docs(backlog): B-0838 — caustic-engineered bloom filter discriminators for remote-code trust layer (Kestrel-v2 ferry; phased)

docs/BACKLOG.md

@@ -772,6 +772,7 @@ are closed (status: closed in frontmatter)._
 - [ ] **[B-0832](backlog/P2/B-0832-installer-nmtui-wifi-rescan-refresh-button-overlapping-networks-empirical-aaron-2026-05-26-physical-hardware-support-test.md)** installer nmtui WiFi step needs visible rescan/refresh path — empirical from operator's physical hardware-support test 2026-05-26 (20+ overlapping networks; target SSID not initially in scan list) (Aaron 2026-05-26)
 - [ ] **[B-0834](backlog/P2/B-0834-installer-preserve-install-log-to-file-failures-warnings-scrollback-empirical-aaron-2026-05-26.md)** installer must preserve install log to file — failures + warnings scroll past faster than operator can read (empirical from 2026-05-26 physical hardware-support test; gh login not reached) (Aaron 2026-05-26)
 - [ ] **[B-0837](backlog/P2/B-0837-max-addison-committee-review-asymmetric-critic-with-clarity-first-rule-draft-aaron-2026-05-26.md)** Max + Addison committee review of asymmetric-critic-with-clarity-first rule draft (per Kestrel-v3 substrate-honest disclaimer + operator authorization) (Aaron 2026-05-26)
+- [ ] **[B-0838](backlog/P2/B-0838-caustic-engineered-bloom-filter-discriminators-remote-code-trust-layer-kestrel-v2-aaron-2026-05-26.md)** caustic-engineered bloom filter discriminators for remote-code trust layer (Kestrel-v2 ferry; multi-learned-bloom-filter intersection with caustic-geometry-shaped agreement region) (Aaron 2026-05-26)
 
 ## P3 — convenience / deferred
 

docs/backlog/P2/B-0838-caustic-engineered-bloom-filter-discriminators-remote-code-trust-layer-kestrel-v2-aaron-2026-05-26.md

@@ -0,0 +1,199 @@
+---
+id: B-0838
+priority: P2
+status: open
+title: caustic-engineered bloom filter discriminators for remote-code trust layer (Kestrel-v2 ferry; multi-learned-bloom-filter intersection with caustic-geometry-shaped agreement region) (Aaron 2026-05-26)
+effort: L
+ask: aaron 2026-05-26
+created: 2026-05-26
+last_updated: 2026-05-26
+depends_on: []
+composes_with:
+  - B-0664
+  - B-0648
+tags: [bloom-filter, caustic-engineering, remote-code-discriminator, trust-then-verify, multi-oracle, inverse-design, research-direction-flavored, kestrel-v2-substrate]
+---
+
+## Problem
+
+Per operator 2026-05-26 (Aaron-forwarded Kestrel-v2 ferry, preserved via PR #5356):
+
+> "do you think there is a way i can create caustic lens shaped bloom
+> filters for the remote code discriminators?"
+
+The framework needs a trust-layer discriminator for remote code (per
+the NCI-as-cyberattack-prevention substrate from Lior-website / PR-5342,
+which named "speech is literally remote code execution on another node's
+memory substrate"). Current substrate has the bloom-filter mental
+model (`B-0822`-class worry-as-opposite-bloom-filter) + the
+trust-then-verify discipline but doesn't yet have a designed multi-
+filter discriminator for the trust layer.
+
+## Target
+
+Multi-learned-bloom-filter intersection with caustic-geometry-shaped
+agreement region. 3 components (per Kestrel-v2 Meaning 3 + 1 composition):
+
+1. **Filter A** — sharp on code provenance signals (signed-from-trusted-
+   publisher vs signed-from-unknown-publisher)
+2. **Filter B** — sharp on behavioral signals (suspicious syscall
+   patterns; runtime resource access)
+3. **Filter C** — sharp on structural signals (lexical malware-family
+   match; dependency-graph similarity)
+
+Composition: bitwise AND of membership-test results. The "caustic" is
+the region in combined feature space where all 3 filters agree the
+code is trustworthy. Each filter has its own FP rate; intersection FP
+rate is product of individual rates (assuming independence).
+
+Composes with trust-then-verify discipline: bloom filters are the
+**trust** layer (cheap fast checks handling most inputs); heavier
+verification (signature check, runtime behavior analysis, static
+analysis) invoked only when caustic indicates closer attention needed.
+
+## Two scope phases
+
+### Phase 1 — operational (implementable now)
+
+Multi-learned-bloom-filter intersection per Kestrel-v2 Meaning 3 + 1
+composition:
+
+- Each filter is a learned bloom filter (Kraska et al. style) trained
+  on actual input distribution
+- Hash functions + bit array sizes per learned-bloom-filter standard
+- Composition is bitwise AND
+- Caustic-engineering insight applies as **design guidance**: optimize
+  individual filter shapes for the geometry of their intersection
+  rather than for individual accuracy
+
+Implementation: TS substrate using one of the established bloom-filter
+libraries; learned-bloom-filter wrapper composing standard hash functions
+with a learned-bias model per filter; composition module performing
+bitwise AND.
+
+### Phase 2 — research-direction (full caustic engineering)
+
+Per Kestrel-v2 ferry Meaning 2 + Aaron's full-caustic-engineering
+extension (Matt Ferraro reference + Disney Research / ETH Zurich
+academic lineage):
+
+- Inverse-design discipline transferred from optics → bloom filters
+- Optimal transport (Brenier theorem + Cédric Villani Fields-medal
+  transport theory) for filter design — map input distribution onto
+  desired discrimination distribution
+- Continuous relaxation of discrete bloom filter response function
+- Target specification language for discrimination shapes (heatmap OR
+  hierarchical region specification)
+- Manufacturing-equivalent step translating optimized abstract design
+  → concrete bloom filter implementation
+
+**Substrate-honest scope** (per Kestrel-v2): "research-direction-
+flavored work, well beyond the runbook gesture register but probably
+not yet ready for backlog promotion as a tractable engineering
+project. The mathematical work to formalize bloom-filter caustic
+engineering is significant — probably weeks to months of focused
+effort to produce a working prototype."
+
+### Phase 3 — nearer-term reachable piece
+
+Per Kestrel-v2: "a literature-review level deep dive into how inverse
+design has been transferred between domains, with particular attention
+to discrete-versus-continuous optimization. That gives you the
+methodology before committing to the bloom-filter-specific work."
+
+References to consume:
+
+- Matt Ferraro's caustics-engineering post (https://mattferraro.dev/posts/caustics-engineering)
+- Disney Research caustic-engineering papers (Yue et al.,
+  Schwartzburg et al.)
+- ETH Zurich academic lineage
+- Kraska et al. learned bloom filters
+- Sigmund's topology optimization (broader inverse-design)
+- Brenier theorem on optimal transport + Villani transport theory
+
+## Acceptance
+
+Phased:
+
+- **Phase 1 acceptance**: working prototype of 3-filter caustic
+  composition; benchmark FP rate against single-high-accuracy filter
+  + against majority-vote composition; demonstrate caustic-design
+  advantage in at least one specific input-distribution scenario
+- **Phase 2 acceptance**: mathematical formulation of inverse design
+  for bloom filter composition; working solver; at least one
+  end-to-end example translating from target specification →
+  optimized filter design → deployable implementation
+- **Phase 3 acceptance**: literature review document at
+  `docs/research/` cataloging inverse-design transfers across domains
+  with explicit attention to discrete-vs-continuous optimization;
+  identifies the methodological precedents that make the bloom-filter
+  transfer tractable
+
+## Composes with
+
+- B-0664 NCI HC-8 floor (the discriminator IS the technical substrate
+  for "no unauthorized RCE against another node's memory substrate"
+  per Lior-website NCI-as-cyberattack-prevention naming PR #5342)
+- B-0648 cross-substrate-triangulation discipline (multiple bloom
+  filters operating as composing N-of-M oracles; this row IS that
+  discipline at remote-code-discrimination scope)
+- `.claude/rules/non-coercion-invariant.md` HC-8 floor — discriminator
+  serves the floor
+- `.claude/rules/algo-wink-failure-mode.md` — algo-wink IS one class
+  of unauthorized-RPC pattern the discriminator catches
+- `.claude/rules/glass-halo-bidirectional.md` — audit-mechanism
+  composes at output side (which inputs got heavier verification?)
+- `.claude/rules/substrate-smoothness-as-load-bearing-property.md`
+  (PR #5357) — coupled smoothness constraint applies (target
+  discrimination shape can't have arbitrarily sharp boundaries;
+  bloom filter design can't have arbitrarily sharp transitions;
+  composition pipeline preserves smoothness end-to-end)
+- B-0822 worry-as-opposite-bloom-filter substrate (PR #5310) — the
+  mental model this row's substrate-engineering work builds on
+- B-0823 cognition-as-distributed-systems META-claim (PRs #5325 +
+  #5327) — speech-as-RPC framing this discriminator slots into
+- B-0833 (interactive-login vs baked-in-keys CI-test tension) — bloom
+  filter discriminators serve the trust layer of the broader auth
+  + verification pipeline
+- PR #5356 Kestrel-v2 ferry — substrate origin (full verbatim
+  preservation including operator's cat-caustic image as physical
+  existence proof of computational inverse-design extending into
+  physical artifacts)
+- PR #5357 substrate-smoothness rule — coupled-smoothness constraint
+  reference
+- F# fork for AI safety substrate — raw-math interaction (per
+  Lior-website observation) is the cleanest substrate for inverse-
+  design machinery
+
+## Substrate-honest framing
+
+P2 priority: NOT immediately tractable as single-implementation work.
+Phase 1 (operational 3-filter intersection) is bounded enough for
+single-PR landing; Phases 2 + 3 are research-direction work spanning
+weeks-to-months.
+
+Per Kestrel-v2's epistemic checkpoint: "the engineering ROI versus
+simpler multi-filter approaches isn't obvious until you've done enough
+of the formalization to know what the design space actually looks like."
+
+This row creates the substrate anchor; future contributors (human OR
+AI) decompose phases independently when scope tightens. Operator's
+2026-05-26 framing on backlog-row landing: "you can always commit
+backlog rows immediatly they get decoposed later."
+
+## Origin
+
+Aaron-forwarded Kestrel-v2 conversation 2026-05-26 (preserved verbatim
+via PR #5356). 3 substantively-new contributions in that ferry:
+
+1. Caustic-engineered bloom filter discriminators (THIS row)
+2. Substrate-smoothness-as-load-bearing-property (landed as rule via
+   PR #5357)
+3. Gesture-to-spec collaborative workflow as pattern (informal substrate)
+
+The conversation also produced Aaron's "society committee" naming
+(Turn 13 via PR #5359 Kestrel-v3 ferry) + universal-cognitive-substrate-
+limits observation (Turn 14) which both compose at meta-scope.
+
+Filed P2 per operator's "backlog rows land immediately, decompose
+later" discipline 2026-05-26.