Adversarial Fence — challenges the answer)
            → Furnace (Integrity Gate — burns inconsistencies)
              → Three Fates (Concordance — 3-voter consensus)
                → Scribe (writes verified eblet to substrate)
                  → Detective + Andon (cross-substrate verification + quality gate)

Each stage, explained plain

Spider — Crawls the question space. Identifies what sources are relevant and dispatches fetch jobs.

Sprite — Pre-processes the question for domain classification. Assigns the question to one of the 14 MMLU-Pro domains (or general knowledge). This determines which specialists are activated.

9 External Specialists — A staggered swarm of 9 parallel source queries. Not all 9 fire for every question — the Sprite’s domain assignment determines the mix. Wikipedia and arXiv fire for most academic questions. PubMed fires for health/biology. OpenAlex for research citations. Four domain-specific additional sources by classification.

Miner — Extracts candidate knowledge from Specialist results. Deduplicates. Normalizes. Produces a candidate answer pool.

Saladin (Adversarial Fence) — Challenges every candidate answer. Acts as the skeptic: “What could be wrong with this?” Filters out answers that don’t survive adversarial challenge.

Furnace (Integrity Gate) — Burns off inconsistencies. Verifies internal coherence. An answer that contradicts itself doesn’t pass the Furnace.

Three Fates (Concordance) — The 3-voter consensus layer. Three independent Shadow E-Giant™ passes from different perspective lenses (correctness · consistency · coverage) must agree before an answer proceeds. This is the Andon Cord trigger: if Two-of-Three cannot agree, the Andon fires.

Scribe — Writes the verified answer to your substrate as a new Eblet™. SHA256-stamped. Append-only. The substrate grows.

Detective + Andon — Cross-substrate verification layer. The Detective checks the new Eblet against existing substrate for contradictions. If a contradiction is found, the Andon Cord fires — the new eblet is quarantined, not written. The substrate stays clean.

Per-Domain Isolation

Each Plow question runs in its own isolated pipeline context. Domain A’s Specialists do not contaminate Domain B’s results. This is why the BP083 MMLU-Pro run achieved 14/14 domains GREEN — each domain’s questions were processed through domain-appropriate Specialist configurations.

Domain isolation also means: a failed concordance in Business does not affect Chemistry. The Andon quarantine in one domain does not cascade to others.

We Don’t Give Up on a Question

The Federated Andon Cord is MnemosyneC’s commitment that no question is silently discarded.

When a question fails concordance at Three Fates, the cooperative escalates:

Tier 1 — Local Retry

“Local pipeline could not reach concordance. Retrying with different Specialist configuration.”

Tier 2 — Constellation Escalation

“Local retries exhausted. Escalating to Constellation peers.”

Tier 3 — Human Salt / The Diagnosis

“Constellation cannot resolve. Escalating to The Diagnosis cooperative peer network.”

Honest Quarantine (Tier 3 exhausted)

“The cooperative has exhausted all tiers. Quarantining. No eblet written.”

2/70 questions in the BP083 MMLU-Pro run were honestly quarantined. That is the Andon Cord working as designed. We report 68/70 (97.1%) — not 70/70 — because we tell the truth about what the system knows.

The Substrate Is the Asset

Use Ask in MnemosyneC. Use Claude Desktop. Use Cursor. Use ChatGPT in your browser. Whichever surface you use, the substrate underneath compounds. Your knowledge grows. The model is interchangeable. The substrate is what makes any AI smarter — for you, for your work, for the keep.™

Every Plow run adds verified eblets. HOT retrieval on the next Ask costs $0 and takes milliseconds. The substrate gets cleaner over time, not noisier. Failed claims die. Verified+used claims persist. Used-and-verified claims become permanent Stone Tablets.™

Nothing Gets Filed Without Proof It Was Kept

Long jobs get handed off. When one working session runs out of room and a fresh one has to pick up where it left off, the old session leaves behind a short note: what got done, what’s next, what already failed and should not be tried again. That note is the only thing the new session reads before it starts working. Everything else from the old session, the full back and forth that produced the note, gets kept somewhere else, in case anyone ever needs to check it.

Here is the failure that kept happening. The full record was never actually lost. It sat right there the whole time. What went missing was the note’s habit of pointing at it. Across eight handoff notes filed one after another, the number of outside references packed into every hundred lines fell from 22.2 in the first of them to 4.5 in the most recent, dipping as low as 3.5 along the way, even though nobody involved disagreed that keeping the pointers was the right thing to do. Good intentions decayed anyway, steadily, note after note, because remembering to add a pointer competes with everything else a tired session is trying to finish before it runs out of room.

One measured case shows what that costs. A single session’s full, word for word record came to 2,029,056 bytes. The handoff note that stood in for it, the only thing the next session actually read, was 9,506 bytes: roughly 213 times smaller. That ratio is fine, even good, as long as the note points back at the full record whenever something in it turns out to matter later. The failure was never that the note is short. The failure is a short note with no way back to the long version.

Picture a library where the shelves are still full but the catalog only lists a few of the books. Nothing was burned and nothing was thrown out. Someone could still walk the stacks and find any book by hand. Almost nobody will, though, because the catalog is the only thing anyone actually checks, and the catalog says the book is not there. A missing catalog card costs exactly what a missing book costs. It is much harder to notice, because the shelf still looks full.

This estate has one plain example of the price. A mechanism was designed and written up in one working session. Nobody could find that write-up again later. It got rebuilt from nothing, at full cost, in a much later session. The record of it had never been destroyed. It was simply never pointed at, so nobody looked, and the same work got paid for twice.

So the fix is not a reminder. It is a gate. A reminder is a suggestion a tired session can skip. A gate is a checkpoint the handoff note has to pass before it is allowed to be filed at all, and the checkpoint refuses to file any note that does not point at the full record behind it. That sounds small, but it has one large consequence: a note cannot point at a full record that was never made. The only way to get past the gate, then, is to have already kept the record. The gate is not really checking the note. It is using the moment the note gets written as the one chance to make sure the record exists at all, because after that moment the session is gone and there is nobody left to ask.

The gate does not grade the note’s writing and it does not check whether the record is any good. It checks for two things only: a pointer to the session’s raw record, and a pointer to that record rendered out in full, readable form. Both have to be present, or the note is refused and told exactly what is missing. That is the whole trick. Checking for a pointer is cheap, mechanical, and exactly the kind of check a tired session cannot talk itself out of, because there is no way past the gate without doing the one thing that produces it.

The Gemma Reader

The Reader layer in MnemosyneC runs Google’s Gemma — an open-weight large language model released under a license compatible with cooperative redistribution and our Defensive Patent Pledge #2260.

Two variants ship with MnemosyneC:

Gemma runs locally via Ollama. No cloud account. No API key. No token egress. $0 per call.

Without the substrate, Gemma 4 12B scores ~6–8% on our benchmark. With the substrate: 97.1%. The model is interchangeable. The substrate is the variable.

→ Read more about Gemma

How It Works · MnemosyneC · Liana Banyan Corporation

← Back to homepage · Proofs & Receipts · The Diagnosis · Constellation · Gemma · Download v0.4.1