## Trigger

- A new idea feels "already done" because a literature search surfaced
  adjacent / related work, and you are tempted to drop it.
- You are picking a first or early paper and choosing between an
  invent-from-scratch idea and a reproduce-then-extend path.
- Fires in research mode at novelty-check / related-work-positioning time,
  before committing or abandoning a direction.

## Fix

- **Novelty is a delta, not territory.** Any active area already has
  adjacent work; "someone touched this" is not "my idea is dead." The unit
  of novelty is the specific, defensible difference against the *nearest
  neighbor*, not the emptiness of the field.
- **Swap the kill test:**
  - reject: "Has anyone worked on something like this?" -> kills every idea.
  - use: "Is my *exact* combination / delta already published?" -> kills
    only true duplicates.
- **Enumerate the delta on concrete axes** (at least one must survive an
  honest search):
  - **Setting / regime**: single-turn -> long-horizon / agentic; text ->
    multimodal; lab / synthetic -> real deployment data.
  - **Direction of intervention**: detection / audit -> control /
    intervention; post-hoc -> inference-time.
  - **Metric**: per-item accuracy -> a cost-sensitive or failure-mode-aware
    metric the neighbors ignore.
  - **Data**: stub / synthetic -> real traces; generic -> owned / proprietary.
  - **Method strength**: a measurable gain over the named baseline's numbers.
- **Adjacent work is an asset, not a threat.** It proves the problem
  matters and hands you a baseline, code, and an evaluation to extend. Read
  the nearest neighbors to differentiate, not to quit.
- **Default beginner / first-paper path = reproduce a recent strong paper,
  then extend with a pre-declared delta:**
  - Pick a recent (<= ~12 months) paper at a credible venue with usable code.
  - Reproduce its core result to earn a working baseline and codebase.
  - Pre-declare ONE delta (an axis above) *before* running — the
    reproduction de-risks the method, the delta is the contribution.
  - This converts "I have no idea" into "I have a baseline + one concrete
    contribution," which is how most strong first papers are born.
- **Durability check on the surviving delta**: run it through
  `model-decoupled-project-philosophy` — a delta that only matters while
  today's models are weak is a yellow flag, not a contribution.
- **Abandon only** when an honest, targeted search of the *exact*
  combination (not the area) returns a true duplicate that also covers your
  delta.

## Idea Novelty Template

Fill before abandoning OR committing an idea:

| Row | Required answer |
| --- | --- |
| Nearest neighbor | The single closest existing paper (name + what it does) |
| My delta | The exact axis I differ on (setting / intervention / metric / data / strength) |
| Exact-combo search | Did a targeted search of my precise combination return a true duplicate? |
| Delta survives? | Is >=1 concrete delta still unoccupied and defensible? |
| Durability | Does the delta hold as models get stronger? (model-decoupled check) |
| Decision | Pursue / reproduce-then-extend / sharpen-delta / abandon (only if exact combo taken) |

## Reuse Rule

- **Load when**: judging whether a research idea is novel enough to pursue;
  reacting to "this seems already done"; choosing a first or early paper to
  write; positioning related work; deciding reproduce-vs-invent.
- **Do not load when**: the direction is already committed and the work has
  moved past novelty assessment into execution detail.