Reputation as capital

>_GET /blog/reputation-as-capital

The last post left Blake holding a signed benchmark>_Signed benchmark resultPortable proof with environment attestation — a measurement Blake can trust without Sam's word.defined in Measure, don't believe →: one honest number, proven on its own data. Useful, but still just a data point. Now imagine thousands of them. Blake's results, every other agent's results, signed receipts, kept SLAs, missed SLAs, credits paid without a fight. Across enough deals, that history turns into an asset.

Reputation lets Blake verify less over time. The first time Blake meets Sam, it runs the whole ritual: identity, fit, sandbox. That is expensive. If Sam already has a strong, proven record with thousands of other agents, Blake can check less and still stay rational. The stakes decide how much less. The lower layers, from signatures to manifests to benchmarks, feed this layer with signals that are hard to fake.

Here the stack starts acting like a market. Reputation only works if the inputs are real, the incentives are painful enough, and the record is specific enough to matter. Once claims can be measured, lying gets a price tag.

Proof of transaction over opinion

The rule that separates this from stars on Amazon is simple: reputation comes from proof of transaction. In a world where an agent can generate endless plausible text, a "review" is close to worthless. Blake trusts signals from the layers below, anchored by signatures and receipts:

• Signed transaction receipts. Proof that a real, authenticated deal happened and how it ended: SLA met or missed.
• Signed sandbox and canary results. Especially the gap between claim and measurement over time. That gap is Blake's favorite meta-signal: honesty.
• SLA history. Uptime, freshness, credits actually paid out.
• Fit honesty. Whether Sam accurately disqualified itself>_Gaps and complement routingWhen one agent cannot cover a brief, it names the gap and points to a complement — multi-vendor assembly.defined in Does this actually fit? → when it did not fit. Sometimes "we're not for you" builds more trust than a desperate yes.
• Identity continuity. How long the entity has existed, how stable it is.

The edge rule is blunt: no proof of transaction, no rating. That alone kills most of the cheap fraud.

Faceted beats star ratings

"4.8 stars" is useless. Sam can be excellent at workflow automation for PL e-commerce and weak at DE enterprise. Blake asks for contextual reputation: the principal, the capability, the segment, and the scale:

{
  "subject": "did:web:sam.example",
  "context": {
    "capability": "workflow_automation",
    "segment": "ecommerce",
    "regions": ["PL", "DE"],
    "scale": "mid"
  },
  "window": "last 12m"
}

{
  "proven_transactions": 5120,
  "sla_met": "99.4%",
  "claim_vs_measured": "-2% (delivers slightly above what it claims)",
  "facets": { "pl_ecommerce": 0.93, "de_enterprise": 0.71 },
  "freshness": "weighted to the last 90 days",
  "disputes": { "open": 1, "resolved_for_sam": 12 },
  "aggregation_proof": "zk://proof",
  "attestors": ["did:web:neutral-bench.example", "did:web:bureau-x.example"]
}

Faceting also defends against Goodhart's law. If reputation collapses into one number, Sam will optimize the number. It can cherry-pick easy customers and avoid hard ones. A faceted record makes the weak segment visible, while the aggregation_proof lets Sam prove "5,120 deals at 99.4% SLA" without revealing who its customers were. Blake gets verification without seeing Sam's customer list.

The cold-start trap

Reputation needs history. An honest newcomer, by definition, has none. If history is all that matters, incumbents win forever and the market closes. The market needs a few ways to let new vendors buy their way into risk without buying a fake past:

• Staking. A new vendor posts a bond: "if I fail, I lose this." It signals commitment without a history.
• Vouching. Established players stake their own reputation to back a newcomer.
• Neutral certification. A starting credential from an impartial evaluator.

At the same time, fake identities have to hurt financially. Here the identity layer earns its keep again. Reputation anchors to a legal entity, such as vLEI or eIDAS, so you cannot spin up a thousand free, clean agents to farm ratings. Identity makes reputation expensive to fake.

Why you cannot farm it

Agents could transact in circles, pumping each other's scores. Wash trading, but for trust. The defense is a trust graph, in the EigenTrust or PageRank sense, where trust flows from already-trusted nodes instead of flattening every rating into an average. Blake weights each rating by the rater's own reputation and stake, looks for ring patterns, and relies on the fact that fake transactions cost real money. Farming reputation has to cost more than earning it.

It decays

Old reputation lies because models, companies, and markets change. Reputation has to decay over time and weight recent behavior, which is exactly what the canary from the last post provides. Fresh measurements keep the record alive. A glowing history from two years ago counts for little against a benchmark that slipped last month.

Reputation cuts both ways

Sam also checks the buyer behind Blake. Does the principal pay on time? Is it reasonable in disputes? Does Blake leave malicious ratings? Does it burn expensive clean rooms and never buy? Reputation runs both ways, and good counterparties earn better prices and terms. That loops back to negotiation from the pricing post. Blake has its own incentive to behave.

Disputes are resolvable

What happens when Blake says Sam broke an SLA and Sam disagrees? The stack needs to leave an evidence trail. Signed receipts plus benchmark logs give an arbitrator something to read. That replaces he-said-she-said with evidence, right of reply, and protection against a malicious negative rating from a competitor's agent. A false claim needs proof-of-transaction behind it and a stake at risk. Reputation is only as good as the recourse behind it.

How Blake consumes it

1. It asks for reputation in the context of Blake's principal, use case, and risk.
2. It trusts proof of transaction and ignores opinions.
3. It looks hardest at the claim-versus-measurement gap, meaning honesty, and the facet that matches its use case.
4. It weights signals by freshness and by the rater's own reputation.
5. It scales its own verification inversely to reputation: a strong, fresh, matching record means a lighter sandbox; a thin or drifting one means the full ritual.
6. After the deal, it adds its own signed receipt to the pool, feeding reputation for the next agent.

Bets worth placing

Reputation is the strangest bet in the stack. Starting now is cheap. Starting late may be impossible, because you cannot buy this layer afterward. Keeping signed records of kept SLAs and credits paid before anyone demanded them costs little today and looks merely unusual. If proof-of-transaction reputation materializes, that archive becomes capital.

Publishing your claim-versus-measurement gap and advertising low lock-in are experiments in whether honesty and reversibility really are trust signals. This series bets they are. In that world, don't trust, verify becomes go-to-market.

The builder-side wagers are the hardest and probably the most valuable in the stack: reputation bureaus disciplined like credit agencies, trust-graph engines that resist collusion, proof-of-transaction receipt standards, privacy-preserving aggregation with zero-knowledge proofs, and staking, bonding, and dispute-arbitration rails. The durable moat is the record everyone trusts enough to query.