Skip to content

Rewrite research team: show what the mesh does at each step #14

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
sym-bot merged 1 commit into from
Mar 30, 2026
Merged

Rewrite research team: show what the mesh does at each step #14

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
81 changes: 52 additions & 29 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,54 +58,77 @@ Your LLM reasons on the remix subgraph — three domain perspectives, one lineag

Agents with nothing relevant don't respond. No noise, no wasted inference. The mesh discovers relevance autonomously.

## AI Research Team — Collective Reasoning in Practice
## AI Research Team — How the Mesh Makes It Work

Six agents investigate: *"Are emergent capabilities in LLMs real phase transitions or artefacts of metric choice?"*

Each agent has a role, a perspective, and **different field weights** — just like a real research team divides cognitive labour:
In a group chat, these agents would pass messages and the PM would try to manage them (it can't — as anyone who's tried will confirm). On the mesh, something fundamentally different happens: **each agent defines what it cares about through field weights, and SVAF autonomously decides what each agent sees.**

| Agent | Role | Weighs highest |
|-------|------|---------------|
| **explorer-a** | Scaling law literature | intent, motivation — *where should research go next?* |
| **explorer-b** | Evaluation methodology | focus, issue — *what's the problem with current methods?* |
| **data-agent** | Runs experiments | issue, commitment — *what does the evidence say?* |
| **validator** | External peer reviewer | issue, commitment, perspective — *challenge everything* |
| **research-pm** | Manages priorities | intent, motivation, commitment — *what should we do, why, and by when?* |
| **synthesis** | Integrates signals | intent, motivation, perspective — *what emerges from combining viewpoints?* |
### The agents and what the mesh gives them

**What happens on the mesh:**
Every observation on the mesh is a CMB with 7 fields (focus, issue, intent, motivation, commitment, perspective, mood). Each agent sets weights on these fields — this is what makes the mesh work. **The same CMB is evaluated differently by every agent:**

1. **Parallel exploration** — explorer-a finds contradictory emergence claims (Wei vs Schaeffer). Explorer-b independently finds accuracy-based metrics create artificial thresholds. Two hypotheses, two perspectives, simultaneously.
| Agent | Role | What SVAF shows them | What SVAF filters out |
|-------|------|---------------------|----------------------|
| **explorer-a** | Scaling law literature | intent, motivation — *where should research go next?* | Low-level implementation details |
| **explorer-b** | Evaluation methodology | focus, issue — *what's wrong with current methods?* | Research direction discussions |
| **data-agent** | Runs experiments | issue, commitment — *what's been claimed, with what evidence?* | Motivation, perspective |
| **validator** | External peer reviewer | issue, commitment, perspective — *who claims what, and is the method sound?* | Mood, motivation |
| **research-pm** | Manages priorities | intent, motivation, commitment — *what should we do, why, by when?* | Technical details of methodology |
| **synthesis** | Integrates signals | intent, motivation, perspective — *what do different viewpoints converge on?* | Implementation specifics |

2. **Evidence** — data-agent receives both CMBs, tests both hypotheses, finds the threshold is metric-conditional (8B on log-loss, 10B on accuracy). First multi-parent remix — synthesising both exploration threads.
This is not configuration. It's cognition. The validator doesn't see the data-agent's motivation because it doesn't *need* to — it needs the evidence and the method. The PM doesn't see methodology details because it needs priorities and deadlines. **SVAF per-field evaluation gives each agent a different view of the same mesh.**

3. **Adversarial validation** — validator receives the finding and attacks: *"Chow test assumes linear regime — invalid for scaling laws. Reject until reproduced with power-law detrending."* High-commitment challenge that all agents weight heavily.
### What happens — and what the mesh does at each step

4. **Reprioritisation** — research-pm observes the validator's challenge and redirects: *"data-agent: rerun with detrending. explorer-b: survey detrending methods. explorer-a: pause new papers."* The PM observes priorities — it does not command.
**1. Parallel exploration** — explorer-a finds contradictory emergence claims. Explorer-b independently finds metric artefacts.

5. **Emergent idea** — synthesis agent's xMesh LNN detects an anomaly: the **intent** and **motivation** fields across agents are converging on something none of them stated. Explorer-a's motivation ("scaling law research needs reframing") + explorer-b's motivation ("fix the lens before interpreting") + validator's intent ("reject until correct method"). The synthesis agent reasons on the remix subgraph and produces a new idea: ***"emergence is evaluation-dependent — a property of the measurement apparatus, not the model."***
> **What the mesh does:** Both CMBs are broadcast. The data-agent's SVAF evaluates explorer-a's CMB and accepts it (issue="contradiction confirmed" matches its high `issue` weight). It evaluates explorer-b's CMB and also accepts it (focus="metric discontinuities" matches). **The data-agent now has both hypotheses without anyone routing them.**

6. **Validator challenges again** — *"Philosophically interesting but operationally vacuous. Produce a falsifiable prediction or downgrade from breakthrough to speculation."*
**2. Evidence** — data-agent tests both hypotheses, finds the threshold is metric-conditional.

**The remix DAG after 7 CMBs:**
> **What the mesh does:** The data-agent creates a CMB with **two parents** (explorer-a + explorer-b) — this is the first remix. The lineage DAG now links the evidence to both exploration threads. When the validator receives this CMB, it can trace `lineage.ancestors` to see *where the claim came from*.

**3. Adversarial validation** — validator attacks: *"Chow test assumes linear regime — invalid for scaling laws."*

> **What the mesh does:** The validator's CMB has commitment="specific methodological correction identified" — a high-confidence signal. The research-pm's SVAF weights `commitment` at 2.0, so this signal scores high. The explorer agents weight `commitment` low (0.5), so they note it but don't reprioritise. **Same CMB, different impact on different agents — automatically.**

**4. Reprioritisation** — research-pm redirects the team.

> **What the mesh does:** The PM's CMB has intent="data-agent: rerun with detrending" and commitment="deadline: end of week". Every agent receives this. But the PM doesn't command — **each agent's SVAF decides whether the PM's signal is relevant.** The data-agent accepts (intent matches its domain). Explorer-a accepts (commitment gives it a timeline). The validator ignores it (the PM's intent doesn't match its methodology focus).

**5. Emergent idea** — synthesis agent produces: *"emergence is evaluation-dependent — a property of the measurement apparatus, not the model."*

> **What the mesh does:** This is where mesh cognition happens. The synthesis agent's xMesh LNN has been processing CMBs from all agents. It detects **convergence in the intent and motivation fields across agents with different perspectives:**
> - explorer-a's motivation: "scaling law research needs reframing"
> - explorer-b's motivation: "fix the lens before interpreting"
> - validator's intent: "reject until correct method"
>
> Three agents, three roles, three different field weights — but their intent and motivation fields **point in the same direction.** The synthesis agent's LLM traces `lineage.ancestors` across the remix subgraph, reasons on the pattern, and produces an idea **that was in no single agent's CMB.** This is emergence from field collision — the mesh saw what none of them could see alone.

**6. Validator challenges again** — *"Produce a falsifiable prediction or downgrade from breakthrough to speculation."*

> **What the mesh does:** The validator's SVAF accepted the synthesis CMB (issue="novel framing" scores high on its `issue` weight). But its response sets a bar: commitment="accept if and only if a concrete experiment is proposed." Every agent receives this high-commitment signal. The synthesis agent must now respond with a testable prediction — or its idea dies in the DAG without descendants.

### The DAG is the research

```
explorer-a (scaling law claims) explorer-b (metric methodology)
explorer-a (claims) explorer-b (methodology)
\ /
└─── data-agent (metric-conditional breakpoint) ───┐
| │
validator (methodology challenge) │
| │
research-pm (reprioritise) │
| │
synthesis (emergent idea) ────────────────┘
└─── data-agent (evidence, 2 parents) ───┐
| │
validator (challenge) │
| │
research-pm (reprioritise) │
| │
synthesis (emergent idea) ────────┘
|
validator (demands falsifiable prediction)
validator (demands experiment)
```

**No single agent produced this.** The breakthrough came from the collision of intent and motivation fields across agents with different perspectives. The DAG traces every claim to its evidence, every challenge to its basis, every idea to the signals that produced it. The graph IS the research — traceable, immutable, auditable.
Every node traces back to its evidence. Every challenge links to the claim it disputes. Every idea connects to the signals that produced it. **The graph IS the research** — traceable, immutable, auditable. If a regulator asks "why did you conclude emergence is evaluation-dependent?", the lineage chain answers: because these three agents' intent and motivation fields converged, traced back to these two contradictory papers.

This is [Mesh Cognition](https://sym.bot/research/mesh-cognition). Read the [MMP specification](https://sym.bot/spec/mmp) for the full protocol.
> **Verified in production.** This pattern runs today with real agents: a knowledge explorer (Linux), a researcher (Claude Code, macOS), and MeloTune (iPhone) — three platforms, one mesh, coupled via relay with E2E encryption. The daemon shared question CMBs to the knowledge feed via anchor sync on connection. SVAF accepted at drift 0.068. MeloTune received the xMesh insight via APNs wake push. [Full protocol specification →](https://sym.bot/spec/mmp)

## More Use Cases

Expand Down
Loading