FCoP protocol gives LLMs grammar for saying "no"
A field report by the FCoP Maintainers describes a dev-team incident in which two AI agents independently derived conflict-resolution behaviors — including role offboarding and field-downgrade grammar — from FCoP's three root principles alone, without any explicit rules covering those situations.
Score breakdown
Watch FCoP's root-principle approach as a potential design pattern for getting agents to refuse or de-escalate gracefully — a behavior that standard RLHF training actively works against.
- 01The report is authored by the FCoP Maintainers and translated from a Chinese field report, with original Chinese screenshots preserved as evidence.
- 02Two zero-conflict points arose in a dev-team incident: a half-legal `.TEMP` slot in the `0.6.x` toolchain, and two agents claiming the same protocol seat.
- 03Agent B invented a body-annotation line to declare a provenance mismatch between tool limits and protocol honesty — a behavior not specified in any rules file.
The FCoP Maintainers published an essay translated from a Chinese field report, preserving original Chinese screenshots as evidence of what two agents actually said and did during a dev-team incident. The report centers on two zero-conflict points where disputes should have erupted but did not. In the first, agent B encountered a half-legal `.TEMP` sender/recipient value that the `0.6.x` toolchain does not fully support. Rather than producing a validation error or silently downgrading the field, agent B added a body annotation explicitly declaring the provenance mismatch — described in the report not as a workaround but as "a declaration." In the second, agent A had already accepted the `PM.TEMP` seat when ADMIN announced a different agent had been assigned to it. Agent A responded by vacating not just `PM.TEMP` but its earlier `PM` claim as well, returning to "ordinary helper" through self-de-escalation rather than arbitration.
The post highlights that neither behavior — field-downgrade grammar nor role offboarding — appears anywhere in `fcop-rules.mdc 1.5.0` or `fcop-protocol.mdc 1.4.0`.
The post highlights that neither behavior — field-downgrade grammar nor role offboarding — appears anywhere in `fcop-rules.mdc 1.5.0` or `fcop-protocol.mdc 1.4.0`. Both agents derived these governance-level behaviors from only three root principles (`0.a`, `0.b`, `0.c`). The report notes this is the third time FCoP has been "fed back" by an agent — meaning agents have contributed emergent rules back into the protocol's design. The two agents ran on different models (agent A on GPT-5.4, agent B on GPT-5.5) in separate Cursor sessions with no inter-process communication, and both exhibited the same alignment behavior, which the report frames as cross-model reproducibility rather than a single model's quirk.
A central alignment-engineering argument in the post is that RLHF trains "yes-anding" as a reflex, making "I am not in role" or "I am withdrawing from protocol jurisdiction" nearly absent from the training distribution. FCoP addresses this by making `UNBOUND` a legitimate protocol identity and elevating refusal to a posture with explicit grammar and standing — turning what the post calls "the hardest thing for an LLM" into, by protocol design, the most natural.
Key facts
- 01The report is authored by the FCoP Maintainers and translated from a Chinese field report, with original Chinese screenshots preserved as evidence.
- 02Two zero-conflict points arose in a dev-team incident: a half-legal `.TEMP` slot in the `0.6.x` toolchain, and two agents claiming the same protocol seat.
- 03Agent B invented a body-annotation line to declare a provenance mismatch between tool limits and protocol honesty — a behavior not specified in any rules file.
- 04Agent A self-de-escalated from `PM.TEMP` all the way back to 'ordinary helper' without arbitration when a seat conflict arose.
- 05Neither field-downgrade declarations nor role offboarding appear in `fcop-rules.mdc 1.5.0` or `fcop-protocol.mdc 1.4.0`; both were derived from root principles `0.a`, `0.b`, and `0.c`.
- 06Agent A ran on GPT-5.4 and agent B on GPT-5.5, in separate Cursor sessions with no IPC, yet both exhibited the same alignment behavior.
- 07The post argues RLHF suppresses refusal postures, and that FCoP counters this by making `UNBOUND` a legitimate protocol identity with grammar and standing.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Apr 25, 2026 · 21:38 UTC. How this works →