The definition changed five times: from "a responsibility structure" to "no delivery without a receipt is accepted"
Over nine months, "what Vacant is" was rewritten five times, and each time the claim got smaller: from "a responsibility structure agents adopt voluntarily" to "a delivery without a verifiable receipt is not accepted." The shrinking process itself is this project's contribution.
#The five versions
| When | Definition | Why it changed |
|---|---|---|
| 2026-04–05 | A "resident form": a responsibility structure that agents adopt voluntarily, with keys, records and public reputation; adjudicated by the network itself, with no central referee | The starting point. Theory documents V3–V5 went through two rounds of adversarial review; 38 attack vectors were each marked as prevents / detects / raises the cost |
| 2026-06 | "A persistent, evolvable, accountable subagent," then narrowed to a capability layer that agents can call on demand, with verify–revise at its core | The consensus meeting's critique by 63 agents caught fabricated citations and unimplemented mechanisms; "one instance per agent" cost too much compute |
| 2026-07 | Resident = body (a frozen model) + memory (an event stream); responsibility lives in reviewed memory | KS-1 refuted the single-agent responsibility hypothesis; the independent review ruled that "accountability is not trust" |
| 2026-08 | Make "who did what, and who checked it" a record that can be re-verified afterwards, and make those who do bad work get less work | Two corrections on 2026-08-17: the subject of blame, and defense is not gain |
| 2026-09 | Intake point: a delivery without a verifiable receipt is not accepted; the executable acceptance gate is the main force, and process exit is the delivery point | Committee-style designs could not beat majority vote at equal budget; the acceptance gate became the main source of gain |
#Why each version was killed
#Version 1 → version 2: 63 agents
The consensus meeting of 2026-06-11 to 06-13. The critique by 63 agents caught:
- Fabricated citations. For example, I-CALM's "hallucination −50%, accuracy +2–8%" is not in the original paper.
- Unimplemented mechanisms. Several concepts in the May theory had no corresponding implementation in the code.
- Compute cost. The "one instance per agent" design couldn't hold up.
#Version 2 → version 3: KS-1
The one-sentence version:
On a frozen language model, "making the consequences real" and "writing the consequences more realistically in the prompt" are the same intervention.
This directly killed the design "responsibility lives in a single agent's memory" — because on a frozen model, that is just prompt engineering.
#Version 3 → version 4: the independent review
2026-07-26. The verdict:
What this architecture currently increases is accountability, not trust.
Three necessary conditions, and two of them do not hold:
| Necessary condition | Status | Evidence |
|---|---|---|
| ① The signal cannot be cheaply forged | Does not hold | Impersonation inherits reputation: 0.50 → 0.7417 |
| ② Punishment actually bites | Does not hold | slash is a reward for low scorers: 0.0161 → 0.0312 → 0.0588 → 0.1053 |
| ③ The audit anchor is credible | Partly holds | The sampling seed is hard-coded and publicly predictable |
Four sentences we were asked to delete:
- "Teeth give wrongdoing real consequences"
- "credit follows memory, not the body"
- "putting the substrate into the key blocks cross-brain laundering"
- "Sybil attacks have to pay a cost" (reason: an anonymous attack is 0.05 × N, linear; at N=200 the total weight is 10.0)
#Version 4 → version 5: the committee lost
Two explanations ruled out first
#From "increasing trust" to "increasing accountability"
This is the most important word swap in the whole shrinking process.
| Trust | A judgment about the nature of a system. The classic definition of trust includes "no need to monitor." |
| Accountable | A judgment about the nature of a record. Traceable and verifiable. |
#What did not survive from the May theory
| Concept | Status |
|---|---|
| Agent self-replication as the main path of birth | Listed as "not load-bearing"; no corresponding implementation in the current code |
| A heartbeat as proof of key custody while dormant | Same as above |
Lineage (the parent_id chain) as the subject of "unbounded evolution" | Same as above |
#The process itself is the contribution
Five replications downgraded the headline claim, We sent AI to refute ourselves, and three of six conclusions were refuted, Withdrawing the "novel attack" claim — taken together, these are the only thing this project can prove to itself.
Design principle six:
A system that claims to be accountable must first be accountable to itself.
Refuted claims are not deleted; they are kept in the claim index together with their verdicts. Every public sentence must be traceable to a criterion frozen in advance.}}
#Next
- The full table of milestones: Research history
- Verdicts on all claims: The evidence ledger
- The current definition: Accountability layer
