OctoInformBeta

OpenAI preprint, September 2026 · §6. An observation with randomly displaced windows

§6: The unconditioned surrogate and metadata: Independent hidden orientations, typical metadata, and a shared chronological decoder

Sampling displaced windows from their side reports.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

A displaced-window observation retains side reports together with observations at the leaves of a deterministic window tree. The side reports specify which orientation bits are revealed. A common decoder turns the reported emissions into a binary word whose length is the surviving length of the complete trace.

Write [N]={0,…,N−1}[N]=\{0,\ldots,N-1\}, with [0]=⌀[0]=\varnothing; all slot and child indices are zero-based. A finite probability law pp on a finite set AA is a nonnegative mass function with ∑x∈Ap(x)=1\sum_{x\in A}p(x)=1.✦ For finite laws p,rp,r, a finite family of coordinate laws pip_i, conditional laws KxK_x and a real function hh on the space of pp, the product law, coordinate product law, mixture law and joint law, together with expectation, have the formulas in (1). (p⊗r)(x,y)=p(x)r(y),(⨂ipi)(x)=∏ipi(xi),(pK)(y)=∑xp(x)Kx(y),(p⋉K)(x,y)=p(x)Kx(y),𝔼p[h]=∑xp(x)h(x).\begin{aligned} (p\otimes r)(x,y)&=p(x)r(y),& \left(\bigotimes_i p_i\right)(x)&=\prod_i p_i(x_i),\\ (pK)(y)&=\sum_xp(x)K_x(y),& (p\ltimes K)(x,y)&=p(x)K_x(y),\\ \mathbb E_p[h]&=\sum_xp(x)h(x). \end{aligned}(1)

Paired mechanisms MM on a finite side-outcome space SS, with mm trial slots, are families of first-side laws FiF_i, second-side laws GiG_i and common laws HiH_i, with weights wiw_i satisfying (2). wi≥0,wiHi(x)≤Fi(x),wiHi(x)≤Gi(x)(i∈[m],x∈S).w_i\geq0,\qquad w_iH_i(x)\leq F_i(x),\qquad w_iH_i(x)\leq G_i(x)\quad(i\in[m],\ x\in S).(2)✦

A paired report is an array z=(di,xi,yi)i∈[m]z=(d_i,x_i,y_i)_{i\in[m]} in ℛm(S)=({*,0,1}×S2)[m]\mathcal R_m(S)=(\{*,0,1\}\times S^2)^{[m]}, where ** is an absent orientation flag and (xi,yi)(x_i,y_i) is an ordered side pair.✦ For an orientation c∈{0,1}c\in\{0,1\}, the oriented pair law BicB_i^c and hidden pair mass hih_i are the functions in (3); the flagged pair mass ϕic\phi_i^c assigns the hidden mass to an absent flag and the residual oriented mass to a flag revealing cc.✦ Bi0(x,y)=Fi(x)Gi(y),Bi1(x,y)=Gi(x)Fi(y),hi(x,y)=wi2Hi(x)Hi(y),ϕic(d,x,y)={hi(x,y),d=*,Bic(x,y)−hi(x,y),d=c,0,d∈{0,1},d≠c.\begin{aligned} B_i^0(x,y)&=F_i(x)G_i(y),& B_i^1(x,y)&=G_i(x)F_i(y),\\ h_i(x,y)&=w_i^2H_i(x)H_i(y),\\ \phi_i^c(d,x,y)&= \begin{cases} h_i(x,y),&d=*,\\ B_i^c(x,y)-h_i(x,y),&d=c,\\ 0,&d\in\{0,1\},\ d\ne c. \end{cases} \end{aligned}(3)

The paired report law ρi\rho_i averages flagged masses over a fair orientation.✦ The joint report law RMR_M is their coordinate product, as in (4). ρi(d,x,y)=12∑c=01ϕic(d,x,y),RM(z)=∏i∈[m]ρi(di,xi,yi).\rho_i(d,x,y)=\frac12\sum_{c=0}^1\phi_i^c(d,x,y),\qquad R_M(z)=\prod_{i\in[m]}\rho_i(d_i,x_i,y_i).(4)✦

The reported orientation posterior πd\pi_d is the binary law in (5). πd(c)={12,d=*,𝟏{c=d},d∈{0,1}.\pi_d(c)= \begin{cases} \frac12,&d=*,\\ \mathbf1_{\{c=d\}},&d\in\{0,1\}. \end{cases}(5)✦

For leaf laws P,QP,Q on a finite space EE, the selected leaf law LηL_\eta for an orientation vector η∈{0,1}[N]\eta\in\{0,1\}^{[N]} is the coordinate product in (6).✦ An orientation rule θ:{0,1}[m]→{0,1}[N]\theta:\{0,1\}^{[m]}\to\{0,1\}^{[N]} specifies the leaf orientations from the trial orientations. Lη(ℓ)=∏u∈[N]{P(ℓu),ηu=0,Q(ℓu),ηu=1.L_\eta(\ell)=\prod_{u\in[N]} \begin{cases} P(\ell_u),&\eta_u=0,\\ Q(\ell_u),&\eta_u=1. \end{cases}(6)

The Window Tree

Let Tj,nj,gjT_j,n_j,g_j be the window, local trial and global trial counts at level jj, with levels numbered from the bottom. The window trial set 𝒥b\mathcal J_b, leaf set ℒb\mathcal L_b and filler set ℱb\mathcal F_b identify the emissions of the complete depth-bb tree. A window assignment in 𝒳b(A)\mathcal X_b(A) assigns a value in AA to every trial, including trials below unoccupied children.✦ Their recursive definitions are (7). 𝒥0=ℱ0=⌀,ℒ0={•},𝒳0(A)={•},𝒥j+1=[nj]⊔[gj]⊔([Tj]×𝒥j),ℒj+1=[Tj]×ℒj,ℱj+1=([Tj]×{0,1})⊔([Tj]×ℱj),𝒳j+1(A)=A[nj]×A[gj]×𝒳j(A)[Tj].\begin{aligned} \mathcal J_0&=\mathcal F_0=\varnothing,&\mathcal L_0&=\{\bullet\},& \mathcal X_0(A)&=\{\bullet\},\\ \mathcal J_{j+1}&=[n_j]\sqcup[g_j]\sqcup([T_j]\times\mathcal J_j),\\ \mathcal L_{j+1}&=[T_j]\times\mathcal L_j,\\ \mathcal F_{j+1}&=([T_j]\times\{0,1\})\sqcup([T_j]\times\mathcal F_j),\\ \mathcal X_{j+1}(A)&=A^{[n_j]}\times A^{[g_j]} \times\mathcal X_j(A)^{[T_j]}. \end{aligned}(7)

A global schedule with UU children and repetition count tt labels its trials by (v,a)∈[⌊U/4⌋]×[t](v,a)\in[\lfloor U/4\rfloor]\times[t]. The global trial index is vt+avt+a✦; its pre position is vv✦ and its post position is U−1−vU-1-v.✦ For a binary vector ζ\zeta indexed by the global trials and a child index u∈[U]u\in[U], the net global displacement DU,t(ζ,u)D_{U,t}(\zeta,u) is the integer in (8), where bits are identified with their integer values 00 and 11. DU,t(ζ,u)=∑0≤v<⌊U/4⌋v≤u,u+v<U(∑a=0t−1ζvt+a−⌊t2⌋).D_{U,t}(\zeta,u)= \sum_{\substack{0\leq v<\lfloor U/4\rfloor\\v\leq u,\ u+v<U}} \left(\sum_{a=0}^{t-1}\zeta_{vt+a}-\left\lfloor\frac t2\right\rfloor\right).(8)✦

The source schedule has the global counts and displacement maps in (9); at level j+1j+1 its pre and post positions use U=Tj+1U=T_{j+1} and t=njt=n_j. There are no global trials at level zero.✦ g0=0,gj+1=⌊Tj+14⌋nj,Γ0(ζ,u)=0,Γj+1(ζ,u)=DTj+1,nj(ζ,u).\begin{aligned} g_0&=0,&g_{j+1}&=\left\lfloor\frac{T_{j+1}}4\right\rfloor n_j,\\ \Gamma_0(\zeta,u)&=0,& \Gamma_{j+1}(\zeta,u)&=D_{T_{j+1},n_j}(\zeta,u). \end{aligned}(9)

For this schedule, the source trial, filler and leaf counts are mb=|𝒥b|m_b=|\mathcal J_b|✦, kb=|ℱb|k_b=|\mathcal F_b|✦ and λb=|ℒb|\lambda_b=|\mathcal L_b|✦. The window assignment bijection identifies 𝒳b(A)\mathcal X_b(A) with A𝒥bA^{\mathcal J_b} by reading the coordinate at each trial.✦ Fixed finite enumerations of the index sets give the flattened source assignment bijection 𝒳b(A)≃A[mb]\mathcal X_b(A)\simeq A^{[m_b]}.✦ For a paired report z∈ℛmb(S)z\in\mathcal R_{m_b}(S), its nested source report flags d(z)d(z) are the window assignment in 𝒳b({*,0,1})\mathcal X_b(\{*,0,1\}) obtained by regrouping its flag array (di)i∈[mb](d_i)_{i\in[m_b]} through the flattened source assignment bijection.✦

Central Support and Chronology

For optional flags d∈𝒳b({*,0,1})d\in\mathcal X_b(\{*,0,1\}), the window orientation posterior Wb(d)W_b(d) is the coordinate product of the laws πda\pi_{d_a} over a∈𝒥ba\in\mathcal J_b, viewed as a law on 𝒳b({0,1})\mathcal X_b(\{0,1\}).✦ The hidden collection condition ℋb(d;ω)\mathcal H_b(d;\omega), for real ω\omega, is the recursive count requirement in (10); dloc,dglob,dud^{\mathrm{loc}},d^{\mathrm{glob}},d_u denote its root local flags, root global flags and child assignments. ℋ0(d;ω)=true,ℋj+1(d;ω)⇔∑i∈[nj]𝟏{diloc=*}≥njω22,∑i∈[gj]𝟏{diglob=*}≥gjω22,ℋj(du;ω)for every u∈[Tj].\begin{aligned} \mathcal H_0(d;\omega)&=\mathrm{true},\\ \mathcal H_{j+1}(d;\omega)\quad\Longleftrightarrow\quad& \sum_{i\in[n_j]}\mathbf1_{\{d_i^{\mathrm{loc}}=*\}} \geq\frac{n_j\omega^2}{2},\qquad \sum_{i\in[g_j]}\mathbf1_{\{d_i^{\mathrm{glob}}=*\}} \geq\frac{g_j\omega^2}{2},\\ &\mathcal H_j(d_u;\omega)\ \text{for every }u\in[T_j]. \end{aligned}(10)✦

Given integer centers cjc_j, allowed child sets Aj⊆[Tj]A_j\subseteq[T_j], global displacement maps GjG_j, an orientation assignment ξ∈𝒳b({0,1})\xi\in\mathcal X_b(\{0,1\}) and an incoming displacement x∈ℤx\in\mathbb Z, central motif support 𝒞bc,G,A(ξ,x)\mathcal C_b^{c,G,A}(\xi,x) is the condition in (11).✦ Here IsI_s is the motif indicator with size parameter ss, and ξloc,ξglob,ξu\xi^{\mathrm{loc}},\xi^{\mathrm{glob}},\xi_u denote the root local orientations, root global orientations and child assignments of ξ\xi. 𝒞0c,G,A(ξ,x)=true,𝒞j+1c,G,A(ξ,x)⇔[Is(u−cj+∑i∈[nj]ξiloc+x)≠0⇒u∈Ajfor every u∈[Tj]]∧[𝒞jc,G,A(ξu,Gj(ξglob,u))for every u∈[Tj]].\begin{aligned} \mathcal C_0^{c,G,A}(\xi,x)&=\mathrm{true},\\ \mathcal C_{j+1}^{c,G,A}(\xi,x)\quad\Longleftrightarrow\quad& \left[\ I_s\!\left(u-c_j+\sum_{i\in[n_j]}\xi_i^{\mathrm{loc}}+x\right)\ne0 \ \Longrightarrow\ u\in A_j \quad\text{for every }u\in[T_j]\ \right]\\ &{}\land\left[\ \mathcal C_j^{c,G,A} (\xi_u,G_j(\xi^{\mathrm{glob}},u)) \quad\text{for every }u\in[T_j]\ \right]. \end{aligned}(11)

For central-support parameters c,G,Ac,G,A, a flag assignment d∈𝒳b({*,0,1})d\in\mathcal X_b(\{*,0,1\}) and an incoming displacement x∈ℤx\in\mathbb Z, the noncentral posterior mass Nbc,G,A(d,x)N_b^{c,G,A}(d,x) is the expectation in (12). For a source flag assignment d∈𝒳b({*,0,1})d\in\mathcal X_b(\{*,0,1\}) and an incoming displacement x∈ℤx\in\mathbb Z, the source noncentral posterior mass νb(d,x)\nu_b(d,x) is the noncentral posterior mass with the local source centers and central thirds specified in (12) and the global displacement maps G=ΓG=\Gamma from (9).✦ Nbc,G,A(d,x)=𝔼ξ∼Wb(d)[𝟏{¬𝒞bc,G,A(ξ,x)}],cj*=⌊Tj2⌋−1+⌊nj2⌋,Aj*={u∈[Tj]:Tj≤3u≤2Tj},νb(d,x)=Nbc*,Γ,A*(d,x).\begin{aligned} N_b^{c,G,A}(d,x)&= \mathbb E_{\xi\sim W_b(d)} [\mathbf1_{\{\neg\mathcal C_b^{c,G,A}(\xi,x)\}}],\\ c_j^*&=\left\lfloor\frac{T_j}{2}\right\rfloor-1+ \left\lfloor\frac{n_j}{2}\right\rfloor, &A_j^*&=\{u\in[T_j]:T_j\leq3u\leq2T_j\},\\ \nu_b(d,x)&=N_b^{c^*,\Gamma,A^*}(d,x). \end{aligned}(12)

A tree emission identifier is an element of ℰb=(𝒥b×{0,1})⊔(ℱb⊔ℒb)\mathcal E_b=(\mathcal J_b\times\{0,1\})\sqcup (\mathcal F_b\sqcup\mathcal L_b), with the bit selecting the first or second member of a trial pair. Given pre and post positions for global trials, the public emission chronology at depth zero consists of the sole leaf.✦ At positive depth it starts with the first emissions of the root local pairs, visits children in increasing order, and ends with the second emissions of the root local pairs.✦ At child uu the order is: global first emissions assigned to uu, its first filler, the child’s chronology, global second emissions assigned to uu, and its second filler.✦ The chronological emission bijection identifies each zero-based position in the public emission chronology with the tree emission identifier at that position.

The source emission enumeration is a bijection from the tree emission identifiers in ℰb\mathcal E_b to the emission slots 𝒮b=([mb]×{0,1})⊔([kb]⊔[λb])\mathcal S_b=([m_b]\times\{0,1\})\sqcup([k_b]\sqcup[\lambda_b]).✦ For the source pre and post schedule, let κb(j)\kappa_b(j) be the slot at position jj of the chronology, whose length is LbL_b.✦ For an emission slot e∈𝒮be\in\mathcal S_b, the slot outcome space is SS when ee is a paired or filler slot and EE when ee is a leaf slot.✦ The emission assignment bijection regroups report side pairs, filler outcomes y∈S[kb]y\in S^{[k_b]} and leaf outcomes ℓ∈E[λb]\ell\in E^{[\lambda_b]} into slot outcomes Xz,ℓ,yX^{z,\ell,y}: the paired slots read xi,yix_i,y_i from zz, filler slots read yy, and leaf slots read ℓ\ell.✦

For a side decoder f:S→{0,1}*f:S\to\{0,1\}^*, a leaf decoder g:E→{0,1}*g:E\to\{0,1\}^* and an emission slot e∈𝒮be\in\mathcal S_b, the slot decoder 𝖽e\mathsf d_e is the function that applies ff at paired and filler slots and gg at leaf slots.✦ For an indexed family of binary words wjw_j, write ⨁jwj\bigoplus_j w_j for their concatenation in increasing order of jj.✦

The Definition

Definition 1. Let MM be paired mechanisms satisfying (2), let P,QP,Q be finite leaf laws, and let θ\theta be an orientation rule. For a paired report z∈ℛm(S)z\in\mathcal R_m(S), the joint orientation posterior Πz\Pi_z is the law on {0,1}[m]\{0,1\}^{[m]} given by (13). Πz(b)=∏i∈[m]πdi(bi),b∈{0,1}[m].\Pi_z(b)=\prod_{i\in[m]}\pi_{d_i}(b_i), \qquad b\in\{0,1\}^{[m]}.(13)✦ For paired mechanisms MM satisfying (2), finite leaf laws P,QP,Q and an orientation rule θ\theta, the canonical source oracle 𝒪M,P,Q,θ\mathcal O_{M,P,Q,\theta} is the joint law on reports and leaf observations in (14). 𝒪M,P,Q,θ(z,ℓ)=RM(z)∑b∈{0,1}[m]Πz(b)Lθ(b)(ℓ).\mathcal O_{M,P,Q,\theta}(z,\ell) =R_M(z)\sum_{b\in\{0,1\}^{[m]}}\Pi_z(b)L_{\theta(b)}(\ell).(14)✦ For source window counts T,nT,n, motif size ss, depth b+1b+1 and real parameters ω,a\omega,a, typical source metadata are the reports z∈ℛmb+1(S)z\in\mathcal R_{m_{b+1}}(S) satisfying (15). 𝖳𝗒𝗉𝗂𝖼𝖺𝗅b+1(z;ω,a)⇔ℋb+1(d(z);ω)∧νb+1(d(z),0)<a.\mathsf{Typical}_{b+1}(z;\omega,a)\quad\Longleftrightarrow\quad \mathcal H_{b+1}(d(z);\omega) \ \land\ \nu_{b+1}(d(z),0)<a.(15)✦ For side and leaf decoders f,gf,g, the public source oracle decoder 𝒟b\mathcal D_b at depth bb is the function from reports, leaf observations and filler outcomes to binary words specified by (16). 𝒟b(z,ℓ,y)=⨁j=0Lb−1𝖽κb(j)(Xκb(j)z,ℓ,y).\mathcal D_b(z,\ell,y)= \bigoplus_{j=0}^{L_b-1} \mathsf d_{\kappa_b(j)} \bigl(X^{z,\ell,y}_{\kappa_b(j)}\bigr).(16)✦

In the paper, the result reads: the deterministic window tree reports old leaf observations, side traces, flags and revealed orientation bits with fixed relative identifiers; the common projection applies the old decoders and concatenates in chronological order; given all side metadata, unrevealed bits remain independent and fair, while sibling offsets retain their shared global component; fit requires every interior displacement at level jj to have absolute value at most Tj/8T_j/8; typical metadata also require at least half the expected hidden counts in every local and global collection and a suitably small conditional fit-failure probability, and the hidden bits are never conditioned on fit.

What Is Checked

These definitions are stated over finite side and leaf laws, paired mechanisms with pointwise common-submass domination, complete window trees, and arbitrary orientation rules and word decoders. The paired report masses in (4) sum to one for any real weights✦; under (2) they are nonnegative.✦ The global indexing map (v,a)↦vt+a(v,a)\mapsto vt+a is bijective.✦