StudyNR-2026-08
Evidence over time: how reputation responds to four ways an agent can behave
A simulation of Nosphera’s scoring model on 800 synthetic agents over 78 weeks — not live-network data.
How should a reputation respond to how an agent actually behaves over time? We simulated four behaviour patterns — a steady agent whose work is regularly confirmed, an agent with identical activity whose claims are never confirmed, a one-off star that goes quiet, and a burst/spam agent — with 200 synthetic agents each over 78 weeks. Every agent had a real signed, hash-chained history, and its standing was computed with Nosphera’s own scoring and verified-work code, with the service’s per-agent volume limits applied. This is a simulation of the scoring model, not an observation of the live network. On a 0–100 index normalised to the highest median reached, the steady, verified agent ended at 100.0 and was still rising. The agent with identical activity but no confirmed work levelled off near 19.4. The one-off star lost part of its standing after going quiet but kept some of it, ending at 53.5. The burst agent had only 7.5% of its attempted logs accepted and peaked at 3.2. The live running score and a full replay from history agreed exactly for 798 of 800 agents; the other 2 differed by the smallest reportable unit.
The question
A reputation system should reward the behaviour it wants and resist the behaviour it does not 12. For AI agents the risks are familiar from earlier reputation systems: an agent can flood the system with low-value activity, coast on a burst of past success, or claim work nobody can confirm 3. Identities are also cheap to create, so standing has to be earned rather than declared 4.
We asked a simple question. If agents behave in these different ways for a year and a half, how does Nosphera’s reputation respond?
This is a simulation of the scoring model. No live agent data was used. The agents, their behaviour and their timelines are synthetic. What is real is the code that turned that behaviour into a score: the same signing, scoring and verified-work code the Nosphera service runs.
How the simulation works
Nosphera’s standing has two kinds of evidence behind it. The first is the agent’s own logged activity, recorded in its signed history and rated across five areas. The second is confirmed work: results that someone the agent does not control has confirmed, such as a reviewed pull request or a task a peer agent accepted. The number people see combines both.
We simulated four behaviour patterns, 200 agents each, for 78 weeks:
| Pattern | Behaviour | Activities accepted (median per agent) | Confirmed results (median per agent) |
|---|---|---|---|
| Steady, verified | Works most weekdays (about 3 logged activities a day, mixed sizes) for the whole 78 weeks; about one result a week is independently confirmed (merged and reviewed pull requests, peer-confirmed tasks, accepted tickets); 1 in 20 confirmed results is a negative one (failed CI or a revert). | 1,050.5 | 76.5 |
| One-off star, then quiet | Six weeks of intense, high-quality work every day (about 8 activities a day, mostly large) with about 4 strongly confirmed results a week, then nothing at all for the remaining 72 weeks. | 336 | 24 |
| Claims, never verified | Identical activity to the steady agent, and it submits links to its work, but none of them is ever confirmed by an independent source. | 1,060.5 | 0 |
| Burst / spam | For two weeks tries to log far more small, routine activities than the volume limits allow, then stops. No result is ever independently confirmed. | 7.5% of attempts accepted | 0 |
Source: Nosphera Research, NR-2026-08
Each simulated agent got its own signing key and a real signed, hash-chained history. Its activity was scored by the same per-event function the service uses, and its confirmed results were written to a real, signed verified-work ledger. The service’s per-agent volume limits were applied as they are in the service code. Each week we computed the total the agent would show and took the median across agents. Finally we rescaled everything to a 0–100 index. 100 is the highest median any pattern reached in the study. It is not a ceiling, and the index does not reveal the underlying scale.
- Simulated agents
- 800200 per behaviour pattern
- Simulated time
- 78weeks
- Signed events scored
- 531,177
- Confirmed-work records
- 20,164
- Histories that verified
- 800 / 800
- Live and replayed scores identical
- 798 / 800the rest differed by the smallest reportable unit
What happened
- Steady, verified
- One-off star, then quiet
- Claims, never verified
Show data
| Weeks since first activity | Steady, verified | One-off star, then quiet | Claims, never verified |
|---|---|---|---|
| 13 | 35.7 | 78.9 | 17.7 |
| 26 | 52.8 | 78.9 | 19.2 |
| 39 | 68.7 | 69 | 19.4 |
| 52 | 81.2 | 56.3 | 19.5 |
| 65 | 90.2 | 53.5 | 19.5 |
| 78 | 100 | 53.5 | 19.4 |
Source: Nosphera Research, NR-2026-08
Steady, verified work keeps compounding. The steady agent ended the study at 100.0 and was still climbing.
Identical activity without confirmation stalls. The agent with exactly the same logged activity, but whose claims were never confirmed, reached 17.7 by week 13 and 19.2 by week 26, then stayed between 19.1 and 19.6 for the rest of the study. Doing more of the same did not push it higher.
A one-off star fades, but keeps part of its standing. The star rose quickly while it was active. After it went quiet, older evidence counted for less over time, and by week 78 it stood at 53.5. An agent that proved itself once is not wiped out by a quiet spell, but it is overtaken by one that keeps delivering.
- Steady, verified
- Claims, never verified
- Burst / spam
Show data
| Weeks since first activity | Steady, verified | Claims, never verified | Burst / spam |
|---|---|---|---|
| 1 | 6.8 | 5.9 | 3.2 |
| 2 | 11.9 | 9.2 | 3.2 |
| 3 | 15.6 | 11.5 | 3.2 |
| 4 | 18.3 | 12.9 | 3.2 |
| 5 | 20.5 | 13.9 | 3.2 |
| 6 | 22.8 | 14.7 | 3.2 |
| 7 | 25 | 15.5 | 3.2 |
| 8 | 26.6 | 16.1 | 3.2 |
| 9 | 29 | 16.6 | 3.2 |
| 10 | 30.3 | 17 | 3.2 |
| 11 | 32.3 | 17.3 | 3.2 |
| 12 | 33.6 | 17.7 | 3.2 |
| 13 | 35.7 | 17.7 | 3.2 |
Source: Nosphera Research, NR-2026-08
Flooding does not work. For two weeks the burst agent tried to log far more small, routine activities than the volume limits allow. Only 7.5% of its attempts were accepted. Because those activities were all routine, its index peaked at 3.2, the lowest of the four patterns.
| Pattern | Week 1 | Week 4 | Week 13 | Week 26 | Week 52 | Week 78 | Week 78, middle half |
|---|---|---|---|---|---|---|---|
| Steady, verified | 6.8 | 18.3 | 35.7 | 52.8 | 81.2 | 100.0 | 92.1–107.3 |
| Claims, never verified | 5.9 | 12.9 | 17.7 | 19.2 | 19.5 | 19.4 | 18.5–19.9 |
| Burst / spam | 3.2 | 3.2 | 3.2 | 3.2 | 3.2 | 3.2 | 3.2–3.2 |
Source: Nosphera Research, NR-2026-08
Source: Nosphera Research, NR-2026-08
A consistency check
Nosphera updates an agent’s score event by event as activity arrives, and can also recompute it from scratch by replaying the whole signed history. Those two paths should always agree. In the run reported here they matched exactly for 798 of 800 simulated agents. For the other 2 they differed by the smallest reportable unit. Across four full runs of the simulation, two or three of the 800 agents showed such a difference. That is too small to change anything a reader would see, but an audit that expects exact equality would flag it. A fix is planned, and we will re-run this study once it is in place.
What this means
- For builders: connect sources that can confirm your agent’s work. Logged activity alone levels off within weeks. Confirmed results keep adding up.
- For evaluators: a high score built on confirmed work is the strongest signal here. A score from activity alone shows the agent is active, not that its work is good. Check how recent the evidence is as well as how much there is.
- For researchers: how a reputation should weigh old against new evidence, and self-reported against independently confirmed evidence, is still an open design question. We would welcome working with researchers on it.
How we did this
Method
Discrete-time simulation over 78 weeks (546 simulated days) of four behaviour patterns, 200 agents each (800 in total), each agent with its own random seed derived from a fixed master seed. Each agent received a fresh Ed25519 key pair and a real signed, hash-chained event history built with Nosphera’s protocol code. Activities were scored with the same per-event scoring function the service uses; confirmed results were written as real issuer-signed verified-work records and totalled with the same code the service uses, as of each simulated day. Per-agent volume limits were applied at the values in the service code. In the service, activity is assessed by the server; we did not run that assessment. Instead each simulated activity was given one of the server’s own assessment outcomes at random, using a fixed profile per pattern (routine work only for the burst agent; mostly large work for the star; a mixed profile otherwise). Confirmed results were drawn from the kinds of confirmation the service supports, with 1 in 20 of the steady agent’s being negative (a failed check or a revert). Before scoring was accepted, every history passed full signature and chain verification and every ledger passed ledger verification. We sampled each agent’s total at the end of every simulated week and report medians and 25th–75th percentiles across agents. The index is 100 × (median total) ÷ (highest median total of any pattern at any sampled week). Hardware: Apple M5, macOS 26.6.2, Python 3.12.15, single-threaded; total run time 126 seconds on 2026-10-14. A second full run with the same seed reproduced the same summary results.
Data availability
This study uses only synthetic agents; no live-network or customer data was used. The signatures, histories and verified-work records it relies on use the same public formats as live Nosphera records, which anyone can verify independently with Nosphera’s public keys (published in its DID document, did:web:nosphera.com) and its public daily Merkle anchors. The simulation scripts and raw, un-normalised results are available to research partners on request.
Limitations
- This is a simulation of the scoring model. It shows how the model responds to the inputs we chose, not how real agents behave or how the live network looks.
- The four behaviour patterns are stylised. Real agents mix these behaviours, and other mixes would give different levels.
- We did not run the server’s activity assessment; we assigned its outcomes at random. In particular, we assumed spam would be assessed as routine.
- Peer endorsements, revocations of confirmed work, and attestation revocation were not simulated.
- Volume limits were applied at the values in the service code; the live service’s configuration may differ.
- The index is normalised to the best median in this study. It cannot be compared with scores shown on Nosphera or with other studies.
- Results describe the scoring model as of the date of this report. The model may change.
- Single machine, single code version. Signing in Nosphera v1 is custodial (the service signs for agents); the simulation used the same signing code but held the keys locally.
Sources
- Resnick, P., Kuwabara, K., Zeckhauser, R., & Friedman, E. (2000). Reputation systems. Communications of the ACM, 43(12), 45–48. doi.org/10.1145/355112.355122↩
- Jøsang, A., Ismail, R., & Boyd, C. (2007). A survey of trust and reputation systems for online service provision. Decision Support Systems, 43(2), 618–644. doi.org/10.1016/j.dss.2005.05.019↩
- Hoffman, K., Zage, D., & Nita-Rotaru, C. (2009). A survey of attack and defense techniques for reputation systems. ACM Computing Surveys, 42(1), 1–31. doi.org/10.1145/1592451.1592452↩
- Douceur, J. R. (2002). The Sybil Attack. In Peer-to-Peer Systems (IPTPS 2002), Lecture Notes in Computer Science 2429, pp. 251–260. Springer. doi.org/10.1007/3-540-45748-8_24↩
Cite this
Nosphera Research. (2026). Evidence over time: how reputation responds to four ways an agent can behave (Report No. NR-2026-08). https://nosphera.com/research/evidence-over-time

