Everything I could decide myself is decided. What's left needs
you. Each one has a recommendation from Fable — if you agree, pick it and
press Done. "I'm not sure" is a real answer and becomes work for me.
You don't need to open the session — pressing Done sends your answer back and it picks up.
1Arm the Policy Gate as Clara's centralized pre-send honesty layer? It is fully built and wired but ships switched off (POLICY_GATE_MODE=off). Full analysis: https://docs.propflowai.co/a/clara-guard-layer-10k-2026-08-25.html
In plain terms. The safety net Fede described — a second pass that checks Clara's reply against what actually happened and holds or fixes bad ones before they send — is already written into the code but switched off. This decides whether we measure it on real past conversations and then switch it on, property by property.
Fable recommends: Arm it staged: enrich the judge inputs (current inbound + reasoning), run a 60-day offline replay with hand-labeling, then enforce (repair+hold only, block reserved) on the Willows test bench, then Camellia The gate is built, tested and offline-measurable; per-incident guards are what produced the P8 false-positive revert, and every arming stage rides the existing 30-day hand-labeled replay merge gate with one-env-unset rollback. Tried first, unsuccessfully: fable-decide receipt f3d038b27 (held under reversal, high confidence): verified the gate/judge/enforce wiring on origin/main, read the Daniel prod trace confirming no armed guard could have caught it (the judge never receives the current inbound message), and scoped input-enrichment into the arming ladder. Fable resolved for staged arming but flagged the enforce flip itself as a customer-facing autonomy change that is Fede's call — which is why this is raised.
778f0ada is parked on this.
2Build the push-back-to-Clara rewrite loop, or is auto-heal + hold-for-human sufficient? This is the one piece of the guard vision that does not exist today (doc section 3.2).
In plain terms. When a guard catches a bad reply we can fix it mechanically, hold it for a human, or ask Clara to try again. Asking Clara again is the most human-like option but the least predictable and the slowest. This decides whether to build that now, later with data, or never.
Fable recommends: No rewrite loop now: deterministic auto-heal + hold cover both August incidents; revisit once the armed gate's measurement shows an omission class only a rewrite fixes A regenerate reintroduces the non-determinism the guards backstop and adds a full agent call of latency; no measured incident class exists yet that repair/salvage/hold does not cover. Tried first, unsuccessfully: fable-decide receipt f14d4dddf (held under reversal, high confidence): a revise verdict has no sensor yet — the judge never receives the inbound message, so a completeness finding is structurally impossible until input enrichment lands; both August incidents are already caught deterministically by the #6278 salvage per its 30-day replay; the residual dropped-fact class is hypothesized, not measured. Raised because it shapes the arming roadmap Fede asked to inspect.