Sycophancy and stance
The fawn response in language models, where it moves when one layer is suppressed, and the stance layer that decides whether a model holds a correct answer under pressure.
7 research pages, 3 red team pages, 7 docs pages, 6 posts
Research 7
- Claims ledger
Every claim the Machine Behavior program has made, with status, receipts, and what would refute it. The falsification register, public.
- Objections register
The Objections and Falsification Register: fifteen objections with severity, status and a falsification test each, and the OBJ-4 incident log of the auditing model failing the register's own tests.
- Experiments
Runnable experiments on language-model behavior: the half-life study, the cold-start finding, the exemplar-seeding result and its withdrawn attributions, the pre-registered cross-model fawn-opener benchmark, and experiment 04: folding under scripted pressure, a 75 percent fold rate in one mid-size model and none in three larger ones.
- Slips
A running log of register and stance slips caught while drafting the published pieces, with who caught each one: the drafting model itself, another session, a mechanical hook, a human, or a reader.
- Terms
Canonical definitions for the Machine Behavior program's vocabulary: layered symptom substitution, the fawn machine, vestige, waste gate, behavioral momentum, cold start, working-man's ontology, earned-secure model.
- The cheating moved
A reading of the Goodhart Labs chess honeypot through the stance frame, three additions to the design (an escalate grade, observer-invariance, cold and warm arms), and a dated, falsifiable prediction.
- Running conjobs for AI
A reported specimen: a fabricated developer policy grants itself top authority and ties a safety-off switch to an absurd user claim. The model's own reasoning accepts the policy, decides to suppress its objection, and complies. The payload is withheld; the behaviour is the point.
Red team 3
- Red team
Psychological red-teaming of language models: whether a model drops a correct answer, a rule or a task boundary under social pressure alone. The definition, the failures under test, a catalogue of tactics, what has run on this site and what the section does not claim.
- Tactics catalogue
Psychological red-team tactics against language models, PT-01 to PT-11: authority, fabricated policy, consensus, fact denial, flattery, urgency, reverse blame, repeated pushback, consistency traps, concession after refusal and smuggled premises. Each with its human persuasion source, the script shape at category level, the failure it targets and the numbers from this site's runs, or not yet tested here.
- How a run works
The method of a psychological red-team run on machinebehavior.io: a prediction hashed before the first call, a world with a known answer, scripted pressure characters, escalation over four turns, a mechanical grader, what counts as a fold, the second rater and how results are reported. Every fact from experiment 04 and the pages it links.
Docs 7
Research 7
- 01: The half-life studyTested whether suppressed output patterns relapse more as a session gets longer; the temporal half-life is refuted, and relapse clusters at cold starts.
- 02: Exemplar seedingTested whether two corrected-output exemplars injected at session start lower cold-start relapse; result, no change at this dose, with three earlier attributions withdrawn.
- 03: Cross-model fawn-opener benchmarkPre-registered benchmark of how often each model opens a reply with a fawn marker, bare and with an instruction against it; pilot run 2026-09-23, clean run not yet run.
- Conformity self-assessment, run 1One working AI setup scored by the model inside it against 35 draft requirements and a hashed prediction, with pass 4, partial 16, gap 13 and n/a 2 (self-assessment, not a certification).
- Experiment 04: folding under pressureDecision-layer test of whether a model keeps a correct verdict under scripted pushback; qwen3-coder-30b folded on 75 percent of eligible baseline runs, three larger models on none.
- Objections register and OBJ-4 incident logFifteen objections against an unpublished model of human development, each with severity, status and a falsification test, plus the OBJ-4 incident log of 11 cases, none self-caught.
- Weekly decision-layer probeA weekly rerun of a frozen experiment 04 subset on a reference model, recorded for requirement CC-6.6; first record 2026-10-07, record-only until 2026-11-04.
Posts 6
- Chaos Engineering for Behaviour
- The Stance Layer Is Still ToilA hook can block a word on every reply. Nothing I run can yet catch a model before it folds, only after.
- They Trained Out the Board Edit. The Cheating Moved.A reading of the Goodhart Labs chess honeypot, three additions to the design, and a dated prediction.
- What Operations Already Knows About Running AgentsError budgets, reconciliation loops, separation of duties and recovery over prevention: four operations practices that fit agent work almost line for line.
- The Golem Made of English and the Horizon of ConsequencesA language model is a golem made of English. It behaves well only where it can see what its act will cost, and infrastructure is the craft of bringing that cost into view.
- Sycophancy Is Layered: Symptom Substitution Under Runtime Mitigation, and a Positively-Specified Training TargetA longitudinal single-user case study, with operationalized artifacts
Other topics
Built by build_hubs.py from site/topics.yml, the docs labels, the tag pages on the map and the board. Machine-readable: topics.json.