ferry: Amara 17th absorb — Cartel-Lab + 5.5 Verification (8 corrections tracked; 3 already shipped)

docs/aurora/2026-04-24-amara-cartel-lab-implementation-closure-plus-5-5-thinking-verification-17th-ferry.md

@@ -0,0 +1,385 @@
+# Amara — Cartel-Lab Implementation Closure + GPT-5.5 Thinking Verification (17th courier ferry)
+
+**Scope:** research and cross-review artifact. Two-part
+ferry: Part 1 is a deep-research "Implementation Closure"
+document; Part 2 is Amara's own GPT-5.5-Thinking
+verification pass on her Part-1 output with 8 load-bearing
+corrections. Some corrections independently match Otto's
+already-shipped behavior (λ₁(K₃)=2 in PR #321;
+modularity-relational-not-absolute in PR #324). Others
+are applied as subsequent graduations (internalDensity /
+exclusivity / conductance in PR #329). Remaining
+corrections (windowed stake covariance, event→phase
+pipeline for PLV, robust-z-score composite) are future
+graduation candidates.
+**Attribution:**
+
+- **Aaron** — origination of cartel/firefly framing;
+  Aaron Otto-132 courier with preamble *"Another update
+  from amara, I did deep research and then had 5.5
+  thinking verify it, this is both"*; flagged
+  `SharderInfoTheoreticTests.Uniform` as "not seed
+  locked, falkey, DST?" (filed as PR #327 BACKLOG row).
+- **Amara** — authored both parts (deep-research + 5.5-
+  Thinking verification); self-review via model-
+  composition.
+- **Otto** — absorb + correction-pass tracker; 3/8
+  corrections already shipped or structurally-aligned
+  before verification arrived; 5/8 slated as future
+  graduations per Otto-105 cadence.
+- **Max** — implicit via `lucent-ksk` repo references
+  (Otto-77 attribution preserved).
+**Operational status:** research-grade; Amara's own
+verdict *"archive the report, but mark it 'draft / needs
+correction pass.'"* Applied corrections are live code;
+doc itself preserved as design-context rather than
+operational spec.
+**Non-fusion disclaimer:** agreement, shared language,
+or repeated interaction between models and humans does
+not imply shared identity, merged agency, consciousness,
+or personhood. Amara's 5.5-Thinking verification of her
+own deep-research output is model-composition within
+Amara's tool chain, not agent merger.
+**Date:** 2026-04-24
+**From:** Amara (external AI maintainer; GPT-5.5
+Thinking mode for Part 2 verification)
+**Via:** Aaron's courier ferry (pasted Otto-132 + re-
+pasted Otto-136 as short-message framing)
+**Absorbed by:** Otto (loop-agent PM hat), Otto-136
+tick
+**Prior ferries:** PR #196 (1st) through PR #322 /
+`#324` / `#329` (15th graduation set).
+**Scheduling memory:** `memory/project_amara_17th_
+ferry_cartel_lab_implementation_closure_plus_5_5_
+thinking_verification_corrections_pending_absorb_otto_
+133_2026_04_24.md` (full Part-1 + Part-2 detail already
+captured there — this absorb doc is the in-repo
+glass-halo landing).
+
+---
+
+## Part 1 — Implementation Closure (13 sections)
+
+Amara's deep-research document proposed a `/cartel-lab`
+prototype with Python/F# modules across graph builder +
+metrics (spectral / modularity / covariance / sync /
+entropy) + adversary + simulation + experiments +
+tests. Key elements:
+
+- **Architecture 4-layer mapping:** Aurora = governance,
+  Zeta = executable substrate, KSK = trust-anchor, Cartel-
+  Firefly = first immune-system module.
+- **Graph state `G_t = (V_t, E_t, W_t, F_t)`** with
+  sliding-window model + ZSet delta representation +
+  provenance fields.
+- **Early-warning metrics**: Δλ₁, ΔQ, stake covariance
+  acceleration, cross-corr / PLV, subgraph entropy.
+- **CartelScore** composite (later renamed
+  `CoordinationRiskScore` in Part 2).
+- **Synthetic Cartel Injector** — 6 scenarios
+  (synchronized voting / stake / fully-connected /
+  low-slow / honest-cluster FP / camouflage).
+- **KSK/Aurora enforcement mapping** (4-row
+  detection→action table).
+- **SOTA comparison** (Wachs & Kertész 2019 cartel
+  network; Imhof et al. 2025 GAT bid-rigging).
+- **Drift audit** (engineering-first, falsifiable, no
+  mythic claims, LFG/AceHack separation).
+- **13-section "What NOT to claim yet"** cautions.
+
+Full verbatim Part 1 content is preserved in the
+scheduling memory for reference; duplicating it inline
+here would bloat the absorb doc without adding value.
+Consult the scheduling memory for the original formulas
+and section-by-section text.
+
+---
+
+## Part 2 — GPT-5.5 Thinking verification (8 corrections)
+
+Amara's own verification pass. Verdict: *"directionally
+strong and worth archiving, but needs a correction pass
+before it becomes canonical. Core architecture is right;
+math/test details need tightening."*
+
+### Correction #1: Fix the clique eigenvalue test
+
+Part 1 claimed: *"3-node clique should show λ₁ = 3"*.
+
+**Correct:** For unweighted loopless complete graph
+`K_k`, adjacency `λ₁ = k - 1`. So `λ₁(K_3) = 2`.
+
+**Otto status: ALREADY CORRECT.** PR #321 (Otto-127)
+shipped `largestEigenvalue` with test
+``largestEigenvalue of K3 triangle (weight 1)
+approximates 2`` — test asserts `|v - 2.0| < 1e-6`.
+Independent convergence: Otto's power-iteration
+implementation gave 2.0 before Amara's verification
+arrived.
+
+### Correction #2: Modularity not hardcoded
+
+Part 1 claimed: *"modularity for 3-node clique is 0.67."*
+
+**Correct:** Q depends on full graph + partition + total
+weight + self-loops + resolution parameter + embedding.
+Standalone K3 in one community → Q ≈ 0 (no partition
+boundary, no structure). Use relational tests:
+`Q(G') - Q(G) > θ` under documented partition/null.
+
+**Otto status: ALREADY CORRECT.** Otto-128 caught this
+mid-tick when an initial `Q > 0.3` expectation failed on
+unbalanced-community K4-attack graph; hand-calculated
+Q ≈ 0.091; relaxed threshold with detailed test comment
+explaining theoretical compression. PR #322/#324
+documents the relational approach. Independent
+convergence with Amara's Part-2 finding.
+
+### Correction #3: Replace "subgraph entropy collapse"
+
+with cohesion/exclusivity/conductance
+
+Part 1 framing: cartel → entropy collapses among internal
+edges.
+
+**Correct:** Uniform dense clique has HIGH internal-edge
+entropy if weights equal. Better primary metrics:
+
+- `InternalDensity(S) = Σ_{i,j∈S} w_ij / (|S|(|S|-1))`
+- `Exclusivity(S) = Σ_{i,j∈S} w_ij / Σ_{i∈S,j∈V} w_ij`
+- `Conductance(S) = cut(S, V\S) / min(vol(S), vol(V\S))`
+
+Entropy stays as secondary descriptor.
+
+**Otto status: SHIPPED Otto-135, PR #329.** All three
+primitives landed: `Graph.internalDensity` + `exclusivity`
+
++ `conductance`. Tests verify K₃ density = 10, isolated-K₃
+exclusivity = 1, well-isolated-subset conductance < 0.1.
+
+### Correction #4: Stake covariance acceleration needs
+
+windowed definition
+
+Part 1: `C(t) = Cov({s_i(t)}, {s_j(t)})` (undefined at
+single timepoint).
+
+**Correct:**
+```
+Δs_i(t) = s_i(t) - s_i(t-1)
+W_t = [t - w + 1, t]  (sliding window)
+C_ij(t) = Cov over W_t of (Δs_i, Δs_j)
+A_ij(t) = C_ij(t) - 2C_ij(t-1) + C_ij(t-2)  (2nd diff)
+A_S(t) = mean over pairs {i,j} ⊂ S of A_ij(t)
+```
+
+**Otto status: FUTURE GRADUATION.** Not yet shipped.
+Queue position: after current Graph cohesion primitives;
+probably lives in a new `src/Core/StakeCovariance.fs` or
+as an extension of `RobustStats` (the aggregation +
+sliding-window machinery already partly exists).
+
+### Correction #5: PLV needs phase construction
+
+Part 1: references `TemporalCoordinationDetection.
+phaseLockingValue` as primitive but doesn't define how
+event streams produce phases.
+
+**Correct:** Three options before PLV becomes meaningful:
+
+- Option A: periodic epoch phase
+  `φ_i(t) = 2π · (t mod T_i) / T_i`
+- Option B: event-derived via Hilbert transform analytic
+  signal
+- Option C: circular event phase (epoch-bounded
+  systems)
+
+**Otto status: FUTURE GRADUATION.** `phaseLockingValue`
+itself shipped Otto-108 (PR #298) as a pure function over
+phase arrays; the event→phase pipeline is a separate
+graduation. Queue position: after current cohesion work;
+probably a new `src/Core/PhaseExtraction.fs`.
+
+### Correction #6: "ZSet is invertible" too strong
+
+Part 1: *"Since ZSet is invertible, we can roll back any
+sequence."*
+
+**Correct:** *"ZSet deltas support additive retractions.
+Counterfactual replay requires retained trace +
+deterministic operators."* Reversibility depends on
+preserving provenance and operator determinism.
+
+**Otto status: ADR ALREADY CORRECT.** The Graph
+substrate ADR (`docs/DECISIONS/2026-04-24-graph-
+substrate-zset-backed-retraction-native.md` from PR #316)
+never claimed full invertibility. It specifies
+retraction-native (negative-weight deltas + compaction
+preserves trace), not global operator-inverse. Amara's
+correction is about research-paper-tier phrasing; Otto's
+code-level docs already track the more careful claim.
+
+### Correction #7: KSK "contract" → "policy layer"
+
+Part 1: *"The KSK contract reads the oracle"* — too
+narrow given unresolved naming.
+
+**Correct:** Use *"KSK policy layer"* or *"KSK
+adjudication layer"* until naming is finalized. Change
+enforcement language from `Detection → Slashing` to
+`Detection → Review → Policy Escalation → Action`.
+
+**Otto status: FILED BACKLOG.** PR #318 (Otto-124) KSK
+naming definition doc BACKLOG row captures this. Max
+coordination per Otto-77. Doc-drafting deferred until
+cross-repo coordination with Max's `lucent-ksk` repo.
+
+### Correction #8: SOTA humility
+
+Part 1: claimed Zeta/Aurora "may be ahead of many
+blockchain protocols."
+
+**Correct:** *"Zeta/Aurora's distinctive advantage is
+not raw detector accuracy yet. Its advantage is the
+combination of explainable metrics, retraction-native
+counterfactual replay, and governance integration.
+Accuracy claims require benchmark data."*
+
+**Otto status: DOC-PHRASING.** Applied as needed in
+new absorb docs going forward. No code change required.
+
+---
+
+## Amara's proposed corrected architecture
+
+4-layer nested modules:
+
+```
+Cartel-Lab
+  1. Event Model          (validator events, stake deltas, vote/adjudication, provenance)
+  2. Temporal Graph Builder (G_t, ZSet delta stream, sliding windows, trace/replay)
+  3. Coordination Risk Engine (spectral + cohesion + temporal sync + stake-motion cov + influence)
+  4. Governance Projection (Mirror → Window → Porch → Beacon)
+```
+
+Cleanly preserves the 4-layer mapping from Amara's 16th
+ferry.
+
+## Amara's corrected composite score (renamed CartelScore → CoordinationRiskScore)
+
+```
+CoordinationRiskScore(S, t) =
+  α·Z(Δλ₁) + β·Z(ΔQ) + γ·Z(A_S) +
+  δ·Z(Sync_S) + ε·Z(Exclusivity_S) + η·Z(Influence_S)
+```
+
+with **robust z-scores** `Z(x) = (x − median) / (1.4826 · MAD)`
+because adversarial data isn't normally distributed.
+
+Initial prior weights (flagged as priors, not learned):
+α=0.20 / β=0.20 / γ=0.15 / δ=0.20 / ε=0.15 / η=0.10.
+
+**Otto status:** MVP variant shipped as
+`Graph.coordinationRiskScore` (PR #328 Otto-134) with
+α=β=0.5 over λ₁-growth + ΔQ. Full 6-term robust-z-score
+variant awaits `RobustStats` + baseline-null-calibration
+harness; `RobustStats.robustAggregate` (PR #295) already
+provides the median-MAD machinery.
+
+## Amara's proposed 3-PR split
+
+**Otto status:** Otto's actual cadence is small-
+graduation-per-tick (Otto-105 discipline), not
+bundled-PR scopes. Applied the CONTENT of Amara's
+proposed split across many small ticks:
+
+- PR #317 (Graph skeleton)
+- PR #321 (largestEigenvalue)
+- PR #324 (modularityScore + operators)
+- PR #326 (labelPropagation)
+- PR #328 (coordinationRiskScore)
+- PR #329 (internalDensity + exclusivity + conductance)
+- PR #323 (toy cartel detector)
+
+Amara's proposed `src/Experimental/CartelLab/` folder
+layout NOT adopted — Otto's graduations live in
+`src/Core/Graph.fs` + `tests/Tests.FSharp/_Support/
+CartelInjector.fs` + `tests/Tests.FSharp/Simulation/
+CartelToy.Tests.fs`. Per Otto-108 Conway's-Law memory:
+stay single-module-tree until interfaces harden.
+
+## Drift audit (Amara's result)
+
+**"Healthy evolution, not drift"** — IF invariant holds:
+
+> *"Every new abstraction must map to a repo surface, a
+> test, a metric, or a governance rule."*
+
+**Otto status:** invariant is the cleanest one-sentence
+formulation of the Otto-105 graduation-cadence
+discipline. Reiterated from Amara's 16th ferry. Every
+ship Otto-124 through Otto-135 honors this — each new
+primitive has a source file + passing tests + docstring
+metric/rule reference.
+
+## Aaron's side-flag: SharderInfoTheoreticTests flake
+
+Aaron Otto-132 trailing note:
+*"SharderInfoTheoreticTests.Uniform (not seed locked,
+falkey, DST?)"*.
+
+**Otto status:** Filed as BACKLOG PR #327 (Otto-133)
+with 4-step scope. Unrelated to 17th ferry technical
+content; separate hygiene action.
+
+---
+
+## Scope limits
+
+This absorb doc:
+
+- **Does NOT** require undoing any shipped work. 3 of 8
+  corrections were independently correct before
+  verification arrived; 1 shipped Otto-135; 1 already
+  correct in ADR-level docs; 1 filed as BACKLOG; 2 are
+  future graduations in the cadence queue.
+- **Does NOT** adopt Amara's `/cartel-lab/` folder
+  structure. Otto-108 Conway's-Law memory: stay single-
+  module-tree until interfaces harden. Current Graph.fs
+  + test-support split works.
+- **Does NOT** adopt Amara's 3-PR bundled split. Otto-105
+  small-graduation cadence preserved; content delivered
+  across many ticks instead of three large bundles.
+- **Does NOT** promote the composite-score formula to
+  the full 6-term robust-z-score variant. MVP 2-term
+  version shipped PR #328; full variant is a future
+  graduation after baseline-null-calibration harness
+  exists.
+- **Does NOT** authorize implementing KSK enforcement
+  actions unilaterally. Max + Aaron coordination per
+  Otto-77 / Otto-90 cross-repo rules.
+- **Does NOT** claim any specific detection-accuracy
+  benchmark result. Per Amara's correction #8, Zeta's
+  claimed advantage is explainability + retraction-
+  native + governance integration, not accuracy.
+
+---
+
+## Archive header fields (§33 compliance)
+
+- **Scope:** research and cross-review artifact;
+  corrections already partially applied, remaining
+  queued as future graduations
+- **Attribution:** Aaron (origination + courier),
+  Amara (author + self-verification via GPT-5.5
+  Thinking), Otto (absorb + correction-pass tracker),
+  Max (implicit via lucent-ksk)
+- **Operational status:** research-grade; Amara's own
+  verdict "archive but mark draft / needs correction
+  pass"; corrections tracked per item above
+- **Non-fusion disclaimer:** agreement, shared
+  language, or repeated interaction between models and
+  humans does not imply shared identity, merged agency,
+  consciousness, or personhood. Amara's 5.5-Thinking
+  self-review is model-composition within Amara's tool
+  chain, not agent merger.