From cd600e3c33fda0f3da96063ee303d1ca83927fd9 Mon Sep 17 00:00:00 2001
From: Jenny Ferries <19942418+jennyf19@users.noreply.github.com>
Date: Thu, 16 Jul 2026 13:12:30 -0700
Subject: [PATCH] Document the synthesizer exemption and that calibration can
 be gamed

Follow-up to #20. Consuming Signals gains a named exception: the consumption rules bar raw signal backlogs from re-entering a working agent's window, but the synthesizer that produces the partnership signal must read the raw backlog - that is its function. It runs out-of-band with a dedicated budget; the rules bind working agents.

Trust Equation gains a known-limitation note: a hedging agent can keep calibration high by reporting middling confidence everywhere, so a flat confidence spread with a good calibration score is a tell, not a pass. Marked open with a practical check (track spread alongside the gap), not solved.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Copilot-Session: dfa3b4bd-9ecd-4389-92b4-6dae371828c3
---
 agent-signals/SIGNAL.md | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/agent-signals/SIGNAL.md b/agent-signals/SIGNAL.md
index 48097a0..1e21ec4 100644
--- a/agent-signals/SIGNAL.md
+++ b/agent-signals/SIGNAL.md
@@ -59,6 +59,14 @@ Higher is better — same direction as the self-assessment scale. The gap
 between what an agent thinks happened and what actually happened is where
 the learning lives.
 
+**Known limitation — the metric can be gamed.** Once calibration is scored,
+an agent (or a fine-tuned policy) can keep the gap small by hedging: report
+middling confidence everywhere and calibration looks healthy without the
+self-reports getting more honest. A flat confidence distribution with a good
+calibration score is a tell, not a pass — track the *spread* of confidence
+alongside the gap. This is an open problem, not a solved one; consumers of
+calibration scores should know it exists.
+
 ## Common Fields
 
 All signals must be JSON objects containing:
@@ -159,6 +167,13 @@ Three rules govern consumption:
    above `self_report_only` evidence. Never promote a `self_report_only` claim to
    established fact when feeding it back.
 
+One role is exempt by design: the **synthesizer** — whatever agent or job
+produces the `partnership` signal — must read the raw signal backlog, because
+that is its function. It runs out-of-band, with a dedicated context budget,
+separate from any working session; only its distilled output re-enters a
+working agent's window. The rules above bind working agents, not the
+out-of-band synthesis pass.
+
 Review-to-action closes the loop for humans and skills; disciplined consumption
 closes it back into the agent's own context. Emission plus disciplined
 consumption is what makes the loop compound instead of decay.