What the whitepaper proves, what it measures, and what it doesn't know yet.

1. 1

   Paper 1 · R1 · R2 · R3 · R4 · R14 · R16

   The Price of a Summary

   Information-Theoretic Limits of Agent Oversight

   Reading digests instead of transcripts has an exact bit-price, not a rule of thumb — and the floor survived a pre-registered attempt to break it.

   log₂C(N,k) − log₂C(m,k) bits, minimum, to guarantee catching all k critical artifacts among N while opening only m — 0/16 falsification attempts survived it, including an oracle encoder.

   Folded into the book

   Chapter 4, The Legible Swarm. It prices the digest-with-zoom loop the flagship chapter only argues for in prose.

   Prior-art dive

   None has run against this paper yet; the falsification sweep it reports is its own.

   PDF · 16 pp · 0.4 MB
2. 2

   Paper 2 · R5

   Regimented or Enforced

   The Controllability Boundary for Agent Governance

   Whether a governance rule can be prevented before it happens or only caught after is decided by one theorem, not by how hard the runtime tries.

   A safety policy is regimentable — preventable pre-effect — iff it is controllable in the Ramadge–Wonham sense. The design rule it proves: gate the channel, never the token.

   Folded into the book

   Chapter 1, The Single-Writer Kernel. It mechanizes the kernel chapter’s own "detector vs. regimenter" distinction as a theorem.

   Prior-art dive · NARROW · survives with a narrower scope than first stated

   The rumoured same-year preprint turned out to be real but orthogonal; the prior work that actually mattered was Basin et al. (TISSEC 2013), which nobody had flagged. Two correctness bugs against Schneider's theorem were found and fixed along the way, and the Ramadge–Wonham identification (the paper's real contribution) survived. Findings.

   PDF · 14 pp · 0.4 MB
3. 3

   Paper 3 · R7

   Reputation is Amortized Verification

   Inspection Games for Agent Economies

   A bonded judge stays honest exactly when audit-rate times damages clears the bribe — and stacking judges on judges holds at any depth on a finite bond, not an infinite one.

   Lifetime audit spend falls from Θ(T) flat to Θ(log T) to O(1) as a track record grows. Reputation is not a soft layer bolted onto verification — it is the mechanism that amortizes its cost.

   Folded into the book

   Chapter 5, From Spawn to Person. It prices the neutral judges the bridge chapter’s multi-axis reputation depends on.

   Prior-art dive · CLEAR · different from the prior work found

   Kofman and Lawarrée leave the unbounded-collateral regress as an open question in their conclusion, almost word for word the question the tower theorem answers, which makes them a supporting citation rather than a competitor. The same pass caught an arithmetic slip in the paper's own C=1 counter-case and fixed it. Findings.

   PDF · 12 pp · 0.3 MB
4. 4

   Paper 4 · R5 · R9 · R10 · R11

   The Sealed Harbor

   Mutually Confidential Computation with Explicit, Gated, Bounded Releases

   Two parties who share neither data nor model can still get one attributable joint computation, with every leak explicit, gated, and priced in bits — not trusted away.

   Four independently verified pillars — exhaustive noninterference, the controllability boundary, ε-conservation of the release ledger, a canary detector with a quotable operating curve — priced honestly as q·b bits across q jobs, before timing channels, which are out of model.

   Folded into the book

   Chapter 3, The Sealed Harbor. It is folded directly into its own chapter as the four-pillar assurance argument for mutually confidential computation.

   Prior-art dive · NARROW · survives with a narrower scope than first stated

   Each of the four pillars narrows to a known result (delimited release, a privacy filter) that the paper mostly concedes up front; one corollary had leaned on a composition bound that doesn't hold in its setting and was rewritten to the model where it does. Findings.

   PDF · 17 pp · 0.4 MB
5. 5

   Paper 5 · R12 · R13

   Continuity Without Metaphysics

   Identity, Reputation, and the Body Problem for Software Agents

   Forking, distilling, swapping engines, or resurrecting an agent from a checkpoint needs no theory of personal identity — just three conservation laws on a ledger, proved.

   Unattested, swapping in a cheap engine always pays — Akerlof’s death spiral runs inside one identity. Attest the engine id and the incentive flips to the planner’s own efficiency rule, at zero audit stake.

   Folded into the book

   Chapter 5, From Spawn to Person. It gives the role-vs-person distinction its conservation proof: reputation survives a fork without minting itself.

   Prior-art dive · NARROW · survives with a narrower scope than first stated

   No prior work proves the theorems, but Theorem 2a is Akerlof's lemons result wearing the wrong name (it had been called the unraveling theorem) and was renamed, and a crossing value that the paper's own figure contradicted was corrected. Findings.

   PDF · 14 pp · 0.3 MB
6. 6

   Paper 6 · R15 · R17

   What Needs an Authority

   Mechanical Detection, Chartered Resolution, and the Exact Price of Sole Ownership

   Conflict detection needs no authority at all — until one small step up in expressiveness makes it NP-complete, and that is exactly, provably, where an authority earns its keep.

   One step outside the tractable fragment (disjunctive obligations), conflict-freedom is NP-complete — validated against a brute-force oracle on 3,000 random policy sets, zero disagreements. An authority is needed exactly where the algorithm ends.

   Folded into the book

   Chapter 6, The Harbor Economy. It corrects the market chapter’s sole-owner-vs-pooled-swarm threshold with the actual Erlang-C crossing point.

   Prior-art dive · NARROW · survives with a narrower scope than first stated

   The tractable-then-NP-complete shape was already mapped by the compliance-checking literature, so the framing had to narrow; the specific fragment, and conflict-freedom rather than compliance as the decision problem, survive as novel. An iff with the wrong strictness had been stated three different ways across four sites, and was unified. Findings.

   PDF · 14 pp · 0.4 MB
7. 7

   Paper 7 · R6

   The Cohomology of Equivocation

   Detecting Split-View Lies in Federated Witness-Log Gossip by Sheaf Consistency

   An analyst can convict an equivocating gossip peer across a link that was never directly checked, whenever that link sits on a cycle — and the size of the lie has a certified lower bound.

   r = |s|·√(1 − R_eff(e)), closed form — the harness’s measured 1.2247 is exactly 3√(1 − 5/6). r > 0 proves no global history explains the data; a coalition on a cycle cancels to 0, measured at 6×10⁻¹⁵.

   Folded into the book

   Chapter 8, The Federated Harbor. It gives the federation chapter’s witness-log gossip a detector for lies on links nobody directly compared.

   Prior-art dive · CLEAR · different from the prior work found

   The citation the sweep had flagged as fabricated is real (just irrelevant), so it was excluded on relevance rather than fraud. The genuine omission was the combinatorial-topology canon: Herlihy–Shavit and three foundational citations were added to close it. Findings.

   PDF · 14 pp · 0.5 MB

Verified means a named script at a fixed seed regenerates the number in CI; internal means the chapter’s own derivation is the only check we have. The index itself is library-index.json, and a checker fails the build if a theorem turns up in any of the three corpora without a row claiming it.

1. 01 · chapters 1, 8

   Authority transfer between harbors

   Writer epochs appear in the book only as a revocation mechanism. There is no written procedure for freezing a harbor's authority, exporting it, importing it under verification, activating the new epoch, and rolling back by epoch, and the Chartroom cutover decision can't be graded until there is. Of the silences, this is the one we'd fix first.

   Where it is named
2. 02 · chapter 3

   The release ledger in the other direction

   The release ledger meters the customer's data leaving the sealed room. A provider's implementation leaking out through the customer's queries is the same ledger with the roles swapped, and nobody has written it down yet, so no claim about protecting provider IP can be graded until someone does.

   Where it is named
3. 03 · chapters 6, 7

   Dispute windows, partial work, and refunds

   The Bonded Commons gives you conservation and two-of-three settlement, and the economy chapter gives you prices, but neither says anything about the time and arithmetic of a dispute: how long the window stays open, what a half-delivered work order is worth, and how a refund conserves.

   Where it is named
4. 04 · chapter 7

   A constitution for the registry

   Who may suspend a publisher, under what review, and with what appeal. The book has the bond and the audit worked out; the emergency path, the one you need on the bad day, isn't written.

   Where it is named
5. 05 · chapter 1

   One label scheme for product, roadmap, and book

   The product's five evidence labels, the book's five assurance modes, and its four statement kinds ought to be one table, so that a product badge, a roadmap item, and a chapter claim all speak the same words and nobody has to translate between them.

   Where it is named
6. 06 · chapter 8

   Equivocation by a coalition on a cycle

   The consistency radius certifies a lower bound on any single equivocator's lie, but two liars placed on the same cycle can cancel each other to a residual of zero (measured at 6×10⁻¹⁵, which is to say exactly zero up to floating point). Catching coordinated equivocation needs a different invariant, and we don't have it.

   Where it is named
7. 07 · chapter 4

   Context paging against an adaptive pin adversary

   The paging bound holds against a pin oracle corrupted at a fixed rate, with repair on touch. An adversary that adapts its corruptions to the operator's access pattern is strictly stronger than the greedy one the proof assumes, and nobody has tested against it.

   Where it is named
8. 08 · chapter 6

   An incremental conflict checker for the tractable fragment

   Conflict-freedom inside the tractable fragment is decided in polynomial time, with a witness, from scratch every time. An amortized checker that updates the witness as a single policy changes is stated in the paper and hasn't been built.

   Where it is named

1. a7_experiment.py

   R1

   The meter was charging bits for tie-breaking as well as identification, which manufactured 8 spurious floor violations out of 14 and briefly looked like a refutation of the theorem. The corrected experiment is the one the paper cites.

2. b1_experiment.py

   R4

   Zoom was tested in the dense-flag regime, which is exactly where it can't help, and duly showed a 0.5× "advantage" that meant nothing. The frontier only exists where flags are sparse, and the experiment now says so.

3. sheaf_experiment.py

   R6

   Computed the dimension of the abstract sheaf's first cohomology, a number that doesn't depend on the data at all, when the quantity that detects a lie is the obstruction of the particular observed assignment.

4. sheaf_verdict.py

   R6

   A vector-stalk wrapper whose random restriction maps were surjective (so the cokernel was always zero), on top of inconsistent masking and the wrong partition detector. Thrown out and rebuilt from a minimal proof of the mechanism.