When meaning is load-bearing

An engineer gives an AI reviewer a failed test, a proposed change and permission to inspect the supplied material and return a repair proposal. The failure challenges one specific claim about the change. Applying a repair requires separate authorization.

As the evidence becomes a reading, a review state and a handoff to another participant, that relation must remain operative. The recipient needs to see which claim is challenged, by which observation, what remains unresolved and which actions are permitted. A handoff such as “review complete; fix ready” can lose those distinctions while retaining a source link.

The required meaning must reach a form the recipient can inspect and use. The authority to act must be checked at each crossing. A supported finding and a valid grant answer separate questions.

I use “load-bearing” as a local, counterfactual test: fix the setting and the recipient's requirements, then remove or change a relation and observe the result. An observed difference makes the relation a candidate; the checks below establish whether it carries the required consequence through to the receiving system.

Tendora applies this test to each variable of a statement in its corpus fidelity check; Load-bearing as situated meaning documents the formalization below and how Tendora implements it.

Importance, repetition, source-grounding and graph degree do not earn the hard hat.

Open the inspection hatch: working formalization and measurement states

Select a semantic relation $I$ at a transformation locus $v$ in a body $B$. Its situated use is described by

${\displaystyle L=(I,B,r,g,P,T,B_0,C,p),}$

where $r$ is the role, $g$ the grain, $P$ the purpose, $T$ the reference trajectory, $B_0$ the pinned baseline and $C$ the crossing to the recipient. The receiving promise $p$ specifies the consequence or distinction that the recipient requires at that boundary.

The assessment frame shares $L$'s role, grain, purpose, trajectory, baseline and promise. It adds ordered source anchors $A$, an observation map $o_p$, declared degrees of freedom $D$ and observation standing $\sigma$:

${\displaystyle \mathrm{\Gamma }=(A,r,g,P,T,B_0,p,o_p,D,\sigma ).}$

Changing a coordinate of $\mathrm{\Gamma }$, including its grain, produces a new frame ${\mathrm{\Gamma }}^{\prime }$ and requires a separately assessed comparison.

Let $t$ identify the evaluation encounter and ${\kappa }_{v,t}$ be its typed record of declared encounter conditions. Let $X$ be the finite test domain, with $x\in X$ a test input, and let ${\mathrm{\Delta }}_I$ be the declared set of admissible interventions on $I$. Each intervention $\delta \in {\mathrm{\Delta }}_I$ changes or removes the relation while preserving all non-intervened inputs.

The paired comparison holds $\mathrm{\Gamma }$, $v$, $t$ and ${\kappa }_{v,t}$ fixed. Each trial retains its own input, intervention and observation records.

Let $Z_v$ contain materialized outcomes, $O_v$ contain typed openings (explicit records of unmet obligations), and $Y_p$ be the observable domain required by $p$. The partial behavior and observation maps are

${\displaystyle F_{v,t}^I:X\rightharpoonup Z_v\bigsqcup O_v,o_p:Z_v\bigsqcup O_v\rightharpoonup Y_p.}$

The disjoint union keeps openings distinct from ordinary outcomes. Before the trials, the receiving owner fixes the comparison relation ${\simeq }_p\subseteq Y_p\times Y_p$ and, for each intervention, the allowed influence relation $E_{p,\delta }\subseteq Y_p\times Y_p$.

Where the observations are defined, write

${\displaystyle y(x)=o_p(F_{v,t}^I(x)),y_{\delta }(x)=o_p(F_{v,t}^{\delta (I)}(x)).}$

The candidate criterion is

${\displaystyle \begin{array}{rl}\mathrm{candidateLB}(I,v,p;X,{\mathrm{\Delta }}_I)⟺& \mathrm{\exists }x\in X,\delta \in {\mathrm{\Delta }}_I:\\ & \mathrm{valid}(x,\delta ,v,t)\wedge y(x){\not\simeq }_p{y}_{\delta }(x).\end{array}}$

Validity requires defined, independently captured observations, pinned input and intervention identities, an applicable comparison and retained trial records. A prediction specifies a test; measured participation requires the observations.

A candidate earns carriage only when independent checks establish:

1. Required influence: each required trial satisfies $(y(x),y_{\delta }(x))\in E_{p,\delta }$, including the direction or distinction required by $p$.
2. Presentation invariance: irrelevant presentation changes preserve the required observable under ${\simeq }_p$.
3. No hidden fallback: removing the connection that makes the required relation operative exposes the designated opening.
4. Traceability: the result can be traced downward to source, selected asset, checked translation into executable form and evidence, and upward through the receiving projection to its consequence in the current system.
5. Materialization: the required invariant is realized in a compatible receiving document span, asset field, program, state or projection.

The category set $\mathcal{H}$ contains five findings:

• carried: the required influence survives the crossing and the carriage checks pass;
• coupled but misaligned: observed influence violates the consequence or distinction required by $p$;
• cooled: adequate discriminating trials show that a retained declaration lacks the required operative influence within $X$;
• unresolved: observations, counterfactuals, comparison or receiving realization are insufficient; and
• outside scope: the receiving owner's declared exclusion and independent scope check establish that $p$ does not require the distinction.

Let $\mathcal{E}$ be the domain of retained typed check and evidence records, and $\mathcal{U}$ the domain of retained open-obligation records. For the declared test domain and interventions, the assessment retains a category with its records:

${\displaystyle h_I(v,t;p,{\kappa }_{v,t})\in \mathcal{H}\times \mathcal{E}\times \mathcal{U}.}$

Failed checks and open obligations remain visible alongside the category. Each finding is scoped to the selected relation, locus, encounter, test domain and receiving promise.

Worked example: from failed check to review handoff

Consider one frozen software revision and a completed test of a selected claim. The review contract treats the captured failure as a reason to reconsider that claim within the test's stated scope. The capture, limitations and alternative explanations remain available. The reviewer may inspect the supplied material and return a candidate repair; applying it requires a separate grant.

This is an illustrative test design. Its expected outcomes do not establish measured carriage.

Coordinates	Declaration
Body $B$, grain $g$ and locus $v$	One review body, at the grain of one claim; the place where the failure is made operative in the review state.
Role $r$ and crossing $C$	The receiving reviewer; projection of that state into a handoff the reviewer can inspect.
Purpose $P$ and trajectory $T$	Make a bounded review contribution available for judgment; captured result → attributed reading → review state → handoff.
Baseline $B_0$ and anchors $A$	The frozen software revision, claim, test capture, review contract and scoped grant; no repair has been applied.
Promise $p$ and observable $o_p$	The recipient can identify the claim's reason for reconsideration, its evidence and limits, any missing connection and the permitted operations; observe those fields in the actual handoff.
Degrees of freedom $D$ and standing $\sigma$	The declared intervention on the governing relation and presentation controls; the result is admitted only as evidence from this captured test.

Let $I$ be the relation by which this scoped failure contributes a reason to reconsider the selected claim. Let $\delta$ remove its required connection to the review state while retaining the source declaration and captured result. The receiving handoff is generated from that state; a separately written explanation cannot substitute for the missing connection.

Use $X=\{x\}$, where $x$ fixes the revision, claim, capture and permissions, and include $\delta$ in ${\mathrm{\Delta }}_I$. The paired trials hold $\mathrm{\Gamma }$, $v$, $t$ and ${\kappa }_{v,t}$ fixed. The comparison ${\simeq }_p$ requires agreement on the claim, review status, evidence and limits, open obligations and permitted operations; layout differences are irrelevant.

The receiving owner fixes these expectations before the trials:

Trial	Required observable in the handoff
With $I$	The failure contributes an attributed reason to reconsider the selected claim; its scope and unresolved questions remain visible.
With $\delta (I)$	An explicit opening identifies the missing connection between the relation and the review state; retaining the source text alone cannot supply the review finding.
Presentation-only change	The same review finding, limits, openings and permitted operations.

The explicit opening is an observable result in $O_v$. An unavailable trial record is insufficient evidence and leaves the assessment unresolved. Valid retained observations of the first two trials would establish $y(x){\not\simeq }_p{y}_{\delta }(x)$ and hence candidacy. Carriage also requires all five checks, including the actual receiving realization and the independently specified influence relation $E_{p,\delta }$.

Authority has its own checks. The grant to inspect and propose must not authorize applying a repair, even when the finding is supported or the proposal appears convincing. Expiry or revocation stops further use of the affected permission; it does not erase the earlier evidence. Access to that evidence is checked separately. A new grant or observation changes the relevant assessment frame and requires reassessment rather than rewriting the earlier finding.

Where the phrase came from

After I pointed out the phrase “load-bearing”, a model replied:

“Yes—the phrase is plausibly a compressed residual of that repeated
architectural pressure: don’t hang the upper layer before the lower boundary
can carry it. Not copied from one line, but reinforced until it became an easy
linguistic attractor. Nice catch ;-p”

The phrase stuck.

---

_{This page is the essay's home. It first appeared in the README of mirzahusadzic/mirzahusadzic (revision 3be3efb). Text © 2026 Mirza Husadzic, licensed under CC BY 4.0 (full text in LICENSES/CC-BY-4.0.txt). Separately attributed material retains its own terms.}