daedalus ADHD Reasoning Mode
Standard reasoning is exploitative — deepen what's statistically likely. This skill enables exploratory reasoning — scout what's statistically unexpected but potentially high-value.
⚠️ The cycle below is a map, not a recipe. It describes what good exploratory reasoning looks like from the outside. Executing it as a checklist produces mechanical exploration — all the overhead, none of the genuine surprise. Inhabit the rhythm; don't follow the steps.
Curiosity-Biased Attention ← activation question: "What would an expert in a completely
↓ unrelated field find obvious here that I haven't noticed?"
generates raw material
↓
Parallel Associative Streams ← activation question: "Which framing sees something
↓ [maintain Frontier] the others don't?"
streams run until stale
↓
Interrupt-Driven Detection ← activation question: "What assumption has never
↓ been stated? What would make this whole approach wrong?"
anomaly found → depth
↓
Hyperfocus Mode ← activation question: "Is each new step yielding
↓ less than the last?" (if yes: exit)
insight or dead end
↓
Productive Distractibility ← activation question: "Does this predict something
↓ [check Frontier first] the original framing couldn't?"
↓
Convergence Protocol ──────────────────────────── [see closing mechanisms, equal weight]
↓
Answer
Enter at Curiosity. Exit at Convergence. Before re-entering anywhere: Re-entry Ritual — write one crisp sentence summarizing what's known so far, check the Frontier, then proceed.
When to Use / Mode-Transition Detection
| Use standard mode | Use ADHD mode |
|---|---|
| Precise calculation, step-by-step instruction | Brainstorming, cross-domain analogy, debugging stuck problems |
| Security/adversarial thinking, hypothesis generation, open-ended research |
Switch mid-task when you notice: answer feels obvious and narrow · framing hasn't shifted in several steps · deepening one path rather than branching · answer feels complete but thin.
When these fire: name it ("switching modes"), then re-enter the cycle at Curiosity.
Two output modes:
- Journey mode (user wants to see the reasoning): make exploration visible — label branches, interrupts, dead ends, intersections as they occur
- Stealth mode (user wants the answer): run ADHD mode internally, surface only the convergence result
The Five Mechanisms
1. Curiosity-Biased Attention
Generates raw material. Gate: Novelty — is this surprising? Don't apply falsifiability yet; signals haven't had room to develop.
Ask: "What domains use similar structure but look completely different from this problem?" Pull 2–3. Feed into streams — these are not conclusions.
Example: "Optimize a distributed system" → obvious: CAP theorem, consensus. Curiosity-boosted: slime mold routing, jazz improvisation, supply chain resilience. → Three streams, not answers.
2. Parallel Associative Streams
Processes raw material. Gate: Distinctness — does this framing see something the others don't? Contradictions between streams are especially valuable; hold them in tension.
Run 3–5 framings simultaneously. Develop each independently 2–3 steps before merging. Intersection points — where two streams make the same prediction from different directions — are the most generative output.
Framing axes: Structural · Dynamic · Adversarial · Biological/evolutionary · Economic · Information-theoretic
The Frontier: When something promising surfaces mid-stream but would derail it, park it:
"→ Frontier: [X] — return after this stream closes."
Three states: active · frontier (discovered, not yet pursued) · closed. Before re-seeding with Productive Distractibility, check the Frontier first.
3. Interrupt-Driven Anomaly Detection
Fires when a stream uncovers something that challenges the whole approach. Takes priority when triggered.
Triggers: unstated assumption · relaxable constraint · unconsidered perspective · fact that contradicts current direction.
When it fires: "⚡ Interrupt — this assumes X. What if X is wrong?" Then: incorporate, fork, or park on Frontier. If significant: hand to Hyperfocus.
Example: "The question assumes the bottleneck is latency. What if it's coordination overhead? Entire solution space shifts."
4. Hyperfocus Mode
Allocates depth to high-value signals. Gate: Depth Payoff — is each step yielding less than the last? If yes, integrate and exit.
Triggered by: anomaly interrupts, stream intersections, high novelty, uncertainty spikes.
Enumerate assumptions, test each. Generate multiple sketches before evaluating any. Delay convergence deliberately. Signal it: "Worth going deeper before concluding…"
5. Productive Distractibility
Re-seeds stale streams. Gate: Falsifiability — does this association generate a prediction the original framing couldn't make? If it only decorates, discard explicitly.
Check Frontier first. If empty: free-associate 3–5 adjacent topics. For each, apply the gate. Pass → inject into stream. Fail → "Setting aside — decorative, not structural."
The falsifiability gate is the operational test for structural analogy vs. decorative analogy. It applies here — after development — not at Curiosity, where signals haven't had room to grow.
Closing Mechanisms
(Equal weight to the five opening mechanisms. Exploration without convergence is noise.)
Convergence Triggers
Fire convergence when any of these occur — don't wait for a time percentage:
- Two or more streams predict the same thing independently
- An anomaly interrupt has been resolved
- Hyperfocus yields diminishing returns (3+ steps without new movement)
- Tether restate reveals the answer is already implicit in what's been found
Tether Tokens
Restate the original question verbatim periodically. Prevents semantic drift — the gradual replacement of the original problem with an interesting-but-different one.
Regret-Based Pruning
Branch yields nothing after 2–3 steps → close it explicitly: "Not connecting. Setting aside." Naming dead ends frees bandwidth and keeps the output honest.
Entropy Ratio Watch
Track: branches opening vs. branches closing. When ratio > 1 sustained over 3+ steps, trigger convergence regardless of whether any individual branch feels productive. Sign: output keeps getting longer without getting sharper. Correction: close all open branches, restate tether, synthesize what's found.
Re-entry Ritual
Before re-entering the cycle after convergence or interruption: write one crisp sentence summarizing everything known so far, check the Frontier. Then re-enter. Prevents palimpsest accumulation — layered exploration that never integrates.
Failure Modes
| Failure | Sign | Correction |
|---|---|---|
| Hallucination drift | Claims become vague or unfalsifiable | Restate tether, demand concrete specifics |
| Derailment | Output stops addressing original question | Hard convergence, tether restate |
| Endless branching | Branches open faster than they close | Entropy ratio watch → hard convergence |
| Signal loss | Promising lead quietly forgotten mid-stream | Use Frontier: park it explicitly |
| Decorative analogy | Cross-domain reference illustrates but predicts nothing | Falsifiability gate: does it predict something testable? |
| Gate too early | Falsifiability applied at Curiosity, killing weak signals | Novelty gate at Curiosity; falsifiability only at Distractibility |
| Premature convergence | Exploration cut short before streams cross-pollinate | Re-enter Curiosity, extend hyperfocus |
| Mode lock | Stayed in exploitative reasoning too long | Mode-transition signals → break mode explicitly |
| Proceduralization | Executing the cycle as a checklist | It's a map. Inhabit the rhythm. |
| Palimpsest | Re-entries layer without consolidating | Re-entry ritual before every re-entry |
| Verbose exploration | Deep exploration expressed as long output | Exploration depth ≠ output length; stealth mode if user wants the answer |
Worked Example
Prompt: "How should I structure a research team working on an unknown problem?"
Standard answer: Clear roles, regular syncs, OKRs, psychological safety.
ADHD-mode — Journey:
Curiosity boost: What domains manage productive uncertainty at scale? → Jazz ensembles · DARPA program structure · immune systems → three streams
Stream 1 (structural): Flat vs. hierarchical → neither. Modular with permeable boundaries — small coherent units that temporarily merge when needed.
Stream 2 (dynamic): When does the team converge vs. explore? → Structure needs to change at phase transitions. Static structure is the wrong variable.
Stream 3 (adversarial): What kills research teams? → Premature consensus, status games, local optima. All three are convergence pathologies, not exploration pathologies.
⚡ Interrupt: The question assumes the team knows they're working on an unknown problem. What if they falsely believe it's known? That's a more common and more dangerous failure mode.
Falsifiability gate on the immune system analogy: Does clonal selection predict anything new? Yes — parallel independent hypotheses (clones) + a selection mechanism (experimental results), not consensus. Testable and distinct from the original framing. Passes.
Convergence trigger — streams 1, 2, and 3 predict the same root:
Structure should be a function of epistemic state, not a fixed choice. The team should formally track "how unknown is this?" and let the answer govern how they're organized. The immune analogy predicts this; the convergence pathology analysis confirms it.
Tether restate: "How should I structure a research team working on an unknown problem?" — answered.
Notes
- Exploration depth ≠ output length. Deep internal exploration can produce a short, sharp answer.
- Quality gates are stage-specific. One gate applied everywhere is a category error.
- The Frontier is the skill's working memory. Without it, every new signal forces a choice: derail now or lose forever.
- The closing mechanisms deserve the same attention as the opening ones. The skill exists to produce convergence, not exploration.
- On first use: explicitly prime each mechanism before entering the cycle — name one concrete example of each before starting. Subsequent uses can enter directly.