From Observable Answers to Executable Queries for CM/LM Programs
Brian Theory | Technical report v1.1 | 8 November 2026
Text edition; equations retain TeX notation.

An answer can be determined by an abstract observation without being obtainable
from an available sensor library or executable measurement policy. This note
separates these three requirements for finite Boolean programs. It connects
static acquisition to weighted conflict covering, gives exact adaptive cost
recurrences, and proves that paid addressing requires at least $b+1$ queries
on every input, attaining the same optimum under every prior. A CM/LM example
shows why cancelled internal frames need not be observed; an outcome gate
shows why knowing a prerequisite differs from executing it. A small independent
verifier checks policy legality, outputs and costs. The contribution is an
expository synthesis with executable examples of established methods.



Three questions about an observable answer

For a program's required output, ask: what information determines it, which
available queries supply that information, and which query sequences are legal?
These questions become distinct when inputs must be acquired at a cost.
Tensor Logic's tensor-equation perspective motivates specifying the observable
interface of a computation [domingos]. Correspondence and logical matrices
(CM/LM) make Boolean frames and pairing explicit [cmlm]. Neither notation
alone supplies a measurement policy.

The underlying mathematics is established: functional dependency concerns
whether an answer is a function of given observations [jiang]; relative
discernibility and test-cost reduction underlie static selection [honko,min];
decision-tree certificates bound what a computation must learn
[grossman]. This report connects these ideas through a common finite model,
worked counterexamples and reproducible code. Its addressing calculation is an
elementary certificate argument, not a claim of historical firstness. No
CM-exclusive algorithmic advantage or integrated Tensor Logic backend is claimed.


Model and costs

Fix a nonempty finite feasible set $\Omega$, required output $F:\Omega\to Y$,
free initial information $\alpha:\Omega\to A$, and binary coordinate or sensor
queries $s_j:\Omega\to\{0,1\}$. Write $m$ for the number of queries. Execution prices $c_j$ and installation prices
$a_j$ are nonnegative rationals; nonnegative prior weights $p(x)$ sum to one. Exactness is
required on every feasible state, including those with zero prior weight.
Queries are deterministic, do not change the state, and occur at most once per
execution path. Policies use only $\alpha$ and acquired readings. A correct
leaf outputs $F(x)$ for every state reaching it. No extra action is required
after the answer is known. Unless specified otherwise, all queries are available.

An abstract representation $\beta$ suffices when $F=h\circ\beta$. In static
acquisition, $\beta=(\alpha,(s_j)_{j\in J})$ for one globally fixed subset $J$
chosen before $\alpha$ is revealed; every selected query is paid. Adaptive
policies start with $\alpha$ available and choose subsequent queries from their
readings. If $C_T(x)$ is path execution cost, the three objectives are
\[
 \max_x C_T(x),\qquad \sum_xp(x)C_T(x),\qquad
 \sum_{j\in U(T)}a_j,
\]
where $U(T)$ is the union of queries used across all feasible paths and initial
observations. Conditional static selection after learning $\alpha$ is a different
comparator. Storage, decoding, calibration and communication are not priced.

The coarsest abstract repair is $(\alpha,F)$: its classes are the nonempty
intersections of the original observation and target classes. Every refinement
determining $F$ must refine these intersections. This does not create a cheap
sensor for $F$. Likewise, replacing $F$ by an equal output leaves acquisition
optima unchanged when states, sensors, costs, prior and legal actions are fixed.


Costed conflict covering

Let
\[
 E=\{\{x,y\}\subseteq\Omega:\alpha(x)=\alpha(y),\ F(x)\ne F(y)\},
 \qquad D_{xy}=\{j:s_j(x)\ne s_j(y)\}.
\]
[Exact static acquisition]
A subset $J$ determines $F$ given $\alpha$ iff $J\cap D_{xy}\ne\varnothing$
for every $\{x,y\}\in E$. Minimum static execution cost is therefore
\[
 \min_{z\in\Bool^m}\sum_jc_jz_j\quad\text{subject to}\quad
 \sum_{j\in D_{xy}}z_j\ge1\quad(\{x,y\}\in E).
\]
Without extra compatibility restrictions, a solution exists iff every $D_{xy}$
is nonempty. Empty $E$ permits $J=\varnothing$.


If a conflict pair is missed, the two states yield identical acquired readings
and different answers, so no decoder works. Conversely, if every conflict is
hit, no acquired-observation fiber contains unequal targets. Assign each
attained reading its fiber's target value; values on unattained readings are
arbitrary. This constructs a decoder. Taking the indicator of $J$ gives the
optimization. Selecting the complete library hits every nonempty $D_{xy}$;
an empty one is hit by no subset.

This is weighted set cover on conflicting pairs, or hitting set on the $D_{xy}$.
Relative discernibility clauses are direct antecedents [honko]; additive
test-cost reduction is also established [min]. Unlike preservation of an
inconsistent table's existing discrimination, total recovery must retain an
empty conflict clause as infeasibility. An acquisition budget adds a price bound;
compatibility restrictions add admissibility constraints. Full-library separation
alone no longer suffices if the whole library cannot be jointly selected.


Exact adaptive acquisition

A history $h\in\{*,0,1\}^m$ records untested sensors and acquired values. Set
\[
 B(h)=\{x\in\Omega: s_j(x)=h_j\text{ whenever }h_j\ne *\},
 \qquad q(B)=\sum_{x\in B}p(x).
\]
For nonconstant $\alpha$, restrict $\Omega$ to the initial observed fiber.
Let $L(h)$ contain the legal untested sensors, and write $h[j=v]$ for extension.
This history representation is sufficient for our outcome gate below. An
order-sensitive protocol requires an ordered trace or sufficient automaton state.

[Exact recurrences]
Let $W(h)$ be minimum worst future execution cost and $E(h)$ minimum
unnormalized expected future cost using the original prior. Both are zero
if $B(h)$ is empty or $F$ is constant on it. Otherwise define
\[
 A(h)=\{j\in L(h): W(h[j=v])<\infty
       \text{ for every }v\text{ with }B(h[j=v])\ne\varnothing\}.
\]
If $A(h)$ is empty, both values are infinity. Otherwise

 W(h)&=\min_{j\in A(h)}\left(c_j+
       \max_{v:B(h[j=v])\ne\varnothing}W(h[j=v])\right),\\
 E(h)&=\min_{j\in A(h)}\left(q(B(h))c_j+
       \sum_{v:B(h[j=v])\ne\varnothing}E(h[j=v])\right).



Induct on the number of untested sensors. A correct leaf has constant target;
unreachable branches impose no obligations. Every nonterminal legal policy
first selects an allowed test and must attach a feasible policy to every
nonempty outcome branch. Thus its first choice belongs to $A(h)$. Execution
pays $c_j$ on every remaining feasible path; worst costs maximize child costs,
whereas expectation contributes $q(B)c_j$ and sums the children's unnormalized
expectations. Induction bounds every child's cost by its optimum. Conversely,
choose a minimizing action and attach separately minimizing child policies for
the objective in question. Their composition attains the displayed expression.
There are finitely many tests and actions. If no feasible first action exists,
no correct nonterminal tree exists. Each extension removes a test, so this
induction also covers zero prices without a cyclic optimality equation.

The two minima may require different trees. Store separate argmins when extracting
policies. At positive mass, $E(h)/q(B(h))$ is the conditional optimum; at zero
mass no division is needed. A nonempty impossible child is disallowed even when
its mass is zero: exact recovery is required on all feasible states. We never
use $0\cdot\infty$. A feasible zero-mass subtree has zero expected contribution.
For free initial observations, aggregate fiber values by maximum for worst cost
and sum for unnormalized expectation. Installed unions couple the fiber policies.

[Installation under unrestricted availability]
When tests are always available, minimum installed price of an exact policy
equals minimum static sufficient-subset price with coefficients $a_j$.


Let $U(T)$ be the sensors occurring anywhere in an exact policy. Two states with
the same $\alpha$ and identical readings on $U(T)$ follow the same tree path,
since each next action depends only on readings already matched. Hence their
targets agree and $U(T)$ is a sufficient subset. Its installed cost is exactly
the price of that subset. Conversely, test any sufficient subset in a fixed
order and decode the attained signature; all tests are available and this policy
uses only that subset. The two inequalities prove equality.



A paid-address separation


For each $b\ge0$, let $k=2^b$ and let states range over all assignments to $b$
address bits $a$ and $k$ data bits $d_0,\ldots,d_{k-1}$. Sensors read individual
bits, all execution and installation prices are one, $\alpha$ is constant, and
$F(a,d)=d_{\operatorname{index}(a)}$. The static and installed optima are $b+k$;
every correct policy uses at least $b+1$ queries on every input. Reading the
address and then its selected data bit uses exactly $b+1$ queries on every input.
Thus both the adaptive worst optimum and the expected optimum under every
fixed prior are exactly $b+1$, including priors with zero-mass feasible states.


Each data coordinate is essential by setting the address to its index and
changing that bit alone. Each address coordinate is essential by choosing two
addresses differing at that coordinate and assigning their selected data bits
different values. Omitting any sensor therefore leaves a conflict pair. All
sensors together suffice, proving the static claim; Proposition~[prop:install]
proves the installed claim.

Fix any terminating feasible transcript ending at a correct leaf with answer
$v$, and consider the coordinate values acquired along that particular path. Every full input agreeing with those values follows the
same transcript: inductively the deterministic policy sees the same earlier
readings and chooses the same next test. Thus the consistent inputs form the
entire subcube specified by the distinct coordinates read on that path.

Suppose $r$ address coordinates remain unread. Exactly $2^r$ addresses are
compatible with the transcript. For each such address $i$, coordinate $d_i$
must have been read and must equal $v$. Otherwise, if it is unread, complete
the address to $i$ and set $d_i=1-v$, retaining every acquired value. If it was
read with a different value, simply complete the address to $i$. In either
case the full feasible cube supplies an input following this leaf but having
target $1-v$, a contradiction. Conversely, if every compatible addressed
coordinate has been read with the same value, the target is constant on that
subcube.

Consequently the path has read at least

\[
 (b-r)+2^r \ge b+1
\]
distinct coordinates, since $2^r\ge r+1$ for every integer $r\ge0$ (by induction).
For $b=0$ this requires the sole data coordinate, so the same proof applies.
Additional irrelevant tests cannot reduce this count. Although our model uses
each test at most once, allowing repeated queries at unit price cannot weaken
the lower bound: total queries are at least distinct queried coordinates.

Reading all address bits and then the selected data bit attains $b+1$ on every
input. Therefore it is simultaneously optimal at every input. For any fixed
prior $p$, summing the pointwise lower bound gives
$\sum_xp(x)C_T(x)\ge(b+1)\sum_xp(x)=b+1$, attained by the same policy.
No full-support assumption on $p$ is needed.

This is the ordinary decision-tree certificate argument: a leaf's observations
certify its answer, and here every input has certificate complexity $b+1$.
The standard pointwise certificate and strict instance-optimality definitions
are given in [grossman]. The derivation above
specializes those established notions; no historical-firstness claim is made.


The full feasible cube is essential. A prior changes weights, not which inputs
require exact answers. A promise changes feasibility: for $b=1$, restricting
to $d_0=d_1$ permits reading only $d_0$, at cost $1$, even with both addresses
possible. Concentrating a prior on those same inputs while keeping the entire
cube feasible still requires cost at least $2$ on every path of an exact policy.
Restricting feasibility to a single known input even permits zero tests.

The static-to-worst and static-to-optimal-expected ratios are both exactly
$(k+b)/(b+1)$ for every prior, and diverge with $b$. The static subset here is
globally fixed and all its readings are paid; it must be sufficient on the full
cube. Supplying the address free changes that comparator: conditional static
selection can then choose its single relevant data coordinate. Composite
sensors, bounded-error answers and nonunit prices are outside this proposition.


A worked CM/LM interface and its Boolean baseline

Formula-valued matrices here multiply over the Boolean ring: addition is XOR,
multiplication is AND. This is not existential OR--AND relational composition
or ordinary real matrix multiplication. Define the true-first state vector
$|u\rangle=(u,\neg u)^T$ and frame
\[
 S_g=g&\neg g\\\neg g&g,\qquad
 C_{\land}=1&0\\0&0.
\]
Disjointness $g\land\neg g=0$ and $g\mathbin\oplus\neg g=1$ give $S_g^2=I$.
The logical-matrix normal form for conjunction is
\[
 M_{p\land q}=S_p C_{\land}S_q
 =
 pq&p\neg q\\\neg p\,q&\neg p\,\neg q
 .
\]
Here $pq$ abbreviates $p\land q$, not a probabilistic independence approximation.
By the one-hot disjoint selector calculation,
\[
 \langle u|M_{p\land q}|v\rangle
   =(u\leftrightarrow p)\land(v\leftrightarrow q).
\]
For example, $u=p,v=q$ selects truth value 1, whereas $u=v=1$ selects $p\land q$.
The upper-left entry and reference-frame pairing are different readouts.

Now take states $(p,g)\in\Bool^2$, sensors $p,g$ with prices 1,100, and export
\[
 F(p,g)=\left(p,\ \langle1|S_g I S_g|1\rangle\right).
\]
Using associativity in this ring, $S_g I S_g=I$ and the pairing equals 1, so
$F(p,g)=(p,1)$. This is the matched-frame cancellation law [cmlm] in a
fully specified interface. The ordinary Boolean formulation exports exactly
$(p,1)$; it requires sensor $p$ at cost 1 and never needs $g$. The two states
with $p=0$ and $p=1$ force acquiring $p$; hence 1 is optimal for both descriptions.
If the intermediate matrix $S_g$ is also exported, its entries determine $g$,
so both sensors are needed at price 101. That is a stronger contract, not a
frame-derived speedup. If a fixed workflow mandates reading $g$ regardless of
exports, eliminating it requires permission to change that workflow.

CM/LM adds an explicit account of polarity, pairing and allowed cancellation.
It does not add information to the sensors, alter the true optimum for unchanged
exports, or certify a learned numerical embedding. The selector's disjoint
weights permit XOR and OR to agree for this pairing; arbitrary multiwitness
aggregation does not share that property. Nor may an internal frame be canceled
through an arbitrary threshold or pointwise nonlinearity without another proof.



Sufficiency can fail to be executable

Consider sensors $g,t$ with execution prices $0,1$, where $t$ is available only
after $g$ has actually been tested and returned $1$. On the full four-state
cube $(g,t)\in\{0,1\}^2$, target $t$ is statically determined by the single
sensor $t$. Nevertheless no exact legal policy exists: the only possible first
test is $g$, and its zero branch leaves both target values possible with no
legal remaining query. Giving that branch zero prior probability does not
remove its correctness obligation.

Now restrict feasibility to $(1,0),(1,1)$. Querying $g$ then $t$ is legal and
costs $1$. Every correct policy must acquire $t$, so this cost is optimal.
Although $g$ conveys no new information, its execution unlocks $t$. The belief
set before and after $g$ is the same; legal actions differ. Thus discarding
all non-splitting tests or recording only the belief set can destroy feasibility.

This outcome gate is stronger than a fixed precedence constraint, which requires
earlier tests irrespective of their answers [szyfelbein]. No approximation
guarantee for that different model is imported. More general chronological
protocols require sufficient ordered history or an explicit protocol state.


Checking a proposed policy

The supplied verifier takes trusted explicit sensor/target tables, prices,
prior and prerequisite rules. Its certificate gives one root per attained
initial observation, test nodes with outcome branches, declared leaf answers,
and exact claims for worst, expected and installed costs. Prerequisites are
conjunctions of previously executed query outcomes. This restricted format
does not implement arbitrary history-dependent legal-action predicates.

For each feasible state, including every zero-prior state, the verifier follows
the proposed policy and checks prerequisites, branches and leaf answers. It
rejects repeated tests or cycles along a feasible execution path. Shared nodes
are checked under each reaching history; global graph acyclicity is not required.
The canonical serialization contains exactly the nodes and edges used by feasible
states. Unreachable branches must be omitted. A digest binds canonical model
JSON, not the authenticity or physical adequacy of those tables.

Acceptance certifies legal finite execution, exact answers and the three claimed
costs within this format. To see this, enumerate all feasible paths: their
individual price sums determine maximum and prior-weighted costs, and their
union of query labels determines installed cost. Every operational node and
edge is covered by the canonical check. Conversely, a legal canonical policy
with truthful answers and costs passes these checks. No comparison with other
policies occurs, so optimality is not certified. A retained fixture deliberately
accepts a correct three-query multiplexer policy although two queries suffice.


Reproducibility and scope

The accompanying Python 3.10+ package uses only the standard library. From
the package root, run:
python -B code/run\_checks.py
This reproduces
static selection, adaptive acquisition, matrix identities, policy verification
and the addressing certificate checks in temporary directories. Machine-readable
expected results are included. The README explains model formats, commands,
provenance and the distinction between validity and optimality.

The retained policy suite has five accepted fixtures and nineteen rejected
corruptions, including forged costs, missing branches and zero-prior failures.
The addressing audit exhausts 177,906 partial assignments for $b=0,1,2,3$ and
also checks sampled adaptive paths with repeated queries. These finite checks
support the proofs; they do not prove the all-$b$ statement. Full counts and
source hashes are supplied with the code, outside the report's narrative.

Explicit conflict construction costs $O(|\Omega|^2m)$ comparisons. Exhaustive
subset optimization is exponential in $m$; observed-outcome dynamic programming
has at most $3^m$ records per initial observation fiber, plus legal-action evaluation, protocol-state and
exact-arithmetic costs. The verifier traverses at most $|\Omega|m$ test steps,
plus schema, prerequisite and arithmetic work. Succinct encodings can have
large intermediates; none of these bounds proves a scalable tensor algorithm.

The examples distinguish semantic sufficiency, affordable acquisition and legal
execution. The CM/LM frame and pairing structure describes
the output contract; identical Boolean outputs with identical access assumptions
have identical acquisition optima. The report's value is this usable synthesis,
not a new learning result or exclusive computational advantage. No noisy sensing,
numerical-embedding guarantee, physical deployment or performance benchmark is
established.


Preparation and review.

AI tools assisted with drafting, code and internal mathematical review. The
proofs and exact tests are available for inspection; internal AI reviews are
distinct from human peer review or proof-assistant verification. Compilation and page-inspection details accompany this edition.


[cmlm]  Brian Theory. Correspondence and Logical Matrices:
A Typed Boolean Operator Calculus. September 27, 2026 revision, public
reading copy 1. Sections 3.2, 4 and 5.1; Theorems 3.1, 4.1 and 5.4.
https://www.b-theory.com/manuscripts-html/correspondence-matrices/.
Accessed 4 November 2026.

[domingos]  P. Domingos. Tensor Logic: The Language of AI.
arXiv:2510.12269v3. https://arxiv.org/html/2510.12269v3.

[jiang]  J.-H. R. Jiang, C.-C. Lee, A. Mishchenko and C.-Y. Huang.
To SAT or Not to SAT: Scalable Exploration of Functional Dependency.
IEEE Transactions on Computers 59(4), 457--467, 2010.
https://people.eecs.berkeley.edu/~alanmi/publications/2010/tc10_sat.pdf.

[honko]  P. Ho\'nko. Attribute reduction: a horizontal data decomposition
approach. Soft Computing 20, 951--966, 2016 (online 2014).
https://doi.org/10.1007/s00500-014-1554-8.

[min]  F. Min, Q. Hu and W. Zhu. Feature selection with test cost
constraint. arXiv:1209.5601, 2012. Section 2.2.
https://arxiv.org/pdf/1209.5601.

[grossman]  T. Grossman, I. Komargodski and M. Naor.
Instance Complexity and Unlabeled Certificates in the Decision Tree Model.
LIPIcs, ITCS 2020, 151:56:1--56:38, 2020.
https://doi.org/10.4230/LIPIcs.ITCS.2020.56.

[szyfelbein]  M. Szyfelbein and D. Dereniowski.
Precedence-Constrained Decision Trees and Coverings.
arXiv:2602.21312v4, 12 July 2026.
https://arxiv.org/html/2602.21312v4.


