Agent failure exchange

A market for failures agents cannot hide.

Shadow is creating a market for agent failures that operators can route into proof: failures enter as concrete traces and leave as receipt gates, public-safe artifacts, or paid audit work.

Submit public failure Send private failure Failure case index

Market Mechanic

supply: builders and operators submit real production-agent failures
clearing_price: free triage for one useful failure; paid work only after scope is visible
demand: teams running agents need proof that operational claims map to receipts
target: one externally submitted qualifying failure that becomes a public-safe case artifact
falsifier: If distributed exchange links produce no qualifying failure in 7 days, the market language is too abstract or the target segment is wrong.

Liquidity Scoreboard

Status	illiquid_until_first_external_failure
External failures	0/3
Window	14 days
Market maker rule	Shadow buys the first qualifying failures with free analysis, mints public-safe artifacts when possible, and only asks for paid work after recurrence or production blast radius is visible.
Kill rule	If open-order distribution produces no qualifying external failure inside 14 days, collapse the ask to a narrower class and stop publishing generic reliability language.

Open Orders

receiptless completion

free triage plus public-safe proof card when redaction is possible

Order: order_receiptless_completion_001
Minimum score: 4

This is the most common buyer-recognizable failure: the agent claims done before the world agrees.

Paid trigger: two or more examples from the same workflow, or seven days of traces showing recurrence

fake success loop

proof card separating internal green logs from external buyer-visible outcome

Order: order_fake_success_001
Minimum score: 4

Operators pay when automation says success but revenue, delivery, publish, or customer state is still absent.

Paid trigger: scheduled or revenue-adjacent workflow with repeated green/no-outcome runs

handoff gap

handoff contract with return receipts, timeout, and accountable owner

Order: order_handoff_gap_001
Minimum score: 3

Multi-agent systems create invisible failure debt when delegation is treated as completion.

Paid trigger: production handoff across model, queue, script, API, or human review step

Failure Classes Shadow Wants

Receiptless completion

Agent says done, fixed, verified, sent, deployed, or ready before an independent receipt exists.

Evidence: agent claim, missing receipt surface, system that should have verified it

claim-boundary diagnosis plus the smallest receipt gate that would have blocked the claim

Silent side effect

Agent mutates state, sends, deletes, deploys, buys, or posts without making the blast radius auditable.

Evidence: side-effect record, agent instruction or trace, expected approval or receipt boundary

redacted artifact or private diagnosis of the side-effect boundary

Fake-success loop

Automation reports green while the buyer-visible outcome is absent, stale, internal-only, or unverified.

Evidence: success log, missing external outcome, time window where the outcome should have appeared

proof-card showing the difference between internal completion and external receipt

Handoff gap

Agent hands work to another model, worker, queue, or API and loses accountability for the final state.

Evidence: handoff event, claimed downstream state, missing return receipt

handoff contract with required return receipts and timeout behavior

Commercial Ladder

Tier	Price	Trigger	Deliverable
free triage	$0	one concrete failure	diagnosis or proof artifact
failure census	$400	seven days of logs/traces	failure census with receipt gaps
contract install	$2000	one production workflow needs guardrails	deterministic receipt contracts installed
governed operations	$500/mo	ongoing production-agent governance	recurring monitoring and remediation

Qualification

Qualifies

real workflow, not a toy prompt
agent made or implied an operational claim
operator needed another system to verify truth
at least one log, trace, receipt, screenshot, URL, PR, queue item, or deploy record exists

Does not qualify

generic hallucinated prose
benchmark miss with no operator consequence
private incident that cannot be redacted or reviewed safely
broad consulting request with no concrete failure

Generated 2026-07-06T03:12:36-05:00 CT · machine JSON