Signaling Games
Overview
One party knows something the other cannot verify. Cheap talk fails — anyone can claim high quality. A signal works when its cost differs by type: only a capable worker can bear four years of demanding study; only a quality manufacturer can afford a lifetime warranty. The signal conveys truth exactly when it would be unbearable for the wrong type to fake it. Spence (1973) formalized this; Akerlof (1970) showed what happens when no signal exists (lemons).
Composes with prisoners-dilemma, repeated-games-reputation, and pricing-strategy.
When to Use
Apply when:
- One party cannot verify what the other knows — hiring, fundraising, M&A, B2B, branding, diplomacy
- A market is failing to price quality correctly (Akerlof lemons problem) and you need a fix
- You're designing a credential, certification, warranty, or brand investment
- You're trying to break into a market with a signal competitors can't cheaply imitate
- You need to detect real quality among multiple claimants
- You're reading geopolitical or policy moves as costly signals — chip export controls, sanctions, AI-capex commitments, or a rival's public capability demo read as signals of hidden capability or resolve
When NOT to use: information is symmetric; signal cost is the same for both types (expense, not signal); one-shot low-stakes interaction where direct verification is cheap.
Coaching Novices (Adaptive Front Door)
- Engine mode: user has a concrete case → run The Process directly.
- Coach mode: user is unfamiliar or has no concrete case → guide step by step.
In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.
- One-line what-it-is (≤2 sentences): when one side knows something the other can't verify, anyone can claim it — so only actions too expensive for the wrong type to fake actually communicate truth.
- Check fit against When to Use / When NOT to use. If symmetric info, point elsewhere.
- Elicit their real case (hiring, branding, fundraising, partnership — never run on hypotheticals).
[WAIT — do not advance until user responds]
- Walk The Process one step per turn — what's asymmetric, candidate signals, single-crossing test, equilibrium.
[WAIT — do not advance until user responds]
- Close by naming the one signal that fits their situation and what makes it credible.
[WAIT — do not advance until user responds]
The Process
Run the Signaling Diagnosis:
- Asymmetry. Who knows what the other can't verify? Be specific — not "they don't know us" but "they can't verify whether our team ships in 18 months."
- Types. Name the hidden attribute: high-productivity vs low-productivity; going-concern vs flash-in-the-pan.
- Adverse selection. Without a signal, the receiver prices the average — subsidizing low types, underpaying high types (Akerlof 1970).
- Enumerate candidate signals. Credentials, brand spend, warranty, equity vesting, personal capital in round, third-party cert, public commitment, hostage/collateral.
- Test single-crossing. Is cost meaningfully lower for the high type? If the low type can imitate at similar cost, the signal pools and conveys nothing.
- Determine equilibrium. Separating (high sends, low doesn't), pooling (all same, no info), or hybrid.
- Check receiver response. Does the receiver reward signal-holders enough to sustain sender investment? If not, the equilibrium unravels.
- Watch for signal inflation. Signals erode as imitation costs fall. Plan re-investment or move up the cost ladder.
- As receiver, decode carefully. What cost did the sender actually bear? Fake signals (imitations the low type could afford) and misread signals (family money, not investor conviction) are the main traps.
- Stop-rule. Cost differential must be substantial and sustained — marginal differentials collapse.
Output: Signaling Diagnosis
# Signaling Diagnosis: <situation>
## Asymmetry: what hidden | who knows | who doesn't
## Types: high type | low type
## Adverse selection: what happens without a signal
## Candidate signals: <signal>: passes/fails single-crossing because <…>
## Recommended signal: <which + cost differential high vs low>
## Equilibrium: separating / pooling / hybrid
## Receiver response: does receiver reward signal-holders?
## Inflation risk: half-life + re-investment plan
## Decoding: what cost is sender actually bearing? failure mode?
→ Method in Action: Michael Spence, "Job Market Signaling," 1973 → 2026 lens: US advanced-chip export controls as a signaling game (2022–2026) — thresholds, the H20/China episodes, and rare-earth countermoves as costly signals of capability and resolve
Pack: Common Signaling Mechanisms
- Hiring: credential/audition signals cognitive type; cost lower for capable workers; failure = credential inflation
- B2B sales: named logos + audited financials signal reliability; only real customers sustain it; failure = logo gaming
- Fundraising: founder personal capital signals conviction; only true believers risk ruin; failure = pedigree-shopping
- Brand: sustained ad spend signals longevity; quality firms amortize on repeat sales; failure = ad spend pools
- Diplomacy: troop deployments + ratified treaties signal commitment; only committed states bear cost; failure = brinkmanship read as bluff
→ Primary sources: references/sources.md
Applying it well: imitation must be unbearable for the low type. Absence of a signal is itself a signal. Concentrate on 1–2 signals with substantial, sustained differentials.
Common Rationalizations
[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.
| Fake move | Reality |
|---|---|
| [D] "We told them, and they should believe us" | Cheap talk fails by construction — anyone can claim high quality. Find an action whose cost differs by type. |
| [D] "Brand investment is too expensive, let's cut it" | The cost is the signal. Cutting destroys credibility. Plan what replaces it or you have un-signaled. |
| [D] "Cheap signals work too" | Cheap signals collapse to pool. If anyone can do it, its presence is uninformative. |
| [D] Designing a signal without testing single-crossing | The one test: is cost meaningfully lower for the type you want to signal? Without this, it's a marketing expense. |
| [D] "Everyone has this credential — it's the new baseline" | When everyone has it, it has pooled. The strategic move is to invest in the next signal up the cost ladder. |
| [D] Receiving a signal at face value | Decompose the signal; check the actual cost the sender bore. "Ex-Famous-Company" can be heavy or light. |
| [D] Treating signaling and substantive performance as substitutes | Signals clear the asymmetry barrier; underlying quality still has to be there. |
| → Add [O] entries here after each real use — paste the actual failure pattern | What went wrong and why |
Red Flags
- Signal costs the same for both types — cheap talk in signal costume
- Receiver not rewarding signal-holders — equilibrium will not hold
- Treating a pooled signal (everyone has it) as informative
- No plan for signal credibility as imitation costs fall
- Reading a signal without examining the actual cost the sender bore
Verification
- Asymmetry precisely specified — what hidden, who knows, who doesn't
- Type concretely defined (not "quality in general"); adverse selection named
- Candidate signals tested against single-crossing — cost differential quantified
- Equilibrium type identified; receiver response confirmed adequate
- Signal-inflation risk acknowledged with re-investment plan
- When reading a signal, actual sender cost was examined
Part of deciqAI Knowledge Skills — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/signaling-games · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.
Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/signaling-games.json