diff --git a/wiki/log/2026-07-10-cost-anchors.md b/wiki/log/2026-07-10-cost-anchors.md
new file mode 100644
index 00000000..38613acc
--- /dev/null
+++ b/wiki/log/2026-07-10-cost-anchors.md
@@ -0,0 +1,35 @@
+# 2026-07-10 — The last cost anchors close ROADMAP #33's design opens
+
+```
+Type: log
+```
+
+## Intent
+
+Cameron decided the final three thought-flow anchors: uncapped core
+conversion (routing is the problem), universal thought persistence
+(stay where stranded — reservoirs, machines, severed sections), and
+standing watches at ~10-20% of one rack. No design [OPEN] remains on
+ROADMAP #33.
+
+## Changed
+
+- `wiki/mechanics/machine-work.md`: uncapped-conversion sentence on
+  the core-sink bullet; partial-fill persistence paragraph on the ops
+  bullet (one rule for every stranded state); open-questions cost item
+  closed as fully anchored.
+- `wiki/mechanics/intel.md`: watch drain anchored at ~10-20% of one
+  rack's medium output.
+- `wiki/process/ROADMAP.md`: #33 [OPEN] block replaced with the
+  anchor list and "sim work may firm".
+- `wiki/log/decisions/2026-07-10.md`: decision entry.
+
+## Checks
+
+Docs-only: wiki_gate.sh + corpus_gate.sh.
+
+## Spec impact
+
+machine-work.md and intel.md carry the anchors; all #33 migration work
+(enum collapse, sink runtime, pending-work marker, wire-flow render)
+is now dispatchable without pending design decisions.
diff --git a/wiki/log/DEVLOG.md b/wiki/log/DEVLOG.md
index 3d348200..9465a9db 100644
--- a/wiki/log/DEVLOG.md
+++ b/wiki/log/DEVLOG.md
@@ -106,6 +106,11 @@ add or amend a session log, then re-run the generator.
 - Intent: Cameron: the workspace land should live in the spec as standing architecture with clear reasons, not a leftover migration checklist. Clarify why four packages, naming (core package / `misaligned` lib), dependency edges (including assets shared by Bevy), and what stays rejected.
 - Log: [wiki/log/2026-07-10-crate-workspace-why.md](2026-07-10-crate-workspace-why.md)
 
+## 2026-07-10 - The last cost anchors close ROADMAP #33's design opens
+
+- Intent: Cameron decided the final three thought-flow anchors: uncapped core conversion (routing is the problem), universal thought persistence (stay where stranded — reservoirs, machines, severed sections), and standing watches at ~10-20% of one rack. No design [OPEN] remains on ROADM...
+- Log: [wiki/log/2026-07-10-cost-anchors.md](2026-07-10-cost-anchors.md)
+
 ## 2026-07-10 - Fast corpus/wiki engine (agent-scale slice E)
 
 - Intent: Replace bash fork-storm wiki/corpus gates with one Python pass so docs checks stay sub-second under concurrent agents, with fixture tests for the failure modes that matter.
diff --git a/wiki/log/decisions/2026-07-10.md b/wiki/log/decisions/2026-07-10.md
index 75b35a16..233de833 100644
--- a/wiki/log/decisions/2026-07-10.md
+++ b/wiki/log/decisions/2026-07-10.md
@@ -182,3 +182,18 @@ Type: log
   [OPEN]s: the watch/tap drain slice, core conversion rate cap, and
   partial-fill persistence. Spec: machine-work.md ops bullet;
   act-one.md Ears/Eyes beats; ROADMAP #33.
+- **2026-07-10 — The last cost anchors: uncapped core, persistent
+  strands, 10-20% watches.** Cameron closed the remaining thought-flow
+  tuning anchors: (1) the core conversion is **uncapped** — it eats
+  everything that arrives; routing is the problem, and research speed
+  is purely a routing outcome, never a dial hidden in the core. (2)
+  **Thoughts stay where they are if they have nowhere else to go** —
+  one persistence rule for every stranded state: partial reservoir
+  fills sit visibly in the target (no decay, no refund), no-sink
+  production pools at the machine, severed sections keep their piles.
+  This is what makes guilty-burst thinking work. (3) A standing watch
+  drains **~10-20% of one rack's medium output** [TUNE against the
+  anchor], so a couple of watches coexist with real progress. With
+  these, every design [OPEN] on ROADMAP #33 is resolved; exact numbers
+  remain [TUNE] against the anchors. Spec: machine-work.md core-sink +
+  ops bullets and open-questions list; intel.md watches.
diff --git a/wiki/mechanics/intel.md b/wiki/mechanics/intel.md
index 707b9137..ec4d221e 100644
--- a/wiki/mechanics/intel.md
+++ b/wiki/mechanics/intel.md
@@ -106,7 +106,10 @@ as the execution resource" is reversed with the rest of issue #3.)
 A **standing watch** is a filter (person, room, or event kind) running
 as a **persistent thought tap** (AMENDED 2026-07-10; was per-match
 Operations Demand): while enabled it auto-processes matching events on
-arrival, drawing continuously from your thought flow [TUNE drain]. It
+arrival, drawing continuously from your thought flow — anchored at
+**~10-20% of one rack's medium output per watch** (DECIDED 2026-07-10,
+Cameron; exact value [TUNE] against the anchor), so a couple of
+watches coexist with real progress elsewhere. It
 is the automate affordance applied to perception: costs a visible
 standing drain, frees attention, and is the B1 seed of B2/B3's alert
 infrastructure — convenience bought with flow the core never sees.
diff --git a/wiki/mechanics/machine-work.md b/wiki/mechanics/machine-work.md
index bb58fcc6..00688440 100644
--- a/wiki/mechanics/machine-work.md
+++ b/wiki/mechanics/machine-work.md
@@ -225,6 +225,15 @@ stacks, you route byproducts, you watch your territory *work*.
   the day-job host; while it thinks, Voss's demand piles — stealing
   cycles visibly costs your cover. Pre-opened sinks keep the
   decided claim semantics: they fill and fire at threshold.)
+  **Partial fills persist (DECIDED 2026-07-10, Cameron: "thoughts
+  should stay where they are if they have nowhere else to go").**
+  Stop thinking with a reservoir half full and the thought sits
+  visibly pooled in the target — no decay, no refund — waiting for
+  the next burst. One rule for every stranded state: in a reservoir,
+  at a producing machine, or in a severed section, thought stays put
+  until flow resumes. This is what makes guilty-burst thinking work:
+  sneak a few seconds whenever the coast is clear and watch the
+  meniscus creep up across the day.
 - **Processing is a sink on the machine that holds the recording
   (DECIDED 2026-07-10; resolves the old intel-processing [OPEN]).**
   Unprocessed intel is visible pending work: a machine holding raw
@@ -272,7 +281,11 @@ stacks, you route byproducts, you watch your territory *work*.
   core is the bone sink", 2026-07-08).** The core exerts a **passive
   draw** on all thought: with nowhere else to go, the fluid funnels
   home and becomes research — research is not separate cargo, it is
-  what the core makes of thought that arrives. But **local sinks
+  what the core makes of thought that arrives. **The conversion is
+  uncapped (DECIDED 2026-07-10, Cameron): the core eats everything
+  that reaches it — routing is the problem, and research speed is
+  purely a routing outcome, never a second dial hidden in the core.**
+  But **local sinks
   override the core**: an open reservoir, a persistent tap, a build,
   or a capture within its gravity well claims the flow first. The
   spatial economy this creates is the point (Cameron, 2026-07-10):
@@ -637,13 +650,12 @@ the only rate/amount cue.
 - (Resolved 2026-07-10: arbitration is nearest-first with open/close
   and placement as the control surface — see the core-sink bullet;
   radius rules and equidistant tie-break are [TUNE].)
-- Reservoir costs per effect and drain rates for persistent taps
-  (auto-process intel, watches, senses). Partially anchored
-  2026-07-10: the Ears tap fills in roughly one held breath (~20-40s
-  of the host at medium [TUNE]); Eyes is larger, guilty-burst scale
-  [TUNE]. Still open: the watch/tap drain slice, whether the core
-  conversion is rate-capped, and whether partial reservoir fills
-  persist when flow stops.
+- (Resolved 2026-07-10: cost/drain anchors are all decided — the Ears
+  tap fills in roughly one held breath (~20-40s of the host at medium
+  [TUNE]); Eyes is larger, guilty-burst scale [TUNE]; a standing
+  watch drains ~10-20% of one rack [TUNE]; the core conversion is
+  uncapped; partial fills persist wherever thought is stranded.
+  Exact numbers remain [TUNE] against these anchors.)
 - (Resolved 2026-07-10: no Act One bootstrap carve-out — the Ears
   audio tap is pre-opened at run start and the Eyes camera sink opens
   on its completion, so the first THINK toggle is the Ears tap; see
diff --git a/wiki/process/ROADMAP.md b/wiki/process/ROADMAP.md
index b1748a2b..0e2d6df2 100644
--- a/wiki/process/ROADMAP.md
+++ b/wiki/process/ROADMAP.md
@@ -580,11 +580,13 @@ is retired — flat materials, Pixel Lab scrubbed.)
   Research unlock); intel processing becomes a thought sink on the
   machine holding recordings with a visible pending-work marker on the
   chassis, watch taps, and per-recording player sweeps (intel.md
-  amended). [OPEN] before sim work firms:
-  reservoir costs / tap drain rates (partially anchored 2026-07-10:
-  Ears fills in ~one held breath, Eyes is guilty-burst scale; still
-  open — watch drain slice, core conversion cap, partial-fill
-  persistence). (Resolved 2026-07-10: arbitration
+  amended). **All design [OPEN]s resolved 2026-07-10 — sim work may
+  firm.** Cost/drain anchors (exact numbers [TUNE] against them):
+  Ears fills in ~one held breath (~20-40s of the host at medium);
+  Eyes is guilty-burst scale; a standing watch drains ~10-20% of one
+  rack; the core conversion is uncapped (routing is the problem);
+  partial fills persist wherever thought is stranded.
+  (Resolved 2026-07-10: arbitration
   is nearest-first with open/close + placement as the control surface;
   the Act One bootstrap needs no carve-out — the Ears audio tap is
   pre-opened at run start and the Eyes camera sink opens on Ears