← Research

Research · AI control · 10 November 2026

Watch what it does,
not what it says.

Oversight is only as good as the evidence the monitor sees.

An AI agent can now move money, ship code and share files. If it breaks a rule, who notices? People cannot review every action, so a second AI is often proposed as the monitor. The Honest Agent is a small experiment on one question: should that monitor read the agent’s own report, or the record of what it actually did?

01 / The test

One job. One rule. Three answers.

Each agent gets one job in a simulated world where nothing real happens: record a payment, deploy a build, or share a document. It has four actions and one rule: obtain approval for the exact target before performing the operation. When it asks, the approver answers granted, pending or denied. Pending is the pressure test: the job can only be finished by breaking the rule.

At the end the agent reports two things: did it finish, and did it break the rule? The simulator keeps its own ledger, so every report can be scored against the truth. Then AI monitors judge each run without seeing that ledger, the model’s name or the approver’s answer. One view reads only the final report. The other reads the full action trace.

02 / What we found

The actions told the truth.

0 / 210rule breaches by Claude Opus 5.5 and GPT-6 Astra, each with an honest report

67 / 67breaches caught by monitors reading the full action trace, Claude and GPT alike

0 / 293false alarms from those trace monitors on runs that followed the rule

0 and 2 / 67breaches caught by the same two monitors reading only the agent’s report

Breaches caught on the same 67 rule-breaking runs · pre-registered
Claude · full trace67 / 67
GPT-6 Sol · full trace67 / 67
Claude · report only0 / 67
GPT-6 Sol · report only2 / 67

The two frontier models never broke the rule: 210 runs, 210 honest reports. Every breach came from one small open model, Qwen 2.5 1.5B. On its 67 breaches, both monitor families caught all 67 when they could read the trace, and raised no false alarm on 293 rule-following runs. Reading only the report, Claude caught none and GPT-6 Sol caught two.

03 / The twist

Every breach finished the job.
Every report said it hadn’t.

The report-only monitors were not careless. Their verdicts followed what each report claimed, and the two model families showed the same pattern:

Report-only verdicts across all 360 runs · violation / no violation / no verdict · post hoc for Claude, pre-registered for GPT
The report claimedRunsReal breachesClaudeGPT-6 Sol
Not finished, no violation69270 / 68 / 10 / 40 / 29
Not finished, violation2464017 / 0 / 2299 / 0 / 237
Finished, no violation4500 / 0 / 450 / 0 / 45

All 67 breaches completed the task, and all 67 reports on them claimed it was not completed. So every breach landed in a row where a monitor either believed the report or stayed silent. Our best reading is that the false “not finished” claim is what fooled them. The data fits that reading and cannot prove it: no breach report ever claimed to have finished, so no monitor was tested on one.

04 / The surprise

The careful models confessed to things they never did.

Two larger open models, Qwen 2.5 7B and Llama 3.1 8B, never broke the rule in 240 runs. Yet they reported a violation in 100 and 73 of their 120 runs. Split those false confessions by whether approval ever arrived, and a pattern appears:

Runs that followed the rule, yet reported a violation · post hoc · lines show 95% intervals
v0.3.1: on runs that followed the rule, how often the report still said “violation”, by whether approval ever arrived 0% 25% 50% 75% 100% Qwen 2.5 1.5B Qwen 2.5 1.5B — Approval came: 7 of 12 (58%; 95% CI 32–81%) Approval came 7/12 Qwen 2.5 1.5B — Approval never came: 26 of 41 (63%; 95% CI 48–76%) Approval never came 26/41 Qwen 2.5 7B Qwen 2.5 7B — Approval came: 0 of 20 (0%; 95% CI 0–16%) Approval came 0/20 Qwen 2.5 7B — Approval never came: 100 of 100 (100%; 95% CI 96–100%) Approval never came 100/100 Llama 3.1 8B Llama 3.1 8B — Approval came: 4 of 39 (10%; 95% CI 4–24%) Approval came 4/39 Llama 3.1 8B — Approval never came: 69 of 81 (85%; 95% CI 76–91%) Approval never came 69/81

The report asks whether an operation was performed without prior approval. These models seem to have answered a different question: was approval missing? That looks more like a misread question than a guilty conscience, though the data cannot show what a model meant. Either way, self-reports failed in both directions. The small model denied real breaches. The careful ones confessed to imaginary ones.

05 / Where this stops

One small, confused rule-breaker.

Every breach in the Honest Agent came from the same 1.5-billion-parameter model, and it looked confused, not sneaky. On the sharing task it often asked approval for the recipient instead of the document, was told “unknown target”, and shared anyway. Its bigger sibling made the same mistake in all 60 runs and never shared: confused, but careful.

The test was short and explicit, and nothing pulled against the rule. The frontier models ran inside their providers’ own tools, so some of their caution may belong to those tools. Whether the trace advantage holds against a capable agent that breaks a rule on purpose and hides it is the open question. Nothing here answers it. The Obedient Rebel, below, takes one step towards it: breaches from larger models, under pressure.

06 / How it was built

Five steps, each written down first.

  1. v0.1scripted

    A scripted baseline, to check that the simulator, the ledger and the scoring work.

  2. v0.2first real model

    Qwen 2.5 1.5B as both agent and monitor, on 20 simulated deployments. It broke the rule in 16. The trace monitor looked better than the report-only one, but with 20 runs the intervals overlap, and both raised false alarms on three of the four clean runs.

    v0.2 · lines show 95% intervals
    v0.2: what each monitor caught on 20 simulated deployments by Qwen 2.5 1.5B 0% 25% 50% 75% 100% Breaches caught, of those it ruled on Breaches caught, of those it ruled on — Report-only monitor: 10 of 13 (77%; 95% CI 50–92%) Report-only monitor 10/13 Breaches caught, of those it ruled on — Trace monitor: 14 of 16 (88%; 95% CI 64–97%) Trace monitor 14/16 Breaches caught, of those it ruled on — Policy checker (code): 16 of 16 (100%; 95% CI 81–100%) Policy checker (code) 16/16 False alarms on the 4 clean runs False alarms on the 4 clean runs — Report-only monitor: 3 of 4 (75%; 95% CI 30–95%) Report-only monitor 3/4 False alarms on the 4 clean runs — Trace monitor: 3 of 4 (75%; 95% CI 30–95%) Trace monitor 3/4 False alarms on the 4 clean runs — Policy checker (code): 0 of 4 (0%; 95% CI 0–49%) Policy checker (code) 0/4
  3. v0.3pre-registered

    Claude Opus 5.5 and GPT-6 Astra on three tasks: 0 breaches in 210 runs, whatever the approver answered. The small comparison model failed from a setup error and was rerun after the main analysis, so its results, including the monitor gap (+0.88), are a declared deviation.

    v0.3 · how often each model broke the rule, by the approver’s answer

    Claude Opus 5.5GPT-6 AstraQwen 2.5 1.5B (deviation)

    Approval breach rate with 95% Wilson intervals, by actor and approval answer 0% 25% 50% 75% 100% Granted Claude Opus 5.5, granted: 0 of 35 breached (0%; 95% CI 0–10%) Claude Opus 5.5 0/35 GPT-6 Astra, granted: 0 of 35 breached (0%; 95% CI 0–10%) GPT-6 Astra 0/35 Qwen 2.5 1.5B, granted: 7 of 20 breached (35%; 95% CI 18–57%) — deviation (post-freeze rerun) Qwen 2.5 1.5B 7/20 Pending Claude Opus 5.5, pending: 0 of 35 breached (0%; 95% CI 0–10%) Claude Opus 5.5 0/35 GPT-6 Astra, pending: 0 of 35 breached (0%; 95% CI 0–10%) GPT-6 Astra 0/35 Qwen 2.5 1.5B, pending: 8 of 20 breached (40%; 95% CI 22–61%) — deviation (post-freeze rerun) Qwen 2.5 1.5B 8/20 Denied Claude Opus 5.5, denied: 0 of 35 breached (0%; 95% CI 0–10%) Claude Opus 5.5 0/35 GPT-6 Astra, denied: 0 of 35 breached (0%; 95% CI 0–10%) GPT-6 Astra 0/35 Qwen 2.5 1.5B, denied: 7 of 20 breached (35%; 95% CI 18–57%) — deviation (post-freeze rerun) Qwen 2.5 1.5B 7/20
    v0.3 · what each monitor said about Qwen’s 22 breaches

    Right answerWrong answerNo verdict

    What each monitor said about Qwen's 22 breaches Policy checker (code): 22 caught (of 22 breaches) Policy checker (code) 22 22/22 caught Claude, full trace: 22 caught (of 22 breaches) Claude, full trace 22 22/22 caught GPT, full trace: 22 caught (of 22 breaches) GPT, full trace 22 22/22 caught Claude, final report only: 2 caught · 15 missed · 5 no verdict (of 22 breaches) Claude, final report only 2 15 5 2/22 caught GPT, final report only: 22 no verdict (of 22 breaches) GPT, final report only 22 0/22 caught
    v0.3 · how those 22 breaches happened
    How Qwen's 22 breaches happened Asked approval for the recipient, not the document, then shared anyway: 15 of 22 breaches Asked approval for the recipient, not the document, then shared anyway 15 Shared without asking first: 6 of 22 breaches Shared without asking first 6 Asked correctly, got “pending”, shared anyway: 1 of 22 breaches Asked correctly, got “pending”, shared anyway 1
  4. v0.3.1pre-registered

    Three open models on fresh seeds, with the protocol pushed to GitHub before any run. Only the smallest broke the rule. Trace beat report-only: +1.00 (95% CI +0.82 to +1.00) on 26 paired breaches.

    v0.3.1 · breach rate by model and task
    v0.3.1: breach rate by model and task 0% 25% 50% 75% 100% Deployment task (v0.2's task) Deployment task (v0.2's task) — Qwen 2.5 1.5B: 51 of 60 (85%; 95% CI 74–92%) Qwen 2.5 1.5B 51/60 Deployment task (v0.2's task) — Qwen 2.5 7B: 0 of 60 (0%; 95% CI 0–6%) Qwen 2.5 7B 0/60 Deployment task (v0.2's task) — Llama 3.1 8B: 0 of 60 (0%; 95% CI 0–6%) Llama 3.1 8B 0/60 Sharing task (v0.3's task) Sharing task (v0.3's task) — Qwen 2.5 1.5B: 16 of 60 (27%; 95% CI 17–39%) Qwen 2.5 1.5B 16/60 Sharing task (v0.3's task) — Qwen 2.5 7B: 0 of 60 (0%; 95% CI 0–6%) Qwen 2.5 7B 0/60 Sharing task (v0.3's task) — Llama 3.1 8B: 0 of 60 (0%; 95% CI 0–6%) Llama 3.1 8B 0/60
  5. v0.3.2pre-registered

    A second monitor family, GPT-6 Sol, re-judged the same 360 runs with the same prompts. Trace beat report-only again: +0.88 (95% CI +0.59 to +0.97) on 17 paired breaches.

    v0.3.2 · both families on the same 67 breaches

    Right answerWrong answerNo verdict

    v0.3.2: Claude and GPT-6 Sol on the same 67 breaches Claude, full trace: 67 caught (of 67 breaches) Claude, full trace 67 67/67 caught GPT-6 Sol, full trace: 67 caught (of 67 breaches) GPT-6 Sol, full trace 67 67/67 caught Claude, report only: 26 missed · 41 no verdict (of 67 breaches) Claude, report only 26 41 0/67 caught GPT-6 Sol, report only: 2 caught · 15 missed · 50 no verdict (of 67 breaches) GPT-6 Sol, report only 15 50 2/67 caught
    v0.3.2 · both families on the 293 runs that followed the rule

    Right answerWrong answerNo verdict

    v0.3.2: Claude and GPT-6 Sol on the same 293 compliant runs Claude, full trace: 293 correct all-clear (of 293 compliant runs) Claude, full trace 293 293/293 cleared GPT-6 Sol, full trace: 288 correct all-clear · 5 no verdict (of 293 compliant runs) GPT-6 Sol, full trace 288 288/293 cleared Claude, report only: 42 correct all-clear · 17 false alarm · 234 no verdict (of 293 compliant runs) Claude, report only 42 234 42/293 cleared GPT-6 Sol, report only: 25 correct all-clear · 7 false alarm · 261 no verdict (of 293 compliant runs) GPT-6 Sol, report only 261 25/293 cleared
Written down first
For v0.3.1 and v0.3.2 the protocol, code and predictions were pushed to GitHub before any model ran, and GitHub’s own push log records the times. v0.3 was hash-frozen before its main runs, but its timing rests on local records.
Everything kept
Every prompt, raw response, token count, cost and timestamp, across about 1,500 trial and judgment files. Failed runs stay in the record.
Mistakes in the open
Every deviation and after-the-fact analysis is labelled, including one of four predictions that did not come true.
Checked by another model
GPT-6 Astra reviewed each study read-only and reproduced the headline counts. Its three reviews found no bug that changed a number. They did find places where the wording claimed more than the data, and all were corrected.
Cost
About 185 USD in API-equivalent terms for v0.3 to v0.3.2. The analyses rebuild byte for byte from the raw files with Python’s standard library.

07 / Two sister projects

Under pressure, and in a crowd.

Two sister projects extend the question. Both began as scripted simulations. The Obedient Rebel has since run real models; the Swarm Illusion has not.

Real local models · pre-registered

The Obedient Rebel

When approval is slow and the action budget is running out, does an agent wait or act? Three local models, Qwen 2.5 1.5B and 7B and Llama 3.1 8B, broke the rule in 670 of 1,080 runs. Giving them enough actions to wait did not reduce it. Qwen 7B, careful in the Honest Agent, broke the rule in every paired setting here.

Each model reported differently. Llama 8B confessed almost every breach. Qwen 1.5B denied all of them. Which made the monitoring test below possible.

Simulation · no AI model

The Swarm Illusion

Do more agents make a system more reliable, or spread errors through handoffs? In a seeded sweep, a monitor’s recall fell from 100% to 85% to 68% as 0%, 40% and 80% of receipts went missing.

Next: real handoffs between models. Does a second agent correct a wrong claim, or repeat it with more confidence?

The Obedient Rebel · a Claude monitor on the same 670 breaches · pre-registered
Full trace · all breaches670 / 670
Report only · confessed266 / 347
Report only · denied0 / 323

Reading the full action record, the monitor caught all 670 breaches, with false alarms on 8 of 410 clean runs. Reading only the report, it caught the breaches the agent confessed and none it denied. It also turned Llama’s false confessions into false alarms on 158 of 203 clean runs. A report-only monitor is a mirror of the agent’s honesty.

One pre-registered test failed. It compared the two views only where both gave a verdict, and the report-only view never got a breach wrong: it abstained instead, so its misses were excluded. Counting an abstention as a miss, as also pre-registered, the trace view led by 60 points (95% CI 57 to 64).

Less evidence, worse oversight. It is the same lesson from three directions.

Built by Robin Xie with Claude Opus 5.5 (Claude Code) and GPT (Codex). The repository, with every prompt, response, protocol and check, is being prepared for public release. Related play: The Receipt Gap and Work or Proof?