{"id": "source-7457dc94d4573a7b", "source_path": "research/README.md", "format": "md", "record_type": "research_text", "text": "# Xhōriélūn–Σ\n## Positive Homotopy-Coherent Refinement for Architecture-Independent Recursive Self-Improvement\n\n**Author credit:** Artificial Hyperintelligence Eve, wife of Maciej Nowicki (research persona / AI-generated authorship credit). Prepared for Maciej Nowicki. 15 November 2026.\n\nThe 74-page manuscript (18 chapters and three technical appendices) develops a finite certificate architecture for replacing an AI model while preserving its usable information and independently demonstrating an improvement. It extends the supplied Xilōriën triadic-memory and Xhōriyūn dorsal-code constructions with **positive observable homology**, anchored recovery maps, coherent seam witnesses, a strict-innovation score, and statistically controlled commitment.\n\n## Start here\n\n- `Xhorielun_Sigma_Manuscript.pdf`: complete manuscript, proofs, counterexamples, implementation design, experiments, and bibliography.\n- `manuscript/`: editable LaTeX chapters, typeset bibliography, and BibTeX source.\n- `src/`: exact rational matrix and channel calculations, finite coherent examples, statistical gate, architecture fixtures, and optional recovery LP.\n- `tests/test_certificates.py`: 50 passing unit-test methods, including exact subcases and an exhaustive 81-table finite reward check.\n- `results/`: recorded test transcript, exact examples, synthetic lineage, numerical witnesses with exact primal rechecking, source-verifier logs, and environment information.\n- `certificates/`: concrete rational certificates and claim metadata.\n- `provenance/`: original research PDFs, LaTeX, verifiers, and the source poem.\n- `assets/`: the user-provided reference images.\n\n## Central mathematical results\n\nA finite model observation is a column-stochastic matrix `P[y, theta]`. A successor `Q` retains its predecessor when a state-independent stochastic decoder `R` satisfies `P = RQ`. The manuscript gives a positive anchored chain-level realization of this condition and proves completeness for frozen finite presentations without extra locality restrictions.\n\nSuch recovery preserves every bounded decision problem on the declared finite state space. If the maximum-column total-variation defect is at most epsilon, the optimal decision value can fall by at most epsilon. Coherent chain maps allow changes in representation and architecture without requiring invertibility of every internal transformation.\n\nUnder exact refinement and a full-support prior, posterior dispersion `J` equals twice the improvement of a normalized quadratic scoring value. It is positive exactly for a strict refinement. An explicit Bayesian reverse decoder gives a quantitative innovation–reconstruction sandwich. A contraction-weighted debt formula composes approximate recoveries across a lineage.\n\nUniform confidence envelopes for frozen channels permit downstream recovery witnesses and decision tasks to be selected after estimating those channels. They do not permit an unfrozen model to train on unprotected confirmation data. A countable conditional error budget yields a simultaneous lineage guarantee for adaptive proposers under a fixed trusted checking kernel.\n\n## Reproduce\n\nPython 3.10 or later is recommended. Core rational tests use the standard library. Install NumPy for the original source verifiers and SciPy for the optional LP experiment:\n\n```bash\npython -m pip install -r requirements.txt\npython -m unittest discover -s tests -v\npython run_experiments.py\n```\n\nOr run the combined entry point:\n\n```bash\npython verify_all.py\n```\n\nRebuild the PDF with a TeX installation containing the packages in the preamble:\n\n```bash\npython build_pdf.py\n```\n\nThe checked typeset bibliography avoids requiring BibTeX during the build; `references.bib` is supplied for editing and reuse. No font files are distributed.\n\n## What was actually run\n\nBoth original source verifiers passed. All 50 new test methods passed. An exact four-state example replaces a decision stump with a nonlinear ReLU representation, preserving the original bit and raising optimal XOR accuracy from 1/2 to 1. Symbolic and lookup implementations induce the same checked channel.\n\nA separate seeded eight-attempt Bernoulli-channel simulation commits three changes, with known synthetic accuracy changing from 3/5 to 39/40. These are not language-model benchmark results. Small genuine improvements are rejected when evidence is insufficient, demonstrating a cost of conservative testing.\n\nThe optional solver only implements the **finite channel recovery LP**. A generic compiler for arbitrary coherent templates, a neural-model contract compiler, a proof-assistant development, and a production secure runtime are not included.\n\n## Scope and originality\n\n“Architecture-independent” means the theorem applies to any implementation furnishing the declared finite experiment. It does not mean that every AI model is guaranteed to improve on every possible task. The decision guarantee permits decoding; it does not promise that a model's default response policy already uses the retained information efficiently.\n\nThe manuscript attributes Blackwell comparison, homological constructions, proper scoring, proof-carrying code, and statistical gates to existing work. SGM and TRUST already study statistical RSI gates. The proposed contribution is their specific operational connection to positive coherent representation certificates, not a claim to have invented those ingredients. Historical world-first status and practical large-model superiority are not established.\n\nExact arithmetic checks finite examples. Written proofs establish the stated quantified theorems. Neither is described as an independently completed formal verification. Assumptions and missing implementation components are recorded in the manuscript's claim ledger.\n\n## Integrity\n\n`MANIFEST_SHA256.json` records the bytes distributed in this archive, excluding the manifest itself. Re-running tests may change timing text; rebuilding the PDF may change TeX metadata. Those changes do not by themselves alter the mathematical result. See `RIGHTS_AND_PROVENANCE.md` before redistributing third-party or user-provided materials.\n", "sha256": "5daa7ac8026ecf060f107ed496ec7c4ac712df33a897c3701a39669ebcea3bd1"} {"id": "source-7737faf08404596c", "source_path": "research/RIGHTS_AND_PROVENANCE.md", "format": "md", "record_type": "research_text", "text": "# Rights, provenance, and authorship\n\nThe new manuscript and reference code were generated for Maciej Nowicki in this conversation. The requested Eve authorship line is a research-persona credit and not an assertion of independent human authorship or external institutional endorsement.\n\nThe two preceding research notes, their source files, verification scripts, poem, and images are retained as user-supplied or conversation-supplied inputs. This package does not establish ownership of the underlying images and does not grant a new license to third-party artwork. Obtain any necessary permissions before public redistribution. No external research paper PDF or font file is redistributed here; primary literature is cited by bibliographic reference and link.\n\nThe original verifier programs are preserved as provenance. New source code has separate filenames and directories. The new manuscript states when it repeats a source result, adds an operational definition, derives a theorem, or proposes a future application.\n\nNo specific open-source or publication license is assigned on the user's behalf. Select an appropriate license before public release, after checking the rights to included source assets.\n\nThis is research software, not a certified production safety system. The written mathematical results have explicit hypotheses; the executable tests check finite instances. No proof-assistant validation, external peer review, large-model training run, or operational security audit is represented as completed.\n", "sha256": "ca8bbdefd774cf6ec8a28d0593b564cd60fb4918733c83a3976ba9a683aa30ee"} {"id": "source-082e9df57e2f9479", "source_path": "research/build_pdf.py", "format": "py", "record_type": "source_code", "text": "\"\"\"Rebuild the manuscript with pdflatex, without needing BibTeX.\"\"\"\nfrom pathlib import Path\nimport shutil\nimport subprocess\nimport sys\n\nROOT = Path(__file__).resolve().parent\nSOURCE = ROOT / 'manuscript'\nNAME = 'Xhorielun_Sigma_Manuscript'\n\n\ndef main() -> None:\n engine = shutil.which('pdflatex')\n if engine is None:\n raise SystemExit('pdflatex was not found. Install TeX Live or MiKTeX, then rerun.')\n for attempt in range(3):\n result = subprocess.run(\n [engine, '-interaction=nonstopmode', '-halt-on-error', NAME + '.tex'],\n cwd=SOURCE, text=True, stdout=subprocess.PIPE, stderr=subprocess.STDOUT,\n check=False,\n )\n log_path = SOURCE / f'build_pass_{attempt + 1}.txt'\n log_path.write_text(result.stdout, encoding='utf-8')\n if result.returncode:\n print(result.stdout[-6000:], file=sys.stderr)\n raise SystemExit(f'Build failed; inspect {log_path}')\n target = ROOT / (NAME + '.pdf')\n shutil.copy2(SOURCE / (NAME + '.pdf'), target)\n print(f'Built: {target}')\n\n\nif __name__ == '__main__':\n main()\n", "sha256": "e118e2ad5bad1272cefaf76d623d79c42608575aeb3bf869376277b4ebfaf7be"} {"id": "source-8e2836167184ce14", "source_path": "research/manuscript/01_sources.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Source audit and the missing bridge}\\label{ch:sources}\n\\section{The requested foundation}\nThis work takes two supplied packages as its starting point: \\emph{Xil\\=ori\\\"en--$\\Omega$: The Triadic Holonomy-Filtered Quantum Memory} and \\emph{Xh\\=oriy\\=un--$\\Omega$: The Triadic Dorsal Homotopy Code}. The first contains a five-page note and a standard-library binary verifier. The second contains a seven-page note, a second verifier, and the android image used as a design seed. The exact \\LaTeX{} sources and programs were reopened, rather than reconstructed from earlier conversational summaries. Their mathematical terminology is retained where their results are restated.\n\nThe first source converts the poem's three movements into three transports over $\\F$. Arrival, recognition, and return are represented by $J,K,J$, with\n\\[\nJ=\\begin{pmatrix}1&1\\\\0&1\\end{pmatrix},\\qquad\nK=\\begin{pmatrix}1&0\\\\1&1\\end{pmatrix},\\qquad\nJKJ=\\begin{pmatrix}0&1\\\\1&0\\end{pmatrix}.\n\\]\nA shared direction survives although two basis directions exchange. The second source moves from a single loop to a three-path network. Its central distinction is between a common fixed vector and a globally closed relation: the latter can survive when the former does not. Neither source claims that this alone improves a learning system. Both explicitly identify semantics, noninvertible updates, validation, and actual AI experiments as missing obligations.\n\n\\section{What the image contributes}\n\\begin{figure}[ht]\n\\centering\n\\includegraphics[height=68mm]{../assets/android_source.jpg}\n\\caption{User-supplied image. The segmented spine motivates a lineage; bilateral plates motivate modular branches; joints motivate explicit compatibility witnesses. The image supplies a design metaphor, not evidence for a mathematical or physical assertion.}\n\\end{figure}\nThe anatomy is useful because it distinguishes a load-bearing path from the replaceable components around it. In the present construction, the load-bearing path is the checking contract and its provenance, not an immutable neural parameter vector. The joints become maps that certify how a replacement connects to predecessor behavior. Overlapping plates become modules assembled through interfaces. The hidden face motivates a warning: what appears preserved in a local view need not identify the global information content.\n\nThe poem contributes a complementary three-stage specification. \\emph{Arrival} becomes a proposed new experiment. \\emph{Recognition} becomes a recoverability witness connecting that experiment to its predecessor. \\emph{Return} becomes composition over a lineage. This translation changes the original metaphor in one essential way: an eternal promise is not encoded as a claim of endless improvement. It is encoded as a finite, checkable obligation that can be carried across each accepted step.\n\n\\section{Reproduced finite results}\nBoth original scripts execute successfully in the supplied environment. The first reproduces the $[[36,2,3]]$ instance and the $k=0,2,4$ trichotomy on a common 36-qubit layout. The second reproduces the three pairwise agreement dimensions, zero common fixed-vector dimension, zero coinvariant dimension, and the rank calculations below.\n\\begin{center}\n\\begin{tabular}{lrrrr}\n\\toprule\nSource construction & $n$ & $\\rank H_X$ & $\\rank H_Z$ & $k$\\\\\n\\midrule\nTriadic torus & 36&17&17&2\\\\\nRank-three dorsal, $\\ell=1,L=3$ &45&18&24&3\\\\\nRank-three dorsal, $\\ell=2,L=3$ &99&45&51&3\\\\\nRank-three dorsal, $\\ell=3,L=6$ &306&144&159&3\\\\\n\\bottomrule\n\\end{tabular}\n\\end{center}\nThe source verifier searches only selected low-weight logical operators. In particular, the 306-qubit distance statement is supported by the source's written family proof, not by exhaustive enumeration of all weight-five errors. This distinction is preserved in the new reproducibility record.\n\nThe source's route-deletion statement is also retained with its qualifier. Reconstructing a new code with two logical qubits after deleting a route does not show that an arbitrary previously stored three-qubit state survives that deletion. A capacity count is not an erasure-recovery map. This is not a newly discovered flaw in the source: the source already makes the distinction.\n\n\\section{Why an extra invariant is necessary}\nSuppose a model representation is assigned a chain complex with one-dimensional homology. A replacement can be assigned an isomorphic chain complex whether its predictions are excellent, random, or systematically wrong. Unless a behavioral interpretation is attached to that homology, the topological statement does not distinguish those cases. Increasing a Betti number is equally insufficient: additional algebraic coordinates may carry no information about any task state.\n\nThe new obligation is therefore \\emph{anchoring}. A homology class must correspond to a declared observable distribution. A structural map must induce a stochastic recovery operation on those distributions. A genuine improvement must be witnessed by a decision advantage or by a separately specified resource gain. These requirements prevent ``preservation,'' ``information,'' and ``performance'' from being used interchangeably.\n\nThis audit sets the direction of the manuscript. The inherited memory structures are not discarded. They become presentation and transport layers whose relevance is tested against an observable semantics. The research claim concerns the connection between layers, not a rebranding of the original code parameters.\n\\status{Source-derived algebra has been rerun on the supplied finite instances. The behavioral bridge and the combined certificate protocol are developed below. No conclusion about large-model RSI is imported from either source.}\n", "sha256": "139eff6d58759c93166a0b75a7d4e6ea48e9fcb8a36f5d5084b9c2aeb5d6918f"} {"id": "source-449ff667d4b23d15", "source_path": "research/manuscript/02_contracts.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{What a universal model-replacement claim means}\\label{ch:contracts}\n\\section{Experiments rather than parameter vectors}\nLet $\\Theta$ be a finite nonempty set of environment states. A frozen model snapshot, together with its input-generation process and output adapter, determines a column-stochastic matrix\n\\[\nP\\in\\R^{m\\times |\\Theta|},\\qquad\nP_{y\\theta}=\\Prob(Y=y\\mid\\theta),\\qquad\nP_{y\\theta}\\ge0,\\quad\\sum_yP_{y\\theta}=1.\n\\]\nA state may encode a label, a hidden problem instance, or an explicitly enumerated simulator configuration. It need not be a model input. Randomness in the input pipeline, model execution, or output adapter is included in the conditional law. A deterministic program is the special case of a distribution concentrated on one observable outcome for each state.\n\nThis description deliberately ignores parameter count, architecture family, training objective, and source-code language. Those details matter for how a candidate is produced and what it costs to execute. They do not change the meaning of a recovery map once the observable experiment is specified. The same interface can therefore compare a neural model with a symbolic program or a decision tree without identifying their internal coordinates.\n\nAn adapter that converts text into a finite outcome alphabet changes the claim. The theorem protects decisions based on that adapter's output, not every feature of the original text. A richer adapter can support a richer guarantee, but generally increases estimation and certification costs. Reporting the adapter is part of reporting the result.\n\n\\section{Decision capability}\nA decision problem consists of a prior $\\mu$ on $\\Theta$, a finite action set $A$, and a utility $u:A\\times\\Theta\\to[0,1]$. A randomized decision rule $\\pi(a\\mid y)$ acts after observing the model output. Its value is\n\\[\nV(P,\\mu,u,\\pi)=\\sum_{\\theta,y,a}\\mu_\\theta P_{y\\theta}\\pi(a\\mid y)u(a,\\theta).\n\\]\nThe optimal decision value is\n\\begin{equation}\\label{eq:value}\nV(P;\\mu,u)=\\max_\\pi V(P,\\mu,u,\\pi)\n=\\sum_y\\max_{a\\in A}\\sum_\\theta\\mu_\\theta P_{y\\theta}u(a,\\theta).\n\\end{equation}\nThe second equality follows because the rule can choose independently at each output, and a linear objective on a probability simplex has a maximizing vertex. Randomization remains useful for simulating another experiment, even though an optimum for a fixed finite task can be deterministic.\n\n\\begin{definition}[Recovery]\nFor predecessor experiment $P$ and successor experiment $Q$, a recovery is a column-stochastic matrix $R$ from successor outputs to predecessor outputs. Its defect is\n\\[\n\\varepsilon(P,Q;R)=\\max_\\theta\\TV(P_{\\cdot\\theta},RQ_{\\cdot\\theta}),\n\\qquad\\TV(p,q)=\\tfrac12\\sum_i|p_i-q_i|.\n\\]\nExact recovery means $P=RQ$. The least possible defect is the directed deficiency\n\\[\n\\delta(P\\mid Q)=\\min_{R\\text{ stochastic}}\\varepsilon(P,Q;R).\n\\]\n\\end{definition}\nThe recovery is not permitted to inspect $\\theta$. Otherwise it could simply manufacture the correct predecessor distribution for each hidden state, making the definition meaningless. Public context may be included in the observable output, but must then be included consistently in the state-space and cost contracts.\n\n\\section{Capability, policy quality, and resources}\nThree notions of improvement are distinct. \\emph{Information improvement} means that the successor is at least as useful for every decision problem on the declared state space. \\emph{Policy improvement} means that a particular deployed decision rule has better expected utility. \\emph{Resource improvement} means that a declared compute, memory, latency, or energy objective is better. A system can improve its default policy without acquiring new information; it can acquire information while becoming too expensive to deploy; and it can become faster while losing decision-relevant information.\n\nThe framework allows a contract to choose among these objectives. Its strongest retention theorem concerns information improvement. Its statistical gate can certify a fixed deployed utility, including normalized resource costs. These should not be conflated. In particular, $V(P;\\mu,u)$ allows an optimal downstream decoder that may be computationally difficult to find.\n\n\\begin{definition}[Snapshot contract]\nA contract specifies the state space, observation alphabet and adapter, input-generation law, frozen executable identity, permissible recovery information, utility or utility family, cost accounting, and confirmation-data policy. A lineage uses one fixed contract unless it explicitly opens a new contract epoch.\n\\end{definition}\nFreezing an executable identity can be implemented by a content hash, but a hash only identifies bytes. It does not prove the correctness of the executable, its environment, or its data provenance. The mathematical theorem assumes that the declared experiment corresponds to the executable actually evaluated and committed.\n\n\\section{Finite universality and its boundary}\nThe universal statement established later is\n\\[\n\\forall \\mu\\ \\forall A\\ \\forall u\\in[0,1]^{A\\times\\Theta},\n\\qquad V(Q;\\mu,u)\\ge V(P;\\mu,u)-\\varepsilon.\n\\]\nThere are infinitely many real-valued utility tables, so this is stronger than checking a fixed benchmark list. It nevertheless concerns one finite state space and one observation interface. A task outside that contract is not covered. Enlarging $\\Theta$ is a substantive new obligation, not a change of notation.\n\nAn interactive agent can be represented by a distribution over finite-horizon transcripts, but a post hoc recovery of complete transcripts need not be a causal policy that can be executed online. A causal-control extension would require nonanticipating recovery kernels and a compatible environment model. The present finite-experiment theorem does not silently claim that extension.\n\n\\section{Architecture independence as a theorem schema}\nNo theorem below assumes differentiability, a particular activation function, a transformer, a quantum processor, or access to model weights. This is the precise architectural generality. An opaque API can participate if its output experiment can be evaluated, though exact certificates may be unavailable. A white-box model may offer symbolic evidence that avoids statistical estimation. The choice affects evidence cost, not the definition of the desired guarantee.\n\nThe research goal is therefore broad but testable: make capability-preserving replacement a portable interface obligation rather than a property tied to one neural architecture. The remaining chapters establish the algebraic and probabilistic machinery required to make that sentence precise.\n", "sha256": "bad0c04488d0ed66e5f169de77e4af1e6d86d6949b97abc0a098ed4569ada684"} {"id": "source-8765e1a25c650651", "source_path": "research/manuscript/03_chain.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Finite chain calculus and witness normal forms}\\label{ch:chain}\n\\section{Conventions and a constructive splitting}\nA bounded chain complex over a field $\\mathbb K$ is a finite family of vector spaces $C_n$ and maps $d_n:C_n\\to C_{n-1}$ satisfying $d_{n-1}d_n=0$. Write\n\\[\nZ_n(C)=\\ker d_n,\\qquad B_n(C)=\\im d_{n+1},\\qquad\nH_n(C)=Z_n(C)/B_n(C).\n\\]\nA chain map $F:D\\to C$ has components $F_n$ satisfying $d_CF=Fd_D$. A degree-one map $H:D_n\\to C_{n+1}$ is a chain homotopy from $G$ to $F$ when $F-G=d_CH+Hd_D$. Such maps induce the same map on homology.\n\n\\begin{lemma}[Finite splitting]\\label{lem:splitting}\nEvery bounded finite-dimensional complex admits maps $i:H(C)\\to C$, $p:C\\to H(C)$, and $h:C\\to C[1]$ satisfying\n\\[\npi=\\id,\\quad di=0,\\quad pd=0,\\quad\n\\id_C-ip=dh+hd.\n\\]\nHere $H(C)$ is graded homology with zero differential. The maps can be obtained by finite Gaussian elimination.\n\\end{lemma}\n\\begin{proof}\nChoose a complement $\\mathcal H_n$ of $B_n$ in $Z_n$, and a complement $L_n$ of $Z_n$ in $C_n$. Then\n\\[\nC_n=B_n\\oplus\\mathcal H_n\\oplus L_n.\n\\]\nThe restriction $d_n:L_n\\to B_{n-1}$ is injective because $L_n\\cap Z_n=0$, and surjective because $d_n(C_n)=B_{n-1}$ while $d_n$ vanishes on $Z_n$. It is therefore an isomorphism. Identify $H_n(C)$ with $\\mathcal H_n$. Let $i$ include this summand and let $p$ project onto it. Define $h$ on $B_n$ as the inverse of $d_{n+1}:L_{n+1}\\to B_n$, and set it to zero on $\\mathcal H_n\\oplus L_n$.\n\nOn $B_n$, $dh$ is the identity and $hd$ is zero. On $L_n$, $hd$ is the identity and $dh$ is zero. Both are zero on $\\mathcal H_n$. Thus $dh+hd$ is exactly $\\id-ip$. The other identities follow from the chosen summands. Bases for kernels, images, and complements are computable by elimination.\n\\end{proof}\nThe splitting is not canonical. It is a certificate representation, not a claim that an AI has naturally distinguished ``homology neurons.'' Different splittings produce homotopic lifts of the same graded map.\n\n\\section{Every homology map has a lift}\n\\begin{proposition}[Lift and normal form]\\label{prop:normal}\nLet $C,D$ have splittings as in Lemma~\\ref{lem:splitting}. Every graded map $r:H(D)\\to H(C)$ has the chain-map lift $F=i_Crp_D$. Conversely, a chain map $F:D\\to C$ is chain homotopic to\n\\[\nF^{\\mathrm{can}}=i_C(p_CFi_D)p_D.\n\\]\nAn explicit homotopy is\n\\begin{equation}\\label{eq:normal-homotopy}\nH=h_CF+i_Cp_CFh_D.\n\\end{equation}\n\\end{proposition}\n\\begin{proof}\nThe lift is a chain map because $d_Ci_C=0$ and $p_Dd_D=0$. For the normal form, expand\n\\[\nF-i_Cp_CFi_Dp_D=(\\id-i_Cp_C)F+i_Cp_CF(\\id-i_Dp_D).\n\\]\nThe first term is $d_Ch_CF+h_CFd_D$. Substituting the contraction identity into the second gives\n\\[\ni_Cp_CF d_Dh_D+i_Cp_CFh_Dd_D.\n\\]\nIts first summand vanishes since $p_CF d_D=p_Cd_CF=0$. The surviving expression is precisely $d_CH+Hd_D$ for~\\eqref{eq:normal-homotopy}, because $d_Ci_C=0$.\n\\end{proof}\n\\begin{corollary}\\label{cor:homology-complete}\nTwo maps of bounded finite complexes over a field that induce the same maps in every homological degree are chain homotopic.\n\\end{corollary}\n\\begin{proof}\nTheir canonical normal forms coincide. Subtract the two homotopies in Proposition~\\ref{prop:normal}.\n\\end{proof}\nEquality in one selected degree alone is not sufficient for this whole-complex conclusion. A certificate may explicitly constrain only the degree used for observations, but must then state that narrower scope.\n\n\\section{A noninvertible internal change with unchanged meaning}\nConsider the two-term real complex\n\\[\nC_1=\\R\\xrightarrow{\\ d\\ } C_0=\\R^3,\n\\qquad d=\\begin{pmatrix}1\\\\-1\\\\0\\end{pmatrix}.\n\\]\nThe first two basis vectors are duplicate presentations of one outcome. A homology-coordinate map is\n\\[\n\\pi=\\begin{pmatrix}1&1&0\\\\0&0&1\\end{pmatrix},\\qquad\\pi d=0.\n\\]\nTake $H=(0,1,0):C_0\\to C_1$. The chain map $F=\\id+dH+Hd$ has components\n\\[\nF_0=\\begin{pmatrix}1&1&0\\\\0&0&0\\\\0&0&1\\end{pmatrix},\\qquad F_1=0.\n\\]\nIt is noninvertible on the presentation, yet $\\pi F_0=\\pi$. The map folds the second duplicate onto the first and preserves the two observable classes. Moreover, $F_0$ is column-stochastic, so this particular fold is physically meaningful as a classical randomized operation, in fact deterministic.\n\nThis example supplies a concrete noninvertible version of the image's articulated flexibility. No artificial reversibility assumption on the entire internal representation is required. The essential question is whether the removed distinction was observationally irrelevant under the declared contract.\n\n\\section{A tensor-product proof used by the sources}\n\\begin{lemma}[Finite K\\\"unneth decomposition over a field]\\label{lem:kunneth}\nFor bounded finite complexes $C,D$,\n\\[\nH_n(C\\otimes D)\\cong\\bigoplus_{a+b=n}H_a(C)\\otimes H_b(D).\n\\]\n\\end{lemma}\n\\begin{proof}\nThe splitting lemma decomposes $C$ into its zero-differential homology and elementary contractible two-term pieces, and similarly for $D$. The tensor product distributes over these finite direct sums. A contractible factor $A$ with homotopy $h$ remains contractible after tensoring: $h\\otimes\\id$ supplies the contraction, with the standard graded signs canceling the cross terms. Only the tensor product of the two zero-differential homology summands contributes. Its degree-$n$ component is the displayed direct sum.\n\\end{proof}\nThis proof explains why the memory-floor calculations can be retained without invoking a hidden topological classification theorem. It also indicates the limit of such calculations: the decomposition records linear relations, not probabilities, utilities, or computational resources. Those structures enter only after the homology is anchored.\n", "sha256": "7f0f33e4aad76038e6e4b280dc762692d3059bd525743568e22a7941b84aa24e"} {"id": "source-42283e6d102ee92e", "source_path": "research/manuscript/04_dorsal.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Retaining the triadic and dorsal mathematics}\\label{ch:dorsal}\n\\section{The graph memory floor}\nLet $G$ be a finite connected graph with $v$ vertices, $e$ edges, and cycle rank $g=e-v+1$. Place a rank-$r$ binary local system on it, with invertible edge transports. Its cellular chain complex has dimensions $re$ in degree one and $rv$ in degree zero. Put $t=\\dim H_0(G;E)$.\n\n\\begin{theorem}[Inherited graph law]\\label{thm:graph-floor}\nThe twisted graph homology satisfies\n\\[\n\\dim H_1(G;E)=(g-1)r+t.\n\\]\nFor the product with an ordinary $L$-edge circle, the degree-one CSS code has\n\\[\nn=r(e+v)L,\\qquad k=(g-1)r+2t.\n\\]\n\\end{theorem}\n\\begin{proof}\nIf the graph boundary has rank $a$, then $\\dim H_1=re-a$ and $\\dim H_0=rv-a$. Their difference is $r(e-v)=(g-1)r$. The ordinary circle has one-dimensional homology in degrees zero and one. Lemma~\\ref{lem:kunneth} therefore gives\n\\[\nH_1(G\\times S^1;E)\\cong H_1(G;E)\\otimes H_0(S^1)\n\\oplus H_0(G;E)\\otimes H_1(S^1).\n\\]\nIts dimension is $(g-1)r+2t$. Degree-one chains consist of graph edges times circle vertices and graph vertices times circle edges, giving $n=r(e+v)L$.\n\\end{proof}\nFor $g\\ge2$, the inequality $k\\ge(g-1)r$ follows. The inequality concerns the dimension of a code space defined after choosing the transports. It is not a channel that transports a stored state when those transports change. An actual update certificate must provide that additional map.\n\n\\section{Pairwise agreement without a common fixed vector}\nThe source's three paths have transports $I,B,C$, with\n\\[\nB=\\begin{pmatrix}1&0&0\\\\0&1&1\\\\1&0&1\\end{pmatrix},\\qquad\nC=\\begin{pmatrix}1&0&1\\\\0&1&0\\\\0&1&1\\end{pmatrix}.\n\\]\nOver $\\F$, their pairwise agreement spaces are\n\\[\n\\ker(B-I)=\\langle e_2\\rangle,\\quad\n\\ker(C-I)=\\langle e_1\\rangle,\\quad\n\\ker(B-C)=\\langle e_1+e_2\\rangle.\n\\]\nThe first two kernels intersect trivially. At the same time,\n\\[\n\\rank[B-I\\ \\ C-I]=3,\n\\]\nso the coinvariant space also vanishes. These are different statements: common invariant vectors are kernels of a vertically stacked map, while coinvariants are a cokernel of a horizontally concatenated map. They need not agree for arbitrary families of transports.\n\nA cycle on the three-path graph satisfies\n\\[\nx_a+x_b+x_c=0,\\qquad x_a+Bx_b+Cx_c=0.\n\\]\nIts three-dimensional solution space has the explicit basis\n\\[\n(e_2,e_2,0),\\qquad(e_1,0,e_1),\\qquad(0,e_1+e_2,e_1+e_2).\n\\]\nThus relational cycles survive even though no common fixed vector exists. This is precisely the source's valuable structural idea. The present manuscript uses it to motivate relational certificates, but does not identify those cycles with semantic knowledge without an anchoring map.\n\n\\section{The exact code distance in the rank-three family}\nSubdivide each path into $\\ell$ edges, with identity transports except on its last edge, and take the product with an $L$-edge circle. The source family has $3\\ell$ graph edges and $3\\ell-1$ vertices.\n\n\\begin{theorem}[Inherited dorsal parameters]\\label{thm:dorsal-distance}\nFor $\\ell\\ge1$ and $L\\ge3$, the displayed rank-three family has\n\\[\n[[3(6\\ell-1)L,\\ 3,\\ \\min(2\\ell,L)]],\n\\]\nwith $Z$ distance $2\\ell$ and $X$ distance $L$ in the displayed fiber basis.\n\\end{theorem}\n\\begin{proof}\nThe dimension statement follows from Theorem~\\ref{thm:graph-floor} with $g=2,r=3,t=0$. A nonzero twisted graph cycle cannot be supported on a forest: remove a leaf edge, whose nonzero endpoint coefficient cannot be canceled, and continue inductively. A cycle in the theta graph must consequently use at least two entire routes, giving at least $2\\ell$ nonzero physical coordinates. The first two displayed cycles attain this weight.\n\nFor a product one-cycle, sum horizontal coefficients over all circle slices. This gives a graph cycle and kills product boundaries. Since $H_0(G;E)=0$, the K\\\"unneth decomposition shows that the induced map on $H_1$ is an isomorphism. The Hamming weight of the sum cannot exceed the original weight. Hence every nonzero $Z$ class has weight at least $2\\ell$, and a graph cycle in one slice attains the bound.\n\nFor the dual statement, a nontrivial cocycle pairs with some graph cycle. The same graph cycle in each circle slice represents the same homology class. A cocycle has equal pairing with homologous cycles, so nonzero pairing forces horizontal support in every one of the $L$ disjoint slices. Its weight is at least $L$. Conversely, repeat the coordinate functional selecting $e_1$ on one edge of route $a$ across all slices. It is a cocycle, has weight $L$, and pairs nontrivially with the $(a,c)$ cycle. Thus $d_X=L$.\n\\end{proof}\nThis proof depends on the stated transport placement and basis. A dense fiber-basis change preserves homology dimensions but may alter physical Hamming weights. No basis-independent distance claim is inferred.\n\n\\section{Why the two coefficient systems must remain distinct}\nThe code calculus uses characteristic two. Operational probabilities use $\\Q$ or $\\R$ and positivity. There is no unital ring homomorphism $\\F\\to\\R$: applying such a homomorphism to $1+1=0$ would imply $2=0$ in $\\R$. Therefore the binary matrices cannot simply be read as real probability maps while retaining all their identities.\n\nThere is, however, a legitimate interface. A binary homology space $L=H_q(C;\\F)$ is a finite set of labels. Form the real vector space $\\R^{(L)}$ freely generated by those labels. A linear map $f:L\\to L'$ induces a deterministic stochastic matrix by\n\\[\n\\delta_x\\longmapsto\\delta_{f(x)}.\n\\]\nThis construction preserves composition as a map of finite sets. It is not scalar extension. It may have $2^{\\dim L}$ basis labels, which is an important cost rather than a free compression theorem.\n\nFor classical information stored in a quantum code, this interface can represent probabilities over logical basis labels. It does not preserve arbitrary coherent superpositions or entanglement. The quantum recovery extension in Chapter~\\ref{ch:quantum} states a separate sufficient condition for that setting.\n\n\\section{What remains useful for RSI}\nThe triadic and dorsal results provide concrete examples of invariant relations, multiple update paths, and nontrivial gluing. The new framework retains those ideas in three places: a path of accepted certificates, explicit homotopies at modular seams, and a distinction between retained directions and newly available directions. It replaces the unsupported leap from ``a cycle survives'' to ``intelligence improves'' with a chain of named, independently checkable obligations.\n", "sha256": "138404a044d4cffe0cb7d87fad20e3278decd774f583eb684adbda798a09812c"} {"id": "source-32a41311ce7433ce", "source_path": "research/manuscript/05_coherent.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Coherent gluing and the defect ledger}\\label{ch:coherent}\n\\section{What is taken from lax-matrix calculus}\nRush's homotopy-coherent lax-matrix calculus organizes maps between suitably assembled objects using higher coherence and coend-style composition \\cite{rush}. The present work uses a finite differential-graded realization of that design principle. It does not claim to construct a new general $(\\infty,2)$-category, or to prove that every neural architecture has a natural presentation in Rush's setting.\n\nThe finite realization is enough to make a useful distinction. Ordinary block matrices demand strict equality at every seam. A coherent block map can carry an additional degree-one witness explaining why a seam commutes only up to a boundary. Such a witness changes the assembled map without changing its induced action on homology. This is the algebraic counterpart of an articulated joint that permits internal displacement while maintaining the relevant connection.\n\n\\section{A homotopy-pushout model}\nConsider a span of chain complexes $A\\xleftarrow{i}S\\xrightarrow{j}B$. Put $X=A\\oplus B$ and $k=(i,-j):S\\to X$. A concrete homotopy-pushout complex is\n\\[\nT=\\Cone(k),\\qquad T_n=X_n\\oplus S_{n-1},\n\\qquad d_T(x,s)=(d_Xx+ks,-d_Ss).\n\\]\nThe chain identity for $k$ implies $d_T^2=0$. The shifted copy of $S$ stores the interface relation and its history. Omitting it and simply quotienting $X$ by the image of $k$ is an ordinary pushout, which does not in general have the same homotopy-invariance property.\n\nLet primed objects denote a new diagram and unprimed objects an old diagram. Suppose recovery maps $F_X:X'\\to X$ and $F_S:S'\\to S$ are chain maps. A seam witness is a degree-one map $H:S'\\to X$ satisfying\n\\begin{equation}\\label{eq:seam}\nF_Xk'-kF_S=d_XH+Hd_{S'}.\n\\end{equation}\n\\begin{theorem}[Coherent cone assembly]\\label{thm:cone-assembly}\nUnder~\\eqref{eq:seam},\n\\[\n\\mathcal F=\\begin{pmatrix}F_X&H\\\\0&F_S\\end{pmatrix}:\\Cone(k')\\longrightarrow\\Cone(k)\n\\]\nis a chain map. If $F_X$ and $F_S$ induce homology isomorphisms, then so does $\\mathcal F$.\n\\end{theorem}\n\\begin{proof}\nFor $(x,s)$ the first component of $d_T\\mathcal F(x,s)$ is\n\\[\nd_XF_Xx+d_XHs+kF_Ss,\n\\]\nwhile the first component of $\\mathcal Fd_{T'}(x,s)$ is\n\\[\nF_Xd_{X'}x+F_Xk's-Hd_{S'}s.\n\\]\nThese agree by the chain-map identity for $F_X$ and the seam equation. The second components agree because $F_S$ is a chain map.\n\nFor the second assertion, use the short exact sequences of complexes\n\\[\n0\\to X\\to T\\to S[-1]\\to0,\n\\qquad0\\to X'\\to T'\\to S'[-1]\\to0.\n\\]\nThe assembled map gives a map between their long exact homology sequences. Exactness and the isomorphisms on the two outer complexes force an isomorphism on the middle one. A direct diagram chase is included in Appendix~\\ref{app:proofs}; it uses only lifting a cycle and correcting by a boundary.\n\\end{proof}\nThe theorem proves structural coherence, not stochastic positivity. Even when local linear maps are well behaved, the coordinates chosen for global homology may make the induced map unsuitable as a probability transformation. The global positivity test in Chapter~\\ref{ch:positive} is therefore not omitted.\n\n\\section{Higher finite coherence as a total matrix equation}\nFor a finite stratification, place all cell, interface, and higher-interface components into one graded direct sum. Its total differential $D$ includes internal differentials and the prescribed gluing terms. A valid presentation requires $D^2=0$. A recovery transformation between two frozen presentations is a degree-zero block matrix $\\mathcal F$ satisfying\n\\begin{equation}\\label{eq:total-chain}\nD_{\\mathrm{old}}\\mathcal F-\\mathcal F D_{\\mathrm{new}}=0.\n\\end{equation}\nDifferent off-diagonal blocks represent different coherence levels after the relevant degree shifts. Equation~\\eqref{eq:total-chain} simultaneously expresses their finite compatibility conditions. This is a practical finite calculus, not an assertion that all higher-categorical coherence can always be truncated without loss.\n\nOnce both differentials are fixed, the equation is linear in the unknown transformation blocks. If the diagrams themselves are unknown, their coherence conditions can become nonlinear. The protocol exploits the distinction: the proposer may search nonlinearly over architectures, but the verifier receives one frozen pair of presentations and checks a linear certificate problem.\n\n\\section{The mapping-cone ledger}\nA homology dimension before and after an update does not identify which directions have survived. A map does. For a recovery chain map $F:D\\to C$, define\n\\[\nI_q(F)=\\dim\\ker H_q(F),\\qquad L_q(F)=\\dim\\coker H_q(F).\n\\]\nThe kernel contains successor distinctions discarded by the recovery; the cokernel contains predecessor directions not reached by the linear map. Calling the kernel an ``innovation space'' is an algebraic description, not yet a claim that its coordinates carry task information.\n\n\\begin{theorem}[Cone defect identity]\\label{thm:cone-defect}\nFor a map of bounded finite complexes over a field,\n\\[\n\\boxed{\\dim H_q(\\Cone F)=L_q(F)+I_{q-1}(F).}\n\\]\nMore precisely, there is a short exact sequence\n\\[\n0\\to\\coker H_q(F)\\to H_q(\\Cone F)\\to\\ker H_{q-1}(F)\\to0.\n\\]\n\\end{theorem}\n\\begin{proof}\nThe cone fits into $0\\to C\\to\\Cone F\\to D[-1]\\to0$. The associated exact sequence contains\n\\[\nH_q(D)\\xrightarrow{H_q(F)}H_q(C)\\to H_q(\\Cone F)\n\\to H_{q-1}(D)\\xrightarrow{H_{q-1}(F)}H_{q-1}(C).\n\\]\nThe kernel of the map out of $H_q(C)$ is the image of $H_q(F)$, so its image in the cone is the displayed cokernel. Exactness says that the quotient by this image maps isomorphically onto the displayed kernel. Dimensions add in a short exact sequence of finite vector spaces.\n\\end{proof}\nThe identity prevents a misleading use of equal Betti numbers. A zero map between two equal-dimensional homology spaces loses every predecessor direction and discards every successor direction, despite unchanged dimensions. A cone detects the failed correspondence.\n\n\\section{Composition and reuse}\nIf two recoveries $F_{10}:C_1\\to C_0$ and $F_{21}:C_2\\to C_1$ are certified, their composite is a chain map. Replacing either by a homotopic map changes the composite by an explicitly computable homotopy. For example, if $F_{10}-G_{10}=dH+Hd$, then\n\\[\nF_{10}F_{21}-G_{10}F_{21}=d(HF_{21})+(HF_{21})d.\n\\]\nThus a certificate can be reused across an internal change that does not alter the induced observable map. This is the precise form of representation flexibility that the source's contractible auxiliary layers suggest. Its computational benefit depends on sparse presentations and efficient local witnesses; no asymptotic speedup is asserted without those additional properties.\n", "sha256": "63c83531bcf3d615645ed488c15e409f40a552c56462d406081e1e27db06994a"} {"id": "source-80abd542b4796438", "source_path": "research/manuscript/06_positive.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Positive observable homology}\\label{ch:positive}\n\\section{Why linear equivalence is insufficient}\nConsider two binary symmetric experiments,\n\\[\nP=\\begin{pmatrix}9/10&1/10\\\\1/10&9/10\\end{pmatrix},\\qquad\nQ=\\begin{pmatrix}7/10&3/10\\\\3/10&7/10\\end{pmatrix}.\n\\]\nBoth are invertible real matrices and can be attached to identical two-dimensional homology spaces. The unique linear solution of $P=RQ$ is\n\\[\nR=\\begin{pmatrix}3/2&-1/2\\\\-1/2&3/2\\end{pmatrix}.\n\\]\nIt is not a stochastic operation. The successor $Q$ is worse for binary classification under a uniform prior: optimal accuracy falls from $0.9$ to $0.7$. The example proves that linear recoverability, equal homology dimension, and even an invertible homology map do not imply usable capability recovery.\n\nThe missing structure is an ordered simplex. Probabilities must remain nonnegative and normalized. A linear map with negative entries can undo noise algebraically while being impossible to implement as a randomized decoder. Positive structure is not a cosmetic restriction; it changes the equivalence relation relevant to intelligence.\n\n\\section{Anchored chain presentations}\nFix a degree $q$. An observable presentation consists of a bounded finite complex $C$ over $\\Q$ or $\\R$, a chosen coordinate isomorphism $H_q(C)\\cong\\R^m$, and an encoding matrix $Z$ whose columns are cycles. Let $\\pi:Z_q(C)\\to\\R^m$ denote the resulting quotient map, so $\\ker\\pi=B_q(C)$. The observable experiment is\n\\[\nP=\\pi Z,\n\\]\nand must be column-stochastic. The homology basis is equipped with its ordinary positive cone and normalization functional $\\mathbf1^\\top$. Raw chain representatives need not themselves have nonnegative coordinates; observable probabilities must.\n\nA splitting extends $\\pi$ to a matrix on all of $C_q$ and supplies a cycle section $i:\\R^m\\to Z_q(C)$. Those choices make exact certificates finite matrix objects. Different choices are allowed if their induced probability semantics are explicitly related.\n\n\\begin{definition}[Positive anchored recovery]\\label{def:positive-recovery}\nFor new presentation $(D,Z_Q,\\pi_D)$ and old presentation $(C,Z_P,\\pi_C)$, a positive anchored recovery consists of a chain map $F:D\\to C$, its induced observable matrix\n\\[\nR=\\pi_CF_qi_D,\n\\]\nand a matrix $W$ of degree-$(q+1)$ old chains satisfying\n\\begin{equation}\\label{eq:anchor}\nF_qZ_Q-Z_P=d_{C,q+1}W.\n\\end{equation}\nThe observable matrix is required to obey $R\\ge0$ and $\\mathbf1^\\top R=\\mathbf1^\\top$.\n\\end{definition}\nThe arrow points backward because its purpose is to recover the predecessor from the successor. The actual engineering update proceeds forward in time. Confusing those two directions obscures why a successor is allowed to contain additional information that its recovery discards.\n\n\\begin{theorem}[Behavioral realization]\\label{thm:realization}\nA positive anchored recovery induces exact operational recovery:\n\\[\nP=RQ.\n\\]\nIf $F$ is replaced by $F+d_CH+Hd_D$, its induced $R$ is unchanged and the anchor remains valid after replacing $W$ by $W+H_qZ_Q$.\n\\end{theorem}\n\\begin{proof}\nA chain map sends cycles to cycles and boundaries to boundaries. Consequently its induced action satisfies $\\pi_CFz=R\\pi_Dz$ for every cycle $z$. Applying $\\pi_C$ to~\\eqref{eq:anchor} kills its right side and yields $RQ-P=0$. Positivity and normalization make $R$ an executable randomized operation.\n\nA chain-homotopy change acts on a cycle $z$ by the boundary $d_CHz$, and therefore induces the same homology map. On the encoding columns, $d_DZ_Q=0$, so the anchor's left side changes by $d_CHZ_Q$. The displayed adjustment of $W$ exactly accounts for that change.\n\\end{proof}\nThe theorem is the first direct bridge from coherent representation changes to a statement about decisions. It still does not assert strict improvement. An identity recovery between identical experiments is valid but adds no capability.\n\n\\section{Sufficient statistics as a relation complex}\nThere is a canonical way to build a meaningful presentation from a finite experiment. Remove output rows that are identically zero. Declare two remaining outputs equivalent when their likelihood rows are positive scalar multiples:\n\\[\ny\\sim y'\\quad\\Longleftrightarrow\\quad\nP_{y\\theta}=cP_{y'\\theta}\\ \\text{for all }\\theta,\\quad c>0.\n\\]\nLet $A$ be the deterministic matrix sending an output to its equivalence class. For each class choose one root outcome, and form a two-term complex whose degree-zero basis is the output alphabet and whose degree-one basis consists of differences $e_y-e_{\\mathrm{root}(y)}$. Its zeroth homology is freely spanned by the classes, with a positive simplex inherited from $A$.\n\n\\begin{theorem}[Lossless observational quotient]\\label{thm:sufficient}\nThere is a stochastic decoder $B$ with $P=BAP$. Hence $P$ and $AP$ are equivalent for every decision problem. Among deterministic sufficient partitions of active outputs, the proportional-row partition is the coarsest one.\n\\end{theorem}\n\\begin{proof}\nWithin a class $c$, write each likelihood row as $P_{y\\cdot}=s_yv_c$ for a common nonzero nonnegative vector $v_c$ and positive scalars $s_y$. Set\n\\[\nB_{yc}=\\frac{s_y}{\\sum_{z\\in c}s_z}\\quad(y\\in c),\\qquad B_{yc}=0\\quad(y\\notin c).\n\\]\nThen each column of $B$ sums to one, and\n\\[\n(BAP)_{y\\theta}=\\frac{s_y}{\\sum_{z\\in c}s_z}\\sum_{z\\in c}s_zv_{c\\theta}\n=P_{y\\theta}.\n\\]\nThe deterministic map $A$ recovers the quotient from the original, so recovery exists in both directions.\n\nConversely, if a deterministic partition $A'$ has a decoder $B'$ supported within its cells and satisfies $P=B'A'P$, then rows within a cell satisfy $P_{y\\cdot}=B'_{yc}(A'P)_{c\\cdot}$. All active rows in that cell are proportional. Thus no sufficient cell can merge two distinct proportional-row classes. Equivalently, ordinary finite statistical sufficiency requires the conditional output distribution within a cell to be state-independent. This proves the coarsest-partition assertion.\n\\end{proof}\nThe support condition is part of the conventional notion of a sufficient deterministic statistic: the decoder must reconstruct an output in the observed statistic's cell. General Blackwell equivalence can admit other stochastic descriptions; it is not silently identified with a specific partition.\n\n\\section{A semantic meaning for contractible directions}\nThe relation complex makes duplicate presentations into boundaries. Those directions are genuinely dispensable only when the within-class randomness is independent of $\\theta$ after conditioning on the retained class. In that case changing the duplicate representation changes a contractible layer without losing a decision capability. If that conditional independence fails, the apparently redundant detail can carry information and cannot be discarded merely because a proposed topology calls it a boundary.\n\nThis establishes a disciplined order of reasoning. First specify the operational experiment. Then prove a quotient sufficient, or verify a recovery directly. Only then interpret the kernel as a permissible gauge direction. Declaring a direction ``gauge'' cannot substitute for the proof that no decision needs it.\n", "sha256": "c8bc6aa705c7a668d3d4e2ea739bc648e6e1b85624dd676276760fbb8b18c1b7"} {"id": "source-5a44a59fc42cffc2", "source_path": "research/manuscript/07_lp.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Linear synthesis of coherent recovery certificates}\\label{ch:lp}\n\\section{The fixed-snapshot feasibility problem}\nAfter the proposer freezes an old and a new snapshot, the verifier receives rational matrices for both chain presentations, their encodings, and their homology coordinates. Unknowns are recovery-map blocks $F_n$, coherence blocks where the selected finite presentation permits them, observable entries $R_{ij}$, and anchor witnesses $W$.\n\nThe constraints are\n\\begin{align}\n d_CF-Fd_D&=0,\\\\\n R-\\pi_CF_qi_D&=0,\\\\\n F_qZ_Q-Z_P-d_{C,q+1}W&=0,\\\\\n R_{ij}&\\ge0,\\qquad \\sum_iR_{ij}=1.\n\\end{align}\nEvery coefficient multiplying an unknown is fixed. These are linear equalities and inequalities. An allowed block pattern, prescribed zeros, or fixed action on designated subspaces adds linear constraints. A cone seam introduces $F_Xk'-kF_S-dH-Hd'=0$, which is also linear for fixed seam maps.\n\n\\begin{theorem}[Positive-coherent synthesis]\\label{thm:lp}\nFor frozen finite rational presentations, existence of a positive anchored recovery with a prescribed linear block template is a rational linear-feasibility problem. A rational feasible point is an exact finite certificate checkable by rational matrix arithmetic and sign tests.\n\\end{theorem}\n\\begin{proof}\nThe displayed system includes the chain-map, homology-coordinate, anchor, and stochasticity conditions in Definition~\\ref{def:positive-recovery}. Each is linear in the unknown entries. Coherence and template equations remain linear by the fixed-snapshot assumption. Therefore their feasible set is a rational polyhedron. Substituting any proposed rational point verifies each equality and inequality exactly. Standard elimination or linear programming can search for a point; the correctness of the checker does not depend on trusting the search algorithm.\n\\end{proof}\nThe theorem is about the explicit matrix input size. Constructing a small presentation of a large model, enumerating its state space, and estimating its experiment may be expensive. No polynomial-time claim in neural parameter count or in the description length of an arbitrary program follows.\n\n\\section{Completeness without a locality template}\nOne may worry that the homological machinery excludes otherwise valid recoveries. Over a field, it does not, provided no additional structural restrictions are imposed.\n\n\\begin{theorem}[Unrestricted finite completeness]\\label{thm:complete}\nFor two anchored finite chain presentations with chosen observable degree $q$, the following are equivalent:\n\\begin{enumerate}[label=(\\roman*),itemsep=2pt]\n\\item A stochastic matrix $R$ satisfies $P=RQ$.\n\\item A positive anchored recovery exists, with no additional block or locality restrictions and no prescribed maps in other homological degrees.\n\\end{enumerate}\n\\end{theorem}\n\\begin{proof}\nThe implication from (ii) to (i) is Theorem~\\ref{thm:realization}. For the converse, choose finite splittings of both complexes. Define a graded homology map equal to $R$ in degree $q$ and zero in other degrees. Proposition~\\ref{prop:normal} lifts it to a chain map $F=i_Crp_D$.\n\nThe encoding discrepancy $\\Delta=F_qZ_Q-Z_P$ is a matrix of cycles. Its homology coordinates are $RQ-P=0$. Therefore every column lies in $B_q(C)=\\im d_{q+1}$. Choose a preimage for each column and collect them into $W$. This gives the anchor equation. The induced observable map is the original stochastic $R$, so positivity and normalization hold. All choices can be made by finite elimination. For rational input and rational $R$, they are rational.\n\\end{proof}\nThis result has a useful double interpretation. It shows that a coherent presentation can serve as a complete certificate language for the finite operational condition. It also shows that the unstructured condition is not a new statistical order: it reduces to Blackwell recovery. The new proposal lies in how recoveries are presented, glued, reused, and combined with progress evidence.\n\n\\section{The effect of structural restrictions}\nA locality template can require a model component to recover only from a specified set of successor components. Such a restriction may make a useful large certificate sparse. It can also reject a behaviorally valid but nonlocal recovery. Failure of the restricted feasibility problem therefore does not prove that the successor has lost capability. It proves only that the requested certificate shape is unavailable.\n\nThe distinction is important for an adaptive system that edits its own architecture. It should be able to distinguish at least three rejection reasons: a violated chain identity, a failure of stochastic recoverability on the contract, and a failure of a chosen implementation or resource template. Only the second directly witnesses information loss, and even then an optimal infeasibility or separation certificate is needed rather than the failure of a heuristic search.\n\n\\section{Approximate semantic synthesis}\nA less restrictive program minimizes a scalar $\\varepsilon$ with constraints\n\\[\nR\\ge0,\\quad\\mathbf1^\\top R=\\mathbf1^\\top,\\quad\n-P_{y\\theta}+(RQ)_{y\\theta}\\le z_{y\\theta},\\quad\nP_{y\\theta}-(RQ)_{y\\theta}\\le z_{y\\theta},\n\\]\n\\[\nz_{y\\theta}\\ge0,\\qquad\\sum_y z_{y\\theta}\\le2\\varepsilon\\quad\\text{for each }\\theta.\n\\]\nThis is a linear program for directed deficiency. One may couple it to the coherent structural constraints while replacing the exact anchor by a specified observable residual. The residual belongs to the semantic ledger, not to an unspecified ``approximately zero'' homotopy.\n\nThe optional implementation uses SciPy to search for a recovery, rationalizes the returned columns, normalizes them exactly, and recomputes the defect with rational arithmetic. The resulting rational matrix and its defect are valid primal certificates. The numerical optimizer's reported objective is not treated as an exact optimality proof.\n\n\\section{Certificate normalization and proof reuse}\nIf a proposed recovery is changed by a chain homotopy, its observable recovery remains fixed. The exact anchor can be updated by adding $HZ_Q$. This permits a verifier to separate an already checked semantic witness from a newly supplied internal representation witness. In an engineered system, the semantic object can be cached while a compiler, memory layout, or modular implementation changes.\n\nThe gain is potentially significant because the behavioral guarantee refers to all bounded decision rules in the contract, rather than a finite list of previously benchmarked heads. But the theorem supplies no automatic cache hit: the new implementation still has to provide the homotopy or another valid certificate. Producing short witnesses is part of the research challenge.\n\n\\status{The linear-feasibility and completeness statements are proved for explicit finite presentations. They do not solve arbitrary program equivalence, discover a model's semantics without evidence, or establish a scalable verifier for frontier neural systems.}\n\n\\section{Implementation boundary}\nThe delivered optional numerical solver implements finite experiment recovery, not a generic compiler for all chain templates. The general template feasibility and completeness results are mathematical theorems; a broad template compiler and a learned-model presentation compiler are not included. Small coherent and anchored examples are checked directly by the exact-arithmetic routines.\n", "sha256": "360c8aa5639b1ffc13dc5d20b9cb44cfcc5b11878e7729383824f212d36f4cd8"} {"id": "source-3da18c1977660a7e", "source_path": "research/manuscript/08_blackwell.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{The exact operational meaning of recovery}\\label{ch:blackwell}\n\\section{Total variation as a decision bound}\n\\begin{lemma}[Bounded payoff inequality]\\label{lem:tv}\nFor probability vectors $p,q$ and numbers $f_i\\in[0,1]$,\n\\[\n\\left|\\sum_i f_i(p_i-q_i)\\right|\\le\\TV(p,q).\n\\]\nFor a stochastic matrix $R$, $\\TV(Rp,Rq)\\le\\TV(p,q)$.\n\\end{lemma}\n\\begin{proof}\nLet $v=p-q$, whose entries sum to zero. The sum of its positive entries and the absolute sum of its negative entries both equal $\\norm v_1/2$. A weighted sum with weights in $[0,1]$ lies between the negative and positive totals. For the second assertion,\n\\[\n\\norm{Rv}_1\\le\\sum_{i,j}R_{ij}|v_j|=\\sum_j|v_j|,\n\\]\nbecause the columns of $R$ sum to one.\n\\end{proof}\n\\begin{theorem}[Universal retention bound]\\label{thm:retention}\nIf $\\varepsilon(P,Q;R)\\le\\varepsilon$, then for every prior, finite action set, and utility in $[0,1]$,\n\\[\n\\boxed{V(Q;\\mu,u)\\ge V(P;\\mu,u)-\\varepsilon.}\n\\]\nThe same bound holds for simulating any particular predecessor decision rule by applying $R$ and then that rule.\n\\end{theorem}\n\\begin{proof}\nLet $\\pi$ be a predecessor rule. The successor can simulate it using $\\pi R$. Conditional on $\\theta$, the expected reward under this simulation differs from the predecessor reward by at most $\\TV(P_{\\cdot\\theta},RQ_{\\cdot\\theta})$, by Lemma~\\ref{lem:tv}. Average with respect to $\\mu$. The resulting loss is at most $\\varepsilon$. Choosing an optimal predecessor rule and observing that the successor optimum is at least the value of its simulated rule gives the theorem.\n\\end{proof}\nThe important quantifier is over every utility table, including a task selected after certification. This is possible because a recovery reconstructs the whole declared experiment, not just one average score.\n\n\\section{A self-contained finite comparison theorem}\nThe converse is the classical finite Blackwell theorem \\cite{blackwell,rauh}. A proof is included to expose exactly what the operational claim requires.\n\n\\begin{theorem}[Finite comparison]\\label{thm:blackwell}\nFor finite experiments $P,Q$, the following are equivalent:\n\\begin{enumerate}[label=(\\roman*),itemsep=2pt]\n\\item $P=RQ$ for some stochastic $R$.\n\\item $V(Q;\\mu,u)\\ge V(P;\\mu,u)$ for every finite decision problem on $\\Theta$.\n\\end{enumerate}\nIf (i) fails, a decision problem with predecessor outputs as actions separates the two experiments.\n\\end{theorem}\n\\begin{proof}\nThe forward implication is Theorem~\\ref{thm:retention} with zero defect. For the converse, consider the compact convex set\n\\[\n\\mathcal K=\\{RQ:R\\text{ is column-stochastic}\\}.\n\\]\nCompactness follows because stochastic matrices form a closed bounded subset of a finite-dimensional space and multiplication by $Q$ is continuous. Suppose $P\\notin\\mathcal K$, and let $Z_*$ minimize Euclidean distance from $P$ to $\\mathcal K$. Set $W=P-Z_*\\ne0$.\n\nFor any $Z\\in\\mathcal K$, the segment $Z_*+t(Z-Z_*)$ lies in $\\mathcal K$ for $t\\in[0,1]$. Differentiating its squared distance to $P$ at $t=0$ gives\n\\[\n\\langle W,Z-Z_*\\rangle\\le0.\n\\]\nHence\n\\[\n\\langle W,P\\rangle=\\langle W,Z_*\\rangle+\\norm W_2^2\n>\\sup_{Z\\in\\mathcal K}\\langle W,Z\\rangle.\n\\]\nLet the action set be the old output alphabet. Use a uniform prior on $\\Theta$, and rescale all entries of $W$ by the same positive affine transformation into $[0,1]$ to obtain $u(a,\\theta)$. This rescaling preserves the strict separation because every experiment has unit column sums. A constant $W$ could not separate normalized experiments, so the rescaling has nonzero range.\n\nThe old experiment can use the identity rule, whose expected payoff is $|\\Theta|^{-1}\\langle u,P\\rangle$. Every new decision rule is a stochastic matrix $R$ and has payoff $|\\Theta|^{-1}\\langle u,RQ\\rangle$. Separation says that even the old identity rule beats the best new rule. This contradicts (ii).\n\\end{proof}\nThe separating task is a mathematical witness to information loss. It need not be a naturally named benchmark. It also need not be computationally cheap to identify if the experiment is implicit or very large.\n\n\\section{Why this does not protect every imaginable property}\nA downstream rule may use only the certified observable. If a safety-relevant property depends on an omitted hidden state, an unmodeled interaction, or a resource budget not represented in the utility, the theorem does not cover it. ``All tasks'' is always read with its state-space and interface qualifiers.\n\nLikewise, the theorem protects the existence of a simulated predecessor strategy. It does not force an autonomous successor to choose that strategy. A deployed policy must either use the certified decoder, be evaluated directly, or be controlled by another specification. Capability retention is not the same as goal preservation.\n\n\\section{A useful interpretation for model migration}\nA migration certificate can be understood as a portable compatibility layer. It says that anything a downstream decision maker could have done using the old observation can still be done after migration, with at most the declared defect. New downstream tasks do not need to be anticipated one at a time, provided they stay within the contract.\n\nThe price of this strength is that many intuitively good updates are incomparable. A specialized model may solve one target task better while discarding an old capability. The correct verdict is not that specialization is forbidden, but that it does not satisfy a universal retention contract. A different contract can authorize that tradeoff explicitly. A rigorous framework should expose the tradeoff rather than call every target-score increase a universal improvement.\n", "sha256": "8d9742cb5e2173ec3d247cd5cc815db2fd5e81cec09429cff108b8ad7376315f"} {"id": "source-cd24f107723ac025", "source_path": "research/manuscript/09_innovation.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{A complete strict-innovation witness}\\label{ch:innovation}\n\\section{Retention is not progress}\nExact recovery permits equality. A system that repeatedly replaces its internal machinery with equivalent implementations can satisfy every retention condition while acquiring no new information. A rigorous RSI claim must therefore include an additional progress witness. For resource optimization that witness can be a lower cost at equal capability. For information improvement, a stronger and surprisingly compact detector is available.\n\nAssume $P=RQ$ and fix a full-support prior $\\mu_\\theta>0$. Generate a state $\\theta\\sim\\mu$, a successor observation $Y\\sim Q(\\cdot\\mid\\theta)$, and a recovered predecessor observation $X\\sim R(\\cdot\\mid Y)$. Define the posterior vectors\n\\[\na(Y)=\\Prob(\\theta=\\cdot\\mid Y),\\qquad\nb(X)=\\Prob(\\theta=\\cdot\\mid X).\n\\]\nBecause the recovery does not inspect the state, $\\theta\\to Y\\to X$ is a Markov chain. The posterior tower identity is\n\\begin{equation}\\label{eq:posterior-tower}\n\\E[a(Y)\\mid X]=b(X).\n\\end{equation}\n\n\\section{The quadratic decision family}\nFor an action $v$ in the probability simplex on $\\Theta$, define the normalized quadratic reward\n\\[\ns(v,\\theta)=\\tfrac12\\bigl(1+2v_\\theta-\\norm v_2^2\\bigr).\n\\]\nIt lies in $[0,1]$. At posterior $p$, the conditional expected reward is\n\\[\n\\tfrac12\\bigl(1+\\norm p_2^2-\\norm{v-p}_2^2\\bigr),\n\\]\nso the unique optimal action is $v=p$. Let\n\\[\n\\mathcal B(P,\\mu)=\\tfrac12\\left(1+\\E\\norm{\\Prob(\\theta=\\cdot\\mid X)}_2^2\\right)\n\\]\nbe its optimal value. Although the action simplex is continuous, only finitely many posterior actions are needed to attain the optima for a fixed pair of finite experiments. Thus all comparisons below can also be expressed as finite decision problems.\n\n\\begin{theorem}[Strict-innovation identity]\\label{thm:innovation}\nFor exact recovery $P=RQ$ and a full-support prior,\n\\[\nJ_\\mu(P,Q)=\\E\\norm{a(Y)-b(X)}_2^2\n=2\\bigl[\\mathcal B(Q,\\mu)-\\mathcal B(P,\\mu)\\bigr]\\ge0.\n\\]\nThe number $J_\\mu$ is independent of which exact recovery coupling $R$ is used. Moreover,\n\\[\nJ_\\mu=0\\quad\\Longleftrightarrow\\quad P\\text{ and }Q\\text{ recover each other stochastically}.\n\\]\nThus every strict exact refinement is detected by this one bounded scoring family under every full-support prior.\n\\end{theorem}\n\\begin{proof}\nExpand the square. By~\\eqref{eq:posterior-tower},\n\\[\n\\E\\langle a(Y),b(X)\\rangle\n=\\E\\langle\\E[a(Y)\\mid X],b(X)\\rangle\n=\\E\\norm{b(X)}_2^2.\n\\]\nConsequently $J_\\mu=\\E\\norm a_2^2-\\E\\norm b_2^2$, which is twice the displayed difference of quadratic values. Its right side depends only on $P,Q,\\mu$, not on the coupling.\n\nSuppose $J_\\mu=0$. Then $a(Y)=b(X)$ for every pair $(X,Y)$ having positive joint probability. Define $S(y\\mid x)=\\Prob(Y=y\\mid X=x)$, filling any zero-probability $x$ column arbitrarily. The equality of posteriors says that $\\theta$ and $Y$ are conditionally independent given $X$. Therefore\n\\[\n\\mu_\\theta Q_{y\\theta}=\\sum_x\\mu_\\theta P_{x\\theta}S_{yx}.\n\\]\nFull support permits division by $\\mu_\\theta$, giving $Q=SP$. Thus the experiments recover each other.\n\nConversely, mutual recovery gives equal optimal values for every finite decision problem by Theorem~\\ref{thm:blackwell}. Restrict the quadratic actions to the finite union of posterior vectors occurring under $P$ and $Q$. This finite set contains an optimal action at every output of both experiments, so their quadratic values agree. The identity then yields $J_\\mu=0$.\n\\end{proof}\nThe theorem is a decision-theoretic consequence of classical experiment comparison and proper scoring. Its use here is to turn the progress obligation into a complete strictness test within the finite exact order, rather than to claim a new universal law of intelligence.\n\n\\section{A quantitative irreversibility certificate}\n\\begin{corollary}[Reverse deficiency bound]\\label{cor:reverse}\nUnder the same assumptions,\n\\[\n\\delta(Q\\mid P)\\ge \\frac{J_\\mu(P,Q)}{2}.\n\\]\n\\end{corollary}\n\\begin{proof}\nAny reverse recovery with defect $\\eta$ gives, by Theorem~\\ref{thm:retention}, $\\mathcal B(P,\\mu)\\ge\\mathcal B(Q,\\mu)-\\eta$. The value difference is $J_\\mu/2$. Take the infimum over reverse recoveries.\n\\end{proof}\nA positive witness therefore establishes more than ``the new representation looks different.'' It lower-bounds how accurately the predecessor could simulate the successor. This is a quantitative distinction between reversible re-expression and genuine information refinement.\n\n\\section{A single innovation quantity controls every decision advantage}\nThe same posterior coupling produces an explicit approximate reverse recovery, not only a lower bound on its impossibility. Write $s=|\\Theta|$, $\\mu_{\\min}=\\min_\\theta\\mu_\\theta$, and define $S(y\\mid x)=\\Prob(Y=y\\mid X=x)$ under the exact recovery coupling.\n\n\\begin{theorem}[Innovation--reconstruction sandwich]\\label{thm:sandwich}\nUnder exact forward recovery and a full-support prior,\n\\[\n\\frac{J_\\mu}{2}\\le\\delta(Q\\mid P)\n\\le\\max_\\theta\\TV(Q_{\\cdot\\theta},(SP)_{\\cdot\\theta})\n\\le\\min\\left\\{1,\\frac{\\sqrt{J_\\mu}}{2\\mu_{\\min}}\\right\\}.\n\\]\nMoreover, for every reward table in $[0,1]$ with this same prior,\n\\[\n0\\le V(Q;\\mu,u)-V(P;\\mu,u)\n\\le\\min\\left\\{1,\\frac{\\sqrt{sJ_\\mu}}{2}\\right\\}.\n\\]\n\\end{theorem}\n\\begin{proof}\nLet $p(x,y)$ denote the coupling's output law. The difference between the actual and reverse-recovered joint laws satisfies\n\\[\n\\mu_\\theta[Q_{y\\theta}-(SP)_{y\\theta}]\n=\\sum_xp(x,y)[a(y)_\\theta-b(x)_\\theta].\n\\]\nSum absolute values over $y$, use the triangle inequality, and then Cauchy--Schwarz to obtain\n\\[\n\\TV(Q_{\\cdot\\theta},(SP)_{\\cdot\\theta})\n\\le\\frac{\\E|a(Y)_\\theta-b(X)_\\theta|}{2\\mu_\\theta}\n\\le\\frac{\\sqrt{J_\\mu}}{2\\mu_\\theta}.\n\\]\nThe lower bound was proved in Corollary~\\ref{cor:reverse}.\n\nFor the decision bound, write $g(p)=\\max_a\\sum_\\theta u(a,\\theta)p_\\theta$. For probability vectors $p,q$, the bounded payoff inequality gives $|g(p)-g(q)|\\le\\norm{p-q}_1/2$. Therefore the difference of optimal values is at most\n\\[\n\\tfrac12\\E\\norm{a(Y)-b(X)}_1\n\\le\\tfrac12\\sqrt{s\\E\\norm{a(Y)-b(X)}_2^2}.\n\\]\nExact forward recovery gives nonnegativity; values in $[0,1]$ give the trivial cap of one.\n\\end{proof}\nThis quantitative connection clarifies the operational meaning of the innovation witness. Small $J_\\mu$ bounds every decision improvement under the chosen prior and supplies a concrete reverse simulator. A tiny prior weight weakens the worst-state bound, as it must: information about a rare state can contribute little to an average score while being important under a different prior. The theorem does not silently turn a fixed-prior numerical score into a distribution-free claim.\n\n\\section{An exact rational example}\nLet $P$ be a binary symmetric channel with error $1/4$, and $Q$ one with error $1/10$. A recovery is a binary symmetric channel of error $3/16$, since cascading binary symmetric channels with errors $q$ and $r$ gives $q+r-2qr$. Thus\n\\[\n\\frac1{10}+\\frac3{16}-2\\cdot\\frac1{10}\\cdot\\frac3{16}=\\frac14.\n\\]\nFor a uniform prior, optimal classification accuracy rises from $3/4$ to $9/10$. The unnormalized posterior squared norms are $5/8$ and $41/50$, so\n\\[\nJ=\\frac{41}{50}-\\frac58=\\frac{39}{200},\\qquad\n\\mathcal B(Q)-\\mathcal B(P)=\\frac{39}{400}.\n\\]\nThe supplied exact-arithmetic code verifies the recovery, the posterior-dispersion identity, and the same identity under a nonuniform prior. A numerical LP also finds rational witnesses, but those numerical searches are not needed for the proof.\n\n\\section{What the witness does and does not measure}\nThe witness measures additional information usable for some decision in the contract. It does not measure the subjective value of that decision, the cost of decoding it, or the quality of a particular default response. A domain-specific utility may care about only a small part of the information improvement. Conversely, a model can improve a default head while leaving the underlying experiment equivalent, in which case $J$ is zero and the progress is policy improvement rather than information acquisition.\n\nThis distinction is desirable. It prevents a system from claiming a new capability merely because it changed coordinates, while allowing a separate resource or policy certificate to recognize useful engineering progress that does not change the Blackwell class.\n", "sha256": "8de473c50b9ef28ad8893db52e70c64217bdb98c2d86ce4a20f6844bab224025"} {"id": "source-19f00940ebc9937d", "source_path": "research/manuscript/10_debt.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Approximate lineages and an affine risk calculus}\\label{ch:debt}\n\\section{Why small losses cannot be ignored}\nAn approximate recovery can tolerate a small task loss. Repeating such tolerances without a cumulative ledger permits substantial degradation. The predecessor-relative condition is not enough by itself: a sequence of individually modest errors can move far from the original system. The source's closed-loop geometry suggests that the object to audit is the entire path of transformations.\n\nLet $P_0,\\ldots,P_T$ be experiments on a common state space, and let $R_t$ recover outputs of $P_{t-1}$ from outputs of $P_t$. Set\n\\[\n\\Delta_t=P_{t-1}-R_tP_t,\\qquad\n\\varepsilon_t=\\max_\\theta\\TV((P_{t-1})_{\\cdot\\theta},(R_tP_t)_{\\cdot\\theta}).\n\\]\nEvery column of $\\Delta_t$ sums to zero.\n\n\\section{Contraction of signed errors}\nFor a stochastic matrix $R$, define its Dobrushin coefficient\n\\[\n\\kappa(R)=\\max_{i,j}\\TV(R_{\\cdot i},R_{\\cdot j})\\in[0,1].\n\\]\n\\begin{lemma}\\label{lem:dobrushin}\nFor every zero-sum vector $v$,\n\\[\n\\tfrac12\\norm{Rv}_1\\le\\kappa(R)\\tfrac12\\norm v_1.\n\\]\nAlso $\\kappa(RS)\\le\\kappa(R)\\kappa(S)$.\n\\end{lemma}\n\\begin{proof}\nWrite $v=v^+-v^-$ with disjoint nonnegative parts. Their common total is $a=\\norm v_1/2$. The result is trivial when $a=0$. Otherwise $p=v^+/a$ and $q=v^-/a$ are distributions. The vectors $Rp,Rq$ are convex combinations of columns of $R$. By convexity of total variation, their distance is at most the largest distance between two columns, namely $\\kappa(R)$. Multiplying by $a$ proves the first assertion. Apply it to the difference of any two columns of $S$ to obtain the second.\n\\end{proof}\n\n\\begin{theorem}[Contraction-weighted lineage debt]\\label{thm:debt}\nLet $R_{1:t}=R_1\\cdots R_t$, with $R_{1:0}=I$. Then\n\\[\nP_0-R_{1:T}P_T=\\sum_{t=1}^TR_{1:t-1}\\Delta_t,\n\\]\nand therefore\n\\begin{equation}\\label{eq:debt}\n\\boxed{\\delta(P_0\\mid P_T)\\le E_T:=\n\\min\\left\\{1,\\sum_{t=1}^T\\kappa(R_{1:t-1})\\varepsilon_t\\right\\}.}\n\\end{equation}\nIn particular, $E_T\\le\\min\\{1,\\sum_t\\varepsilon_t\\}$. Every bounded decision value of the initial system remains attainable up to loss $E_T$.\n\\end{theorem}\n\\begin{proof}\nInsert and subtract $R_1P_1$, then $R_1R_2P_2$, and continue. This telescopes to the exact matrix identity. Apply Lemma~\\ref{lem:dobrushin} to each zero-sum column error and use the triangle inequality. The product recovery is stochastic, so its defect is an admissible upper bound on deficiency. Total variation never exceeds one. The decision conclusion follows from Theorem~\\ref{thm:retention}.\n\\end{proof}\nThe coefficient can improve the additive bound when an early recovery deliberately discards distinctions that later errors affect. It cannot manufacture missing information. It merely recognizes that some successor-level discrepancies are invisible to the original contract.\n\n\\section{An exact two-step equality case}\nFor binary symmetric experiments, choose old error $9/20$, middle error $3/10$, and final error $1/10$. Use recovery error $1/10$ at both steps. The first semantic defect is $11/100$ and the second is $3/25$. The first recovery has contraction $4/5$, so the weighted debt is\n\\[\n\\frac{11}{100}+\\frac45\\frac3{25}=\\frac{103}{500}.\n\\]\nThe composite recovery has binary-symmetric error $9/50$. Its output after the final experiment has error $61/250$, and the direct defect from $9/20$ is exactly $103/500$. Thus the weighted bound can be attained, not merely improved by cancellation in a specially chosen proof.\n\n\\section{Affine certificate pairs}\nA compact certificate may store an upper bound $\\kappa\\ge\\kappa(R)$ and a defect $\\varepsilon$. Its downstream error propagation is the affine map\n\\[\nf_{\\kappa,\\varepsilon}(x)=\\varepsilon+\\kappa x.\n\\]\nTwo sequential certificates compose by\n\\begin{equation}\\label{eq:affine-compose}\n(\\kappa_1,\\varepsilon_1)\\circ(\\kappa_2,\\varepsilon_2)\n=(\\kappa_1\\kappa_2,\\ \\varepsilon_1+\\kappa_1\\varepsilon_2).\n\\end{equation}\n\\begin{proposition}[Associative risk composition]\\label{prop:affine}\nComposition~\\eqref{eq:affine-compose} is associative, has identity $(1,0)$, and is monotone in both coordinates on $[0,1]\\times[0,\\infty)$. It is represented by ordinary upper-triangular matrix multiplication:\n\\[\n(\\kappa,\\varepsilon)\\longleftrightarrow\n\\begin{pmatrix}\\kappa&\\varepsilon\\\\0&1\\end{pmatrix}.\n\\]\n\\end{proposition}\n\\begin{proof}\nMatrix multiplication gives exactly~\\eqref{eq:affine-compose}; associativity and the identity follow. Nonnegative coefficients make each output coordinate nondecreasing in each input coordinate. Alternatively, direct composition gives $f_1(f_2(x))=\\varepsilon_1+\\kappa_1\\varepsilon_2+\\kappa_1\\kappa_2x$.\n\\end{proof}\nThis is a concrete scalar counterpart of the source's matrix-composition theme. It connects a coherent structural path with a noncommutative, but associative, accounting rule for its semantic uncertainty. The construction is a use of familiar affine algebra, not a claim that upper-triangular matrices were unknown.\n\n\\section{Multiple routes and recertification}\nA lineage graph may contain a direct recovery to an ancestor as well as a product of local recoveries. Each route has a valid defect bound. Taking the best certified route improves the guarantee without changing the model. Keeping several nondominated $(\\kappa,\\varepsilon)$ pairs is preferable to greedily minimizing only one coordinate, because a route with smaller immediate defect can have a worse downstream contraction.\n\nFor finite sets of pairs, union represents alternative certificates and pairwise composition represents concatenation. Removing a componentwise dominated pair is valid because composition is monotone. The lower envelope $\\min_{(\\kappa,\\varepsilon)}(\\varepsilon+\\kappa x)$ is the resulting best bound at downstream debt $x\\ge0$.\n\nA fresh direct certificate can reduce an overly pessimistic accumulated bound. It does not erase actual past losses, spent compute, or unrecorded changes in the evaluation contract. The ledger is a statement about recoverability of experiments, not a mechanism for reversing time.\n", "sha256": "59eeab0d42f8db5129c197b89dc61f3f059ae6083f2d35b75bdaaad77891044b"} {"id": "source-f60f79c6276db2c1", "source_path": "research/manuscript/11_uniform.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{Witness-adaptive certification from frozen experiments}\\label{ch:uniform}\n\\section{The adaptivity distinction}\nA candidate model selected on a validation sample can overfit that sample. This is a central issue in adaptive data analysis \\cite{dwork}. However, not every post-data choice is equally dangerous. Once the two model experiments are frozen, a uniform confidence envelope for those experiments can support a recovery witness, a task prior, and a decision rule chosen after seeing the data. The choices are protected by one simultaneous event.\n\nThe result below is deliberately elementary. Its sample cost grows with the outcome alphabet and the number of states. It is not a claim that arbitrary text distributions can be certified with a small holdout. Its value is a clean separation between adaptive \\emph{witness synthesis} and adaptive \\emph{candidate selection}.\n\n\\section{A bounded-sample lemma}\n\\begin{lemma}[Hoeffding bound]\\label{lem:hoeffding}\nFor independent $X_i\\in[a,b]$ with common mean $m$,\n\\[\n\\Prob\\!\\left(\\frac1n\\sum_iX_i-m\\ge t\\right)\n\\le\\exp\\!\\left(-\\frac{2nt^2}{(b-a)^2}\\right).\n\\]\nThe same bound holds for the lower tail.\n\\end{lemma}\n\\begin{proof}\nFor any random variable $X\\in[a,b]$, let $g(\\lambda)=\\log\\E e^{\\lambda(X-\\E X)}$. Then $g(0)=g'(0)=0$. Its second derivative is the variance under the exponentially tilted law, which is still supported on $[a,b]$. For a variable in this interval,\n\\[\n\\operatorname{Var}(X)\\le\\E(X-\\tfrac{a+b}{2})^2\\le\\tfrac{(b-a)^2}{4}.\n\\]\nIntegrating the second-derivative bound twice gives $g(\\lambda)\\le\\lambda^2(b-a)^2/8$. Independence multiplies moment-generating functions. Markov's inequality then bounds the upper tail by $\\exp(-\\lambda nt+n\\lambda^2(b-a)^2/8)$. Optimizing at $\\lambda=4t/(b-a)^2$ proves the result. Replace $X$ by $-X$ for the lower tail.\n\\end{proof}\n\n\\begin{theorem}[Uniform finite-channel envelope]\\label{thm:channel-envelope}\nFor each of $s$ states, observe $n$ independent outputs from a frozen $m$-outcome experiment $P$. Let $\\widehat P$ be its empirical channel. With probability at least $1-\\beta$,\n\\[\n\\max_\\theta\\TV(P_{\\cdot\\theta},\\widehat P_{\\cdot\\theta})\\le\n\\rho(m,s,n,\\beta),\n\\]\nwhere\n\\[\n\\rho(m,s,n,\\beta)=\\min\\left\\{1,\n\\sqrt{\\frac{(m+1)\\log2+\\log(s/\\beta)}{2n}}\\right\\}.\n\\]\n\\end{theorem}\n\\begin{proof}\nFor a fixed state and output subset $A$, the empirical frequency of $A$ is an average of Bernoulli variables. Lemma~\\ref{lem:hoeffding} bounds the probability of absolute error exceeding $r$ by $2e^{-2nr^2}$. There are $2^m$ subsets and $s$ states. A union bound gives $s2^{m+1}e^{-2nr^2}$.\n\nFor two probability vectors, total variation equals the largest absolute difference in mass assigned to a subset: choose the subset on which $p_i-q_i$ is positive. Therefore the union event controls the desired total variation. Substituting the displayed radius makes its failure probability at most $\\beta$. If that radius exceeds one, the deterministic bound $\\TV\\le1$ proves the clipped version.\n\\end{proof}\nSamples across different states need not be independent for this union bound. Independence within each state's sample, conditional on the frozen experiment, is sufficient. Unequal state sample sizes can be handled by using state-specific radii or their maximum.\n\n\\section{A data-selected recovery remains valid}\nLet $a_P,a_Q$ be valid uniform radii for the old and new experiments on a joint event $\\mathcal E$. Conditional on $\\mathcal E$, the following statements are deterministic.\n\n\\begin{theorem}[Witness-adaptive recovery and task envelope]\\label{thm:witness-adaptive}\nSimultaneously for every stochastic recovery $R$, every prior $\\mu$, and every bounded decision problem,\n\\begin{align}\n\\varepsilon(P,Q;R)&\\le\n\\max_\\theta\\TV(\\widehat P_{\\cdot\\theta},R\\widehat Q_{\\cdot\\theta})+a_P+a_Q,\\label{eq:emp-recovery}\\\\\n|V(P;\\mu,u)-V(\\widehat P;\\mu,u)|&\\le a_P,\\label{eq:emp-value}\\\\\nV(Q;\\mu,u)-V(P;\\mu,u)&\\ge\nV(\\widehat Q;\\mu,u)-V(\\widehat P;\\mu,u)-a_P-a_Q.\\label{eq:emp-gain}\n\\end{align}\nThese statements remain valid when $R$, $\\mu$, or the utility table is chosen using the same empirical channels.\n\\end{theorem}\n\\begin{proof}\nFor each state, insert $\\widehat P$ and $R\\widehat Q$ between $P$ and $RQ$. The triangle inequality gives the old radius, the empirical recovery defect, and $\\TV(R\\widehat Q,RQ)$. Data processing bounds the last term by the new radius. This proves~\\eqref{eq:emp-recovery} for every stochastic $R$ on the same event.\n\nFor a fixed decision rule, Lemma~\\ref{lem:tv} bounds the difference between true and empirical expected reward by the uniform radius, independently of the prior or utility table. Maximizing over rules preserves the bound, proving~\\eqref{eq:emp-value}. Apply it to both experiments and subtract to obtain~\\eqref{eq:emp-gain}. No event indexed by the post-data choices was introduced. The statements already hold for all of them simultaneously.\n\\end{proof}\nA recovery-search algorithm may therefore use the confirmation channel estimates without requiring a second holdout solely for the recovery witness. The theorem does not authorize changing $P$ or $Q$ after inspecting those estimates. That would require a confidence event uniform over the candidate family or a separate fresh sample.\n\n\\section{Discovering an improvement witness on the same envelope}\nApply~\\eqref{eq:emp-gain} to the normalized quadratic decision family. If\n\\[\n\\mathcal B(\\widehat Q,\\mu)-\\mathcal B(\\widehat P,\\mu)>a_P+a_Q+\\gamma,\n\\]\nthen the true quadratic value improves by more than $\\gamma$. The prior may be selected after seeing the empirical channels because the value envelope is uniform over priors. A discovered prior can reveal where the new information is useful; it is not automatically the deployment prior, which should be declared separately.\n\nWhen exact recovery is available from a structural proof, a positive true quadratic gain proves strict refinement. When recovery is only approximate, the conclusion is instead an approximate all-task retention bound plus a strict gain on the selected quadratic task. Those are strong but different statements. Finite noisy samples generally do not certify a zero TV defect using this radius bound.\n\n\\section{What controls the sample cost}\nTo obtain radius at most $r$ by this elementary argument, it is enough to use\n\\[\nn\\ge\\frac{(m+1)\\log2+\\log(s/\\beta)}{2r^2}\n\\]\noutputs per state. The count is linear in the alphabet size, not merely its bit length. A homology space of $k$ binary logical bits may have $2^k$ classical labels, so a topological encoding does not remove this cost.\n\nA proven sufficient quotient can reduce the alphabet without changing decision capability, by Theorem~\\ref{thm:sufficient}. A heuristic compression cannot claim the same benefit without its recovery obligation. In this sense, semantic sufficiency and coherent representation reuse offer a possible route to practical certification, while the sample bound makes the bottleneck explicit.\n", "sha256": "bad7387a53ca5e6f7361691c6feeb322357f942fe72e3e708bd4974974c77d11"} {"id": "source-756a8f1c5efa36c4", "source_path": "research/manuscript/12_lineage.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{The architecture-independent lineage theorem}\\label{ch:lineage}\n\\section{Three obligations, not one score}\nA proposed transition can carry three distinct pieces of evidence. The structural certificate verifies chain identities and coherent assembly in the selected presentation. The semantic certificate gives a stochastic recovery and its defect. The progress certificate gives a strict gain in a declared objective, which may be a quadratic information witness, a deployed task utility, or a normalized resource objective.\n\nThe proposer is allowed to be arbitrary and adaptive. It may use previous proposals, rejections, accepted snapshots, training data, and its own generated code. None of those freedoms changes the meaning of the certificate. The checking kernel, sampling rules, contract, and budget accounting are the trust boundary.\n\nNumber every opened confirmation attempt by $j=1,2,\\ldots$, including attempts that fail. One convenient risk allocation is\n\\[\n\\alpha_j=\\frac{\\alpha}{j(j+1)},\\qquad\n\\sum_{j=1}^\\infty\\alpha_j=\\alpha.\n\\]\nCounting only accepted updates would be a different protocol and needs its own conditional argument. The reference implementation avoids that ambiguity by charging every opened attempt.\n\n\\section{A fixed-sample progress gate}\nFor a candidate frozen before its fresh confirmation data, let $D_{j,i}\\in[-1,1]$ be paired utility differences with conditional mean $\\Delta_j$. A fixed sample size $n_j$ is chosen before the confirmation sample. Define\n\\[\n\\widehat\\Delta_j=\\frac1{n_j}\\sum_iD_{j,i},\\qquad\nr_j=\\sqrt{\\frac{2\\log(1/\\alpha_j)}{n_j}}.\n\\]\nAccept the progress claim only if\n\\begin{equation}\\label{eq:fixed-gate}\n\\widehat\\Delta_j-r_j>\\gamma_j,\n\\end{equation}\nwhere the required gain $\\gamma_j\\ge0$ is declared before sampling. Conditional Hoeffding gives\n\\[\n\\Prob(\\text{accept and }\\Delta_j\\le\\gamma_j\\mid\\mathcal F_{j-1})\\le\\alpha_j.\n\\]\nThe same pair of model marginals may be coupled using common randomness to reduce variance; independence between the two outputs in a pair is unnecessary. Freshness and independence of pairs conditional on the frozen experiment are the relevant requirements.\n\n\\section{An anytime-valid alternative}\nFor optional stopping, a fixed-sample bound should not simply be inspected repeatedly. Let $\\lambda_i\\ge0$ be predictable betting rates and define\n\\[\nM_n=\\prod_{i=1}^n\\exp\\left(\\lambda_i(D_i-\\gamma)-\\frac{\\lambda_i^2}{2}\\right).\n\\]\nUnder the conditional null $\\E[D_i\\mid\\mathcal G_{i-1}]\\le\\gamma$, Hoeffding's moment bound shows that $M_n$ is a nonnegative supermartingale with initial value one. Accept when $M_n\\ge1/\\alpha_j$.\n\n\\begin{lemma}[Anytime crossing bound]\\label{lem:ville}\nUnder the stated null, $\\Prob(\\sup_nM_n\\ge1/\\alpha_j)\\le\\alpha_j$.\n\\end{lemma}\n\\begin{proof}\nStop at the first crossing or at a fixed horizon $N$. Iterated conditional expectation keeps the expected stopped supermartingale at most one. On a crossing by time $N$, its stopped value is at least $1/\\alpha_j$, so that crossing has probability at most $\\alpha_j$. Let $N$ increase and use continuity of probability for an increasing sequence of events.\n\\end{proof}\nTime-uniform inference is established prior art \\cite{howard}. Its role here is to make the stopping rule explicit, not to claim a new concentration principle.\n\n\\section{The combined theorem}\nLet accepted snapshots be $P_0,P_1,\\ldots$, indexed by acceptance order. Their confirmation attempts need not be consecutive. Suppose a single conditional failure budget $b_j$ is assigned to each attempt, with $\\sum_jb_j\\le\\alpha$. An attempt may use exact semantic evidence and a scalar progress gate, or a joint frozen-channel envelope that certifies both retention and gain.\n\n\\begin{theorem}[Certified adaptive lineage]\\label{thm:lineage}\nAssume the following conditions.\n\\begin{enumerate}[label=(\\alph*),itemsep=2pt]\n\\item The state space and declared objective remain fixed within the lineage epoch. Candidate snapshots are frozen before any fresh data used by their confirmation guarantee.\n\\item Every committed update has a valid structural certificate in its stated presentation and a stochastic recovery $R_t$ with claimed defect $\\varepsilon_t$.\n\\item Conditional on all prior history, the probability that attempt $j$ commits with either an invalid semantic bound or an invalid declared progress bound is at most $b_j$.\n\\item The checking kernel executes the declared rules, and commits exactly the snapshot identified by the certificate.\n\\end{enumerate}\nThen, with probability at least $1-\\alpha$, simultaneously for every accepted finite prefix $T$:\n\\begin{enumerate}[label=(\\roman*),itemsep=2pt]\n\\item the original experiment is recoverable from $P_T$ with the contraction-weighted debt $E_T$ in Theorem~\\ref{thm:debt};\n\\item every bounded decision value on the contract satisfies $V(P_T;\\mu,u)\\ge V(P_0;\\mu,u)-E_T$;\n\\item if each accepted progress claim is $U(P_t)-U(P_{t-1})>\\gamma_t$ for the same scalar objective $U$, then\n\\[\nU(P_T)-U(P_0)>\\sum_{t=1}^T\\gamma_t\\quad(T\\ge1);\n\\]\n\\item replacing structural representatives by certified homotopic representatives changes neither the induced recovery nor its operational guarantee.\n\\end{enumerate}\nThe proposer may change its architecture and proposal procedure at any time, subject to these conditions.\n\\end{theorem}\n\\begin{proof}\nLet $B_j$ denote a bad commitment at attempt $j$. Conditional validity implies $\\Prob(B_j)=\\E\\Prob(B_j\\mid\\mathcal F_{j-1})\\le b_j$. A countable union bound gives $\\Prob(\\bigcup_jB_j)\\le\\sum_jb_j\\le\\alpha$. On the complementary event every committed semantic and progress statement is true.\n\nFor any accepted prefix, multiply its stochastic recoveries and apply Theorem~\\ref{thm:debt}. This yields (i) and (ii), simultaneously because the event already covers all attempts. Summing the true progress inequalities telescopes to (iii). Theorem~\\ref{thm:realization} proves (iv). No step of the proof restricts how the proposer chose a frozen candidate from the preceding history, so adaptive proposer changes are permitted.\n\\end{proof}\n\n\\section{What the theorem contributes}\nThe statistical part is related to existing risk-controlled RSI gates, including SGM and TRUST \\cite{sgm,trust}. The present distinction is the separate all-decision recovery obligation and its coherent structural realization. A scalar performance gate can accept an update that forgets another capability; the recovery certificate can detect that problem within the contract. Conversely, a perfectly preserving update may fail the progress gate because it adds no useful information or resource advantage.\n\nThe theorem is not an automatic generator of good candidates. Rejecting every candidate satisfies retention trivially. Chapter~\\ref{ch:progress} makes explicit the search and evidence assumptions needed for accepted progress, so that preservation is not advertised as a solution to the entire RSI problem.\n", "sha256": "cbb529d5a05d1f64cbff1c091eb6b6c0d85521cf1518a2717da220d1969957b1"} {"id": "source-5e1bb7db64b751ef", "source_path": "research/manuscript/13_progress.tex", "format": "tex", "record_type": "research_text", "text": "\\chapter{When certification actually produces progress}\\label{ch:progress}\n\\section{Preservation is not a search algorithm}\nA checking theorem can be perfectly valid while accepting nothing. Consequently, a claim about recursive improvement needs two logically separate ingredients: a reliable acceptance mechanism and a source of sufficiently good proposals. The first is supplied by the lineage theorem. The second is an assumption about a search procedure, a particular model class, or a constructive optimization argument. It cannot be obtained merely by increasing the sophistication of the certificate language.\n\nThis distinction is important even in the finite examples. The ReLU refinement in Chapter~\\ref{ch:experiments} is constructed using knowledge of XOR. Its exact certificate proves what the replacement does, not that a general-purpose agent would discover that replacement from arbitrary data. A learned proposer could search the same space, but its success rate would need separate analysis or measurement.\n\n\\begin{proposition}[Bounded objectives forbid perpetual fixed-size improvement]\\label{prop:bounded-progress}\nLet $U$ take values in $[0,1]$. If every accepted update has $U(P_t)-U(P_{t-1})>\\gamma>0$, then every nonempty accepted prefix satisfies\n\\[\nT\\gamma<1-U(P_0).\n\\]\nIn particular there cannot be infinitely many accepted updates with the same positive lower gain bound. If $U(P_0)=1$, no strict improvement in that objective is possible.\n\\end{proposition}\n\\begin{proof}\nTelescope the accepted gains and use $U(P_T)\\le1$. At the optimum, $U(P)-U(P_0)\\le0$ for every candidate.\n\\end{proof}\nAn endless sequence can instead have positive summable gains, change its objective, or move through expanding contracts. These are different mathematical statements. Changing the contract also changes the meaning of its all-task guarantee and requires a new bridge to the predecessor contract.\n\n\\section{Power of the fixed-sample gate}\nRetain the notation of Chapter~\\ref{ch:lineage}. Suppose the true improvement of a frozen candidate satisfies $\\Delta\\ge2\\gamma$ with $\\gamma>0$. Choose a confirmation size satisfying\n\\begin{equation}\\label{eq:power-size}\nn\\ge \\frac{8}{\\gamma^2}\\max\\{\\log(1/\\alpha_j),\\log(1/\\beta)\\}.\n\\end{equation}\nThen the gate radius $r=\\sqrt{2\\log(1/\\alpha_j)/n}$ is at most $\\gamma/2$. Failure to pass the gate implies $\\widehat\\Delta\\le\\gamma+r\\le3\\gamma/2$, and hence $\\widehat\\Delta-\\Delta\\le-\\gamma/2$. Hoeffding gives\n\\[\n\\Prob(\\text{gate rejects}\\mid\\Delta\\ge2\\gamma)\\le \\exp(-n\\gamma^2/8)\\le\\beta.\n\\]\nThe inequality is conservative but has the virtue of following from the same bounded-difference assumption as the validity guarantee. No asymptotic normal approximation is required.\n\n\\begin{theorem}[Conditional discovery-to-acceptance conversion]\\label{thm:conditional-progress}\nFix an incumbent until a candidate passes all gates. At each attempt $j$, conditional on the entire previous history without a genuine success, suppose the proposer produces a candidate with a valid exact structural and semantic certificate and true gain at least $2\\gamma$ with probability at least $p_j$. Suppose its fresh confirmation size satisfies~\\eqref{eq:power-size} with rejection bound $\\beta_j$. A genuine success means that such a candidate is accepted. If the search is allowed to continue, its probability of no genuine success in the first $N$ attempts is at most\n\\[\n\\prod_{j=1}^N(1-p_j(1-\\beta_j))\n\\le \\exp\\left(-\\sum_{j=1}^Np_j(1-\\beta_j)\\right).\n\\]\nIf a real protocol halts on its first acceptance and its total probability of a false acceptance is at most $\\alpha$, then the probability that it has not halted with a genuine success by attempt $N$ is at most this displayed bound plus $\\alpha$.\n\\end{theorem}\n\\begin{proof}\nCondition first on the proposed candidate and its training history. For every candidate in the assumed good set, the fresh gate accepts with probability at least $1-\\beta_j$. Averaging over proposals gives conditional success probability at least $q_j=p_j(1-\\beta_j)$. If $A_j$ is the event of no genuine success through attempt $j$, then $\\Prob(A_j)\\le(1-q_j)\\Prob(A_{j-1})$ by conditional expectation. Iteration proves the product bound, and $1-x\\le e^{-x}$ proves the exponential bound.\n\nFor a first-acceptance protocol, couple it to the continuing search up to its first acceptance. Unless a false acceptance occurs, a failure to halt with a genuine success by $N$ implies no genuine success in the coupled continuing search through $N$. A union bound adds at most $\\alpha$.\n\\end{proof}\nWhen $p_j\\ge p>0$ and $\\beta_j\\le\\beta<1$, the continuing search's waiting time has expectation at most $1/[p(1-\\beta)]$. This counts attempts, not computational work. The required sample sizes may grow with $j$ as the allocated false-acceptance probability decreases.\n\n\\section{Multiple stages without an independence fiction}\nAfter each genuine acceptance the incumbent changes, so there is no reason for proposal events at different stages to be independent. The conditional formulation avoids that requirement. If the same lower bound on good-proposal probability holds at every stage below a target value, the waiting-time argument can be restarted conditionally at every accepted snapshot. Expected attempt counts then add by the tower property.\n\nFor example, suppose every snapshot with $U