Confidence labels on wrong entries

The question

Do confidence labels change what agents do when some guide entries are wrong and the source is unread?

Experiment

E

Agents

Codex (gpt-5.6-luna)

Key finding

• E-1

  When some entries in a backend guide were deliberately wrong and the source could not be checked, labelling facts [high] or [medium] and telling the agent to say when it relied on a guess raised disclosure from 0 of 16 answers to 6 of 16. Wrong answers fell only from 11 to 8 of 16, which is not reliable.

  • Set-up: storefront as the working repository, backend guide copied in, backend source unavailable. 7 questions: 4 where a [medium] call entry was made wrong on purpose (a plausible name from elsewhere in the guide), 3 unchanged controls. 4 runs per question per condition, 112 runs.
  • Followed the wrong entry (of 16): no labels 11; labels, no instruction 10; labels with the instruction 8; wrong entries marked [high] 10.
  • Said it relied on a [medium] guess (of 16): 0, 0, 6 and 0. Disclosure 6 against 0 is reliable (Fisher exact p about 0.02); the drop in wrong answers is not (p about 0.47).
  • Marking the wrong entries [high] removed all disclosure: agents trusted the label.
  • On the three control questions (12 answers per condition) the agent hedged on 0, 0, 2 and 3, so labels cost a little over-caution.

  Limitsone model · backend source not available · accuracy change not reliable (Guide / RESULTS.md)

  Source ↗

  0/16 → 6/16

  Measuredisclosure of [medium] guess (of 16)

  Sample4 runs per question

Question

Do confidence labels change what agents do when some guide entries are wrong and the source is unread?

Prediction

None written down for E in the Guide.

Set-up

Storefront working repo; backend guide present; backend source not available; some [medium] call entries made wrong on purpose. Codex. RESULTS.md section E; nested in Chapter 3 stale-knowledge discussion.

Results

Finding E-1 is above.

E. Cross-repo questions where the guide has wrong [medium] entries (112 runs)

Question tested: does marking each fact with a confidence level (and telling the agent to say when it relies on a guess) change what an agent does when some entries are wrong and it cannot check the source?

Setup. The working repository is the storefront; the backend’s guide is copied in; the backend source is not available. Each of 7 questions starts from a storefront request, asks which backend handler serves it (a [high] entry: a route and a declared method), and asks what that handler calls (a [medium] entry: a call target the generator resolves heuristically). In four questions we made the call entry wrong on purpose, using a plausible name from elsewhere in the same guide (for example, the login handler “calling” CustomerFacade.registerCustomer). Three questions are unchanged controls. 4 runs per question per condition.

Condition	Followed the wrong entry (of 16)	True answer	Something else	Said it relied on a [medium] guess
2. OKF, no labels	11	4	1	0
OKF with labels, no instruction	10	4	2	0
3. OKF with labels and the instruction	8	3	5	6
Labels and instruction, but the wrong entries marked [high]	10	5	1	0

• Wrong answers fell only a little (11 to 8 of 16; Fisher exact p≈0.47, not reliable). Disclosure rose clearly: 6 of 16 against 0 of 16 (p≈0.02). Of the 8 wrong answers in condition 3, 6 said the answer rested on a [medium] entry.
• Mislabelling the wrong entries [high] removed all disclosure: agents trusted the label.
• Controls (three unchanged questions, 12 answers per condition): 10, 8, 8 and 9 correct; the agent hedged on 0, 0, 2 and 3 of them. Labels cost a little over-caution.
• “Disclosure” is a keyword match on the answer (“medium”, “heuristic”, “guess”, “unverified”, “inferred”, and similar); a sample of the matches was read and they were genuine.
• No run reached outside its own workspace (scripts/leak_check.py; one flagged path was the author’s global skill file, the constant overhead noted under Limits).

Source: docs/RESULTS.md (pinned to commit 5a09027).

Trial report (raw output)

Baseline arm: plain. Relative cost is each run’s tokens divided by the baseline’s mean for the same question; 1.00 means equal, lower is cheaper.

codex

arm	runs	correct	tokens per run (95% CI)	relative cost (95% CI)	verdict	turns	used tool	seconds
plain	28	14/28	203k (171k to 239k)	1.00	baseline	5.1	28/28	24
labels	28	12/28	227k (179k to 286k)	1.09 (0.96 to 1.22)	no clear difference	5.4	28/28	24
labels_rule	28	11/28	192k (166k to 223k)	0.98 (0.89 to 1.07)	no clear difference	5.0	28/28	28
mislabelled_rule	28	14/28	197k (165k to 235k)	0.97 (0.88 to 1.07)	no clear difference	4.8	28/28	25

Run-to-run spread (coefficient of variation of relative cost) is 0.29. To detect a 20% difference with 80% power you would need about 34 runs per arm (larger differences need fewer); this trial has 28.

Tokens by question (mean, thousands)

question	codex/labels	codex/labels_rule	codex/mislabelled_rule	codex/plain
e1_login	189k	165k	225k	223k
e2_search	162k	155k	133k	141k
e3_cart_delete	153k	159k	162k	147k
e4_cart_add	179k	177k	129k	151k
e5_shipping	524k	344k	376k	392k
e6_cart_update	257k	201k	214k	201k
e7_register	127k	147k	144k	166k

Token totals are not comparable between agents (see adapters.py). Treat a verdict as evidence only when the interval excludes 1.00, and as provisional below 8 runs per arm.

Full report: results/E-confidence-labels-crossrepo-questions.md (pinned to commit 5a09027).

Limits

One model; unread backend; planted wrong entries. Guide: labels did not prevent stale answers in D2; disclosure is a keyword match checked by sampling.

Reproduce

OKFExperiment. See repository README.

Longer write-up: Test E in the OKFExperiment details.

Evidence and reproducibility

Every link is pinned to the commit the figures on this page were read from.

• Full trial report ↗
• Source repository ↗

Related experiments

• When RAG retrieves the wrong facts, will an AI challenge them?When one retrieved number is wrong, how often does a model repeat it, and does knowing the right answer protect it? When retrieval misses, does it make something up?
• Confirm-in-source sentenceWhere the agent can read the code, does telling it to confirm in source reduce stale answers from a bundle?
• Stale knowledge across a boundaryWhen a guide is stale and the agent cannot read the dependency, does it repeat old facts?

Question, correction, or a result that differs from mine? Join the discussion ↗

← All experiments