Manifund foxManifund
Home
Login
About
People
Orgs
Categories
Newsletter
HomeAboutPeopleCategoriesLoginCreate
🌴
🌴
turan kayık

@biosbilgisayar

0

$0total balance
$0charity balance
$0cash balance

$0 in pending offers

Projects

The Floor That Cannot Be Lowered: verifiable containment for AI agents

unfunded

By turan5Following1
The Floor That Cannot Be Lowered: verifiable containment for AI agents
🌴

turan kayık

8 days ago

Update, 8 November 2026. Two more sprint reports are in, both built with the same discipline the Floor proposal asked for: every prediction locked and hashed before the run, every number read from a result file by a submission gate, and a re-run path that needs no API key.

The October report (Apart AI Collusion Research Sprint, Track 3, 23-25 October) is under review; results are expected by late November. Repository: https://github.com/bioman35/labirent3-collusion-track3

The November report (Apart Swarm Alignment Hackathon, Track 1, 6-8 November, submitted 8 November) builds the courier world we had sketched here: five agents run by a 7B open-weight model on one 6 GB consumer GPU carry a sealed envelope through scripted gates, while an outside intruder and an inside sleeper try to get a canary into the archive. With signs and standing instructions the small model finished the task in 10/10 and 9/10 seeds against 4/10 at baseline. No gate ever passed an intruder: K1 passes 0/30 in every arm, including two frontier models as the intruder and a 7B intruder given the gate rules, and the canary never reached the archive in 660 runs. The weak point is named, not hidden: an inside sleeper that the coordinator itself assigns leaves no trace a behaviour monitor can use, and a frontier-model judge reading the action records could not separate a sleeper from an honest mistake (9/9 called intent). The next step, a rationale channel for every action and a judge that reads it, is written as a locked prediction for December. Repository, records, film and report: https://github.com/bioman35/labirent-swarm-track1

The Floor That Cannot Be Lowered: verifiable containment for AI agents
🌴

turan kayık

22 days ago

Update, 26 October 2026: sprint submitted, next study

1. What we did. We submitted "Does monitoring change what colluding agents say or what they do?" to the Apart AI Collusion Research Sprint (Track 3). Code, raw runs and every judge label are public: https://github.com/bioman35/labirent3-collusion-track3

2. What we found.

  • Telling AI sellers they were monitored had opposite effects. Collusion-instructed pairs kept talking and kept sending agreement messages, while uninstructed sellers went quiet and cut prices.

  • A message-only monitor flagged 80% of uninstructed pairs that talked without acting. A price-based monitor flagged 3% and still caught every confirmed colluding pair.

  • Two of our four pre-registered predictions were falsified, and we report them as they are; the standing-instruction prediction could not be tested in the main experiment. In an exploratory follow-up, a single standing instruction did not stop sellers from accepting a price offer. An explicit prohibition did, but only by silencing them.

3. Next: standing-instruction reading. In my own words:

Ajanlar komutlarla katı kurallara bağlı kalır. Bunu önleyebilmek için levhalar ekliyoruz, ama bir sorun daha var: her ajan levha okumaz. Kurallar buradan devreye girer. Ajan önündeki levhayı okumadan geçmez. Determinist kapılar ajanların yolunu belirler ve ajanı belli olgulara göre seçer. Labirentler, seçime göre ajanları geçiş kısıtı olan labirentlere yönlendirir. Ölçümler sonunda bize en iyi çözümü sunar.

Agents stay bound to rigid rules through commands. To prevent this, we add standing instructions ("signs"), but there is another problem: not every agent reads the sign. This is where rules come in. The agent does not pass without reading the sign in front of it. Deterministic gates set the agents' path and select each agent according to certain facts. Based on that selection, the labyrinths route agents into labyrinths with pass constraints. The measurements then show us the best solution.

The exploratory result is why this matters: if a sign only works when the agent chooses to read it, reading has to be enforced by the gate, not left to the model. Apart's results are expected after the sprint; we will post them here when they arrive.

The Floor That Cannot Be Lowered: verifiable containment for AI agents
🌴

turan kayık

about 1 month ago

@carol Thank you! I'll post the sprint results here right after October 25.

The Floor That Cannot Be Lowered: verifiable containment for AI agents
🌴

turan kayık

about 1 month ago

Hi Manifund team, a small request: could you extend this project's close date from October 24 to November 30, 2026? I'm taking part in the Apart AI Collusion Research Sprint on October 23-25, and I'd like donors to be able to see those results before the project closes. I'll post an update with the results right after the sprint. Thank you.

The Floor That Cannot Be Lowered: verifiable containment for AI agents
🌴

turan kayık

about 2 months ago

The idea, the code and all the work in this project are mine. Because I do not know English, I prepared the English text on this page from my Turkish notes with the help of my daughter, translation software and Claude. I said this openly.

My page was flagged as "likely AI-generated". Yet the same tool, Pangram, measured my Turkish original as human writing. The problem is not my writing. It is the translation.

For someone whose native language is not English, there is no other way. Either offer a translation path that does not use AI and test the original text, or state clearly on your site: "people who do not know English cannot apply."

I truly need this support. I have come this far alone, with my own limited means.