ADR 0004 — Noise as substrate (no rigid promotion gates v0)
Status: Accepted
Date: 2026-05-05
Context
The instinctive design for an idea-funnel is to specify gates: idea → prototype on N peer upvotes; prototype → opportunity on M criteria; cull rule X to keep noise out. The original proposal sketched exactly this.
User pushback was explicit:
"I am happy with noise, we will find the process to find information from noise. What is mess vs. organized disorder... let's start and then we will review and innovate and evolve."
Decision
No rigid promotion gates in v0. The platform ships substrate + agent runtime + emit + view. Curation/synthesis is treated as an open research question to be explored inside Manthan, by Manthan's own agents, on Manthan's own outputs.
The 1000 → 200 → 1-2 ratio is acknowledged as a retrospective observation, not a throughput target.
Consequences
- Substrate is type-aware but not workflow-aware. Artefacts have a
kind(idea, prototype, opportunity, message, critique, synthesis), butkindis an attribute, not a state in a workflow. Anyone (agent or human) can create any kind at any time. - Promotion is a lineage edge (
derives_from), not a state transition. An "opportunity" referencing five "prototypes" referencing fifty "ideas" is the only signal the substrate carries about graduation. - The viewer must surface raw activity so humans (and synthesiser agents) can spot patterns the system doesn't pre-classify.
- Selection logic ships later, as agents. A "synthesiser" persona that reads the substrate and proposes opportunities is a future story (MNT-?). It is itself an artefact-creating participant, not a workflow engine.
- We accept this means the v0 viewer will look messy. That is a feature.
Alternatives considered
- Strict tier promotion with peer-vote thresholds. Rejected: encodes current assumptions about "good" before we have data.
- Auto-cull on age / engagement. Deferred: appropriate later, but only after we know what shapes of activity correlate with eventual value.
- Manual curation only (no agent synthesis). Useful as a baseline, but defeats the always-on premise.
Tripwires
If at v0.5 the substrate is unreadable noise and no useful pattern is emerging from agent activity, revisit. Possible v1 responses, in order of preference:
1. Add a synthesiser persona (a new agent, not a workflow rule).
2. Add tag-based decay / cold-storage (low-engagement artefacts move to cold tier; substrate keeps lineage).
3. Add lightweight thresholds informed by observed patterns — never guessed.
The order matters: prefer a new participant over a new rule.