Compendium of the Harbor coordination program's executed results R1–R17 — information floor, split-digest theorem, derived regret head, digest-zoom frontier, hypervisor-as-supervisory-control, sheaf equivocation verdict (with the consistency-radius theorem), inspection tower, work-unit machine, sealed-room noninterference, ε-ledger conservation, canary/SPRT detection power, no-mint reputation inheritance, engine substitution and resurrection soundness, costly-escalation tuning band, Erlang-C specialization boundary, context paging under a corrupted pin oracle, and the tractable deontic-conflict fragment — with exact statements, verified-vs-internal numbers, honest boundaries, paper mapping, and the tacit lessons behind them. Use when extending, citing, presenting, implementing, stress-testing, or planning follow-on work (papers, proofs, product claims) on any of these results. NOT for general math/CS reference, Port Daddy product features outside these results, or writing style (see harbor-exposition).
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Der Befehl bleibt in einer Zeile. Scrollen Sie horizontal, um ihn vor dem Kopieren vollständig zu prüfen.
Sie bevorzugen eine lokale Kopie? Laden Sie die Dateien herunter, die SkillsMP derzeit vorliegen.
Datei-Explorer
23 Dateien
SKILL.md wird angezeigt
SKILL.md
Quellanweisungen · Schreibgeschützte Vorschau
license
Apache-2.0
name
harbor-results
description
Compendium of the Harbor coordination program's executed results R1–R17 — information floor, split-digest theorem, derived regret head, digest-zoom frontier, hypervisor-as-supervisory-control, sheaf equivocation verdict (with the consistency-radius theorem), inspection tower, work-unit machine, sealed-room noninterference, ε-ledger conservation, canary/SPRT detection power, no-mint reputation inheritance, engine substitution and resurrection soundness, costly-escalation tuning band, Erlang-C specialization boundary, context paging under a corrupted pin oracle, and the tractable deontic-conflict fragment — with exact statements, verified-vs-internal numbers, honest boundaries, paper mapping, and the tacit lessons behind them. Use when extending, citing, presenting, implementing, stress-testing, or planning follow-on work (papers, proofs, product claims) on any of these results. NOT for general math/CS reference, Port Daddy product features outside these results, or writing style (see harbor-exposition).
allowed-tools
Read,Write,Edit
metadata
{"category":"Research & Knowledge","tags":["harbor","coordination","theorems","results","multi-agent","oversight"],"version":"1.0.0","pairs-with":[{"skill":"harbor-exposition","reason":"The presentation style these results ship in"},{"skill":"falsification-first","reason":"The discipline that produced and validated these results"}]}
Harbor Results: The Executed Corpus (R1–R17)
Seventeen results, executed and reproducible (program seed 20260816), each with a one-breath statement here and full depth in the references.
When to Use
✅ Use for: citing or restating any R1–R17 result precisely; choosing which result underwrites a product claim; extending a result or checking whether a "new" idea is already covered or already refuted; packaging results into papers; onboarding a collaborator to the program.
❌ NOT for: general information theory / game theory / security reference unrelated to these nine; Port Daddy engineering outside the results (kernel schemas, CLI); exposition mechanics (harbor-exposition); running the validation discipline itself (falsification-first).
The Seventeen, in one breath each
R1 — Information floor. No digest below log₂C(N,k) − log₂C(m,k) bits can guarantee catching all k critical items among N while opening m; survived falsification (0/16 violations).
R2 — Split-digest. One scalar summary serves two readers iff their orders are comonotone; successor-agent and operator provably are not; joint zero-miss floors are super-additive (≈2.13×, not 2×).
R3 — Derived regret head. stakes × irreversibility × anomaly = exact expected unrecoverable loss iff anomaly is a calibrated posterior; optimal surfacing is the likelihood-ratio test; reputation enters only through the posterior.
R4 — Digest-zoom frontier. Two-constraint rate-distortion R(δ,f) with the old floor as its zero-miss corner (R(0)=H(p), e.g. 0.286 bits at p=0.05); adaptive zoom needs ≈k·log₂(F/k) opens vs F flat — only in the sparse-flagged regime.
R5 — Hypervisor = supervisory control. A policy is preventable (regimentable) iff controllable w.r.t. the uncontrollable event set (Ramadge–Wonham K̄Σᵤ∩L̄⊆K̄); "no egress after reading a secret" IS regimentable — gate the channel, never the token.
R6 — Sheaf verdict. Cohomology detects equivocation beyond pairwise comparison iff the missing edge lies on a cycle (cocycle sum ≠ 0); on a cut edge, never. Conditional commit with that exact scope.
R7 — Inspection tower. ρ* = G/(dB); sealed sampling from C cliques makes bribery uneconomical (profitable iff G_k > C·B), corruption decays (1−ρd)^k; reputation is amortized verification (Θ(log T) or O(1) audit spend).
R8 — Work-unit machine. Six safety invariants hold in all 536 reachable states; all five guards proven load-bearing by mutation (shortest crimes: 4, 2, 4, 1, 7 steps).
R9 — Sealed-room noninterference. Erin's view identical across equal-parity secrets under every interleaving to depth 7; schedule secret-independence checked separately; leaky-gate and bypass mutations caught with witnesses.
R10 — ε-ledger conservation. The release ledger's atomic append+add conserves σ = Σ_Λ εᵢ ≤ εmax under every concurrent interleaving (single-writer serialization); sequential/advanced DP composition gives the sum its meaning; recorded spend only — mediation is R5's job.
R11 — Canary power + SPRT. k smuggled canaries are caught w.p. 1−β^k; uniform planting gives the hypergeometric operating curve Pr(detect)=f(leak size); Wald's SPRT turns leak intensity into expected time-to-detection with errors at or below target.
R12 — No-mint inheritance. Fork priors conserve iff split = transfer (the source is debited): total live creditable reputation never exceeds total witnessed; copy-full inheritance mints 8.2× from one episode — the quorum attack the invariant blocks.
R13 — Engine substitution. Unattested, swapping a cheap engine behind a good record pays Δc at every price (Akerlof unraveling inside one identity, threshold μ*); engine attestation flips the incentive to Δc−Δθ — the planner's rule at zero audit stake; resurrection across providers is sanction-sound iff all three migration clauses hold, with a shortest attack per dropped clause.
R14 — Escalation tuning band. Monotone benefit and validation give a unique escalation threshold u*(δ) (Spence-style separating equilibrium; u* = δw/(1+δw) uniform-linear); the feasible debit set between alarm fatigue and distress silencing is an interval that can be EMPTY — then the fix is capacity or evidence quality, not tuning.
R15 — Specialization boundary. A sole specialist beats the pool iff μs/μg ≥ g(ρ,c) = cρ + c(1−ρ)/(c(1−ρ)+C(c,ρ)) — the whitepaper's proposed threshold falsified in both directions; with mortality ξ and succession rate η, pooling dominates at EVERY skill premium once ξ/η > D* — the price of the succession rule.
R16 — Context paging. Landlord's k/(k−h+1) competitiveness imports to variable-size context spans; a pin oracle corrupted at rate φ costs only an ADDITIVE φN·c_refetch with repair-on-touch — and exactly one blind-trusted corruption costs Θ(N); φ is R1's forgeable-feature budget, shared.
R17 — Deontic conflict fragment. Inside Horn + ground intervals + difference constraints, commitment-conflict detection is polynomial with a witness per conflict; one expressive step out (disjunctive obligations) is NP-complete — the fragment boundary is the exact price of proposal-time checking.
R6 theorem (CR). The completion residual r is a certified lower bound on any injected lie (equality at the completion), localizes to cycles through the equivocator, and computes in Õ(|E|·L); single-equivocator only — coalitions on a cycle can cancel to r = 0.
Routing
flowchart TD
Q[Query about the program] --> A{Need the precise statement,\nnumbers, boundary, or paper home?}
A -->|yes| C[references/results-compendium.md\nload the matching R-section]
Q --> B{Extending, re-deriving, implementing,\nor building a similar experiment?}
B -->|yes| L[references/l3-tacit-lessons.md\nread the matching lesson FIRST\nto avoid re-hitting known traps]
Q --> P{Planning papers or next wave?}
P -->|yes| C2[compendium end sections:\npaper map + remaining roadmap]
Anti-Patterns
The Unpriced Side Channel
Novice: "My simulation's oracle knows the answer; I only charge bits for output resolution."
Expert: Information bounds constrain what a channel carries; charge bits for identification, not tie-breaking. A sim that smuggles identity past the meter will appear to beat the floor (R1's 8/14 spurious violations before the fix).
Detection: Any experiment "beating" an information-theoretic bound.
Cohomology of the Sheaf, Not the Data
Novice: "dim H¹ of the sheaf detects the inconsistency."
Expert: dim H¹ of the abstract sheaf is a property of restriction maps, data-independent. Equivocation is an obstruction of the observed assignment (is the disagreement cochain a coboundary?). And the value-add exists only where the missing edge lies on a cycle.
Timeline: Both wrong turns hit and documented 2026-08; they located the boundary.
Detection: Computing H¹ without the observed data anywhere in the computation; testing on a bridge/tree topology.
Zoom Everywhere
Novice: "Group testing always beats flat inspection."
Expert: It pays ≈F/(k·log₂(F/k)) only when positives are sparse in the flagged set; dense flags make splitting overhead dominate. Zoom is for the miss-averse over-flagging regime — exactly where a safe digest operates.
Detection: Claiming zoom savings without stating flagged-set density.
References
references/results-compendium.md — Load when citing, restating, packaging, or checking scope of any R1–R17: boxed statements, [verified] vs [internal] numbers, applications, boundaries, paper map, reproducibility pointers, and the treatise-corrections consensus.
references/l3-tacit-lessons.md — Load BEFORE extending or re-implementing anything here, or designing a related experiment: the hard-won lessons (bugs hit, boundaries found, conventions that change answers) with transferable rules.
Scripts (regenerate every [internal] number)
Self-contained; deps: numpy, scipy, matplotlib, networkx; seed 20260816 fixed inside each. Run from the skill root when re-verifying a number before citing it, or after modifying any claim these underwrite.
python3 scripts/b6_probation.py — B6: front-loaded probation dominance — 0 dominating schedules in 4,000 random instances (76,000 schedules tested), matching the closed-form exchange argument; exit nonzero on violation.
python3 scripts/zoom_bound_check.py — Paper 1's zoom theorem verified: Q <= 2k*ceil(log2(F/k)) + 4k on every tested instance (random + adversarial placements), exact tightness at F=4096,k=32 (511), bound dominates the measured ~163 opens at the b1 point; exit nonzero on violation.
python3 scripts/b7_escalation_band.py — R14: costly-escalation threshold equilibrium (single crossing, closed forms u*=δw/(1+δw) and Lambert-W) + the two-sided debit tuning band tied to R3's constants; empty-band regime exhibited; zero-debit and excessive-debit mutants caught; exit nonzero on violation.
python3 scripts/b4_deontic_fragment.py — R17: polynomial witness-producing conflict checker for the Horn+interval+difference deontic fragment (3000-instance oracle sweep, 0 disagreements) + the 3-SAT NP-completeness reduction one expressive step outside it, verified both directions on 16/16 instances; Horn-propagation mutant misses 100/885 conflicts; exit nonzero on violation.
python3 scripts/b5_engine_substitution.py — R13: Akerlof unraveling inside one identity (swap gain price-independent, μ* threshold, death spirals), the IC flip under engine attestation (swap gain Δc−Δθ, audit stake → 0), and resurrection soundness (747-state migration machine, 0 Def-III.6.1 violations intact; 4 mutants caught with shortest crimes 1,2,2,2); exit nonzero on violation.
python3 scripts/sheaf_consistency_radius.py — R6 theorem (W8 tail): CR-1/2/3 verified — soundness (r = exact minimum lie, 0/600 violations; closed form r = |s|·√(1−R_eff), the R6 1.2247 number derived), localization (328/328 noise-free, electrical-flow support), complexity (per-coordinate Laplacian decomposition, CG slope 0.65 vs dense 1.88); coalition-cancellation boundary exhibited; D2 mutant caught; exit nonzero on violation.
python3 scripts/b9_context_paging.py — R16: Landlord k/(k−h+1) import verified vs exact OPT (122 pairs, cycles at equality) + linear φ-degradation under a corrupted pin oracle (additive φN·c_refetch bound, 0 violations incl. adaptive corruption; slope 1.9% of ceiling, R²=0.994); blind pin-trust mutant catastrophic (Θ(N) staleness from one corruption); exit nonzero on violation.
python3 scripts/b8_specialization.py — R15: exact Erlang-C specialization boundary g(ρ,c) (falsifies the proposed 1+(c−1)ρ/(1−ρ) threshold in both directions, crossing at ρ=(3−√5)/2) + succession-price theorem W_bd closed form, D* = ηK/(1−ηK); DES/matrix-geometric/CTMC cross-checks, 60-instance sweep, breakdown-blind mutant caught; exit nonzero on violation.
