Return to the full study

KEDDEH SYSTEMS · K-WP05 · E2

Observer-Anchored Measurement Algebra

A formal separation of source, frame, foreground, environment, and retained residue

Abstract

A measurement becomes difficult to interpret when a reported value is detached from its source, observer, instrument, foreground or conditions. This paper develops an observer-anchored record and forward-model discipline in which those roles remain distinct. Its organising expression preserves the source-defined argument order, while a deliberately conventional linear comparison asks when separate source and foreground contributions can be recovered. A null-space argument proves non-identifiability for one blended view; a second, differently transformed view gives a full-rank system. Exact editorial calculations reproduce the example and show why nearly repeated transforms can still amplify measurement error. The result specifies when the comparison model can separate its unknown contributions. The paper supplies operational definitions, units, evidence boundaries and a test programme for transferring the record contract to instruments, simulations and cosmological inference. The current literal origin designation .1 and the distinction between unobserved and nonexistent remain intact.

1. Research question and interpretation boundary

The question is not simply whether an instrument returns a number. It is whether the number can support the particular inference claimed for it. A detector may accurately report a blended signal even when the contributions to that signal cannot be separated. Calling that total an individually measured source value would make a stronger claim than the observation supports.

The July source therefore distinguishes source identity, observer context, structured foreground, environment and retained residue. Its central thesis motivates source-resolved inference: the source and translation conditions must remain distinguishable in the reasoning. This edition makes an important qualification explicit. An aggregate measurement can remain scientifically useful without uniquely identifying every latent component; what fails in that situation is the unsupported decomposition. [1]

The current authored interpretation places the observed entity, observer and observable in distinct roles, with spatial, temporal and instance context. Equal reported values do not establish equal relational records. An unavailable observation does not establish nonexistence. The origin designation .1 is an exact nonnumeric label for the shared referent. It is neither 0.1 nor a numerical offset, and this paper does not perform arithmetic on it. Historical numerical conventions in the source remain historical. [2, 3]

The July paper's conventional linear comparison provides a concrete way to investigate this question. Its response matrices and unknowns are defined below, and its rank result applies under those conditions. The native ordered relation is retained as the source record; the comparison is its separately identified mathematical application.

2. Source-defined model and operational meanings

The three original equations are retained unchanged:

O_r=T(S,M_o,F,H,R) (1)

y_o=M_o s+F_o f+ε_o (2)

rank([I I])=n<2n; rank([[I,I],[I,2I]])=2n (3)

Equation (1) names a translation rather than defining a universal algorithm. S identifies the source and its domain variables. M_o records observer frame, motion and instrument. F is structured intervening signal. H states the translation medium and environmental conditions. R is retained residue or memory when a defined state law exists. O_r is the decoded observation, accompanied by uncertainty and its source relation. T must be supplied by the application. Its ordered arguments remain S, M_o, F, H, R. [1]

The wider source lineage makes two further dependencies explicit. Draft XXXVII §XXV.8, “Mathematical Working: Observer-State Translation,” adds C for calibration weighting and U for uncertainty propagation to the observer relation. Applied in this edition's comparison, calibration is carried by the declared response coefficients and their version, while uncertainty accompanies the observation and its covariance. An application modelling those as independent inputs should retain the expanded source relation. This explains their placement here without changing the original five-argument equation. [7]

The source uses M_o both for the broad observer record in (1) and for a matrix in (2). This edition retains that notation but distinguishes the meanings by equation. In (2), M_o is the declared linear source-to-observation map; it cannot stand for an arbitrary context object without a specified conversion. Likewise, F in (1) is a foreground role, while F_o in (2) maps a foreground vector f into the measured space.

For the comparison model, let s contain p source components, f contain q foreground components, and y_o contain m observed components. Then M_o has m rows and p columns; F_o has m rows and q columns; ε_o has m components. These dimensional declarations are editorial specifications of the comparison. No matrix dimensions are inferred for the complete T.

Every summand in a row of (2) must share that observation's units. If a source component has unit U_s and its observed contribution has unit U_y, the corresponding response coefficient carries U_y/U_s. Dimensionless coefficients are justified only when the quantities have already been expressed in compatible units. A multi-channel vector may contain different units by row; its covariance must retain the corresponding pairwise units.

Equation (2) does not explicitly include H and R. A particular application may fix them through controls or include their effects in a declared response model. It cannot treat their omission as proof of absence. Adding a residue term requires a state definition, units, update law and observability argument; otherwise the term only renames an unexplained residual.

Metrological terminology supplies a useful comparator: a measurement result includes relevant information alongside attributed values, and a measurement model relates the quantities involved in the measurement. The proposed Keddeh contribution is to retain these dependencies as inspectable computational state across subsequent transformations. This is an architectural objective, not a claim that established metrology ignores context. [4]

3. Derivation: one mixture and two independent transforms

The following derivation expands source equation (3) using conventional real linear algebra. Assume a noiseless comparison with p=q=m=n, known response matrices, and a fixed ordered unknown vector x formed by placing s before f. For one view, y=s+f, the design matrix is A₁=[I I]. It has n rows and 2n columns.

For any n-component vector h, replacing s by s+h and f by f−h leaves y unchanged. Equivalently, A₁ maps the nonzero vector (h,−h) to the zero vector whenever h is nonzero. These directions form an n-dimensional null space. The first I block also supplies n independent columns, so the rank is exactly n. The observation constrains a sum, not a unique pair.

Now add a second view y₂=s+2f while retaining y₁=s+f. Stacking the observations in the stated order produces the source's block matrix [[I,I],[I,2I]]. Subtracting the first block row from the second gives [[I,I],[0,I]]. The two diagonal identity blocks establish rank 2n. Direct reconstruction is

f=y₂−y₁; s=2y₁−y₂.

These reconstruction formulas are new editorial derivations of the original comparison equation. They use its declared conventional arithmetic. They neither reorder the native relation nor assert a general inverse for T.

A second copy of y₁ gives [[I,I],[I,I]], whose rank remains n. Repeated observations with independent noise may reduce uncertainty in the same measured combination, but repetition alone does not identify the two separate components. “Independent transforms” therefore means linearly informative responses here; it does not merely mean two files, instruments or acquisition times.

The rank result also assumes the responses are known. If the coefficient 2 is itself unknown and unconstrained, the two observations need not identify every source, foreground and calibration parameter. When an experiment estimates those additional quantities, its enlarged parameterisation needs a new identifiability analysis.

4. Worked comparison and uncertainty

4.1 New editorial example: exact source and foreground recovery

Choose one source component s=2 U and one foreground component f=3 U, where U is an arbitrary common signal unit. Let both responses be dimensionless. Then y₁=5 U and y₂=8 U. The reconstruction gives f=8−5=3 U and s=2×5−8=2 U. A pair such as s=6 U, f=−1 U would produce the same first observation but not the second.

Negative values in this illustration are conventional signed signal components. They are not claims about native active states or physical negative intensity. A domain permitting only nonnegative components should enforce that constraint explicitly; constraints can reduce a solution set without automatically making every mixture unique.

4.2 New editorial example: full rank with weak separation

Replace the second response by y₂=s+(1+δ)f. For nonzero δ, reconstruction becomes f=(y₂−y₁)/δ and s=y₁−f. The model is structurally identifiable, yet small δ amplifies the error in the difference.

For δ=0.001, the preceding source and foreground give y₁=5 U and y₂=5.003 U. Add an error of 0.000001 U only to y₂. The recovered foreground changes by +0.001 U and the source by −0.001 U. Full rank therefore establishes uniqueness in an idealised noiseless problem; it does not establish useful precision in the actual instrument.

For the original second response, suppose the two measurement errors are independent, zero-mean and each has variance σ². The reconstructed foreground error is ε₂−ε₁ and the source error is 2ε₁−ε₂. Their variances are respectively 2σ² and 5σ², and their covariance is −3σ². These are editorial consequences of the declared error assumptions. Correlated errors, uncertain calibration or model mismatch require their own covariance propagation; a rank count cannot remove them.

The negative covariance is instructive. Increasing the allocated foreground can decrease the inferred source while leaving their sum well constrained. A publication should therefore report the joint inference or relevant covariance, rather than presenting two marginal error bars as if the estimates were independent.

5. Architecture and cross-domain use

This formulation uses the state record defined in WP-01 §2, equation (3), under the domain-mapping conditions established in its §5:

X_t = ⟨id,q_t,payload_t,observer_t,environment_t,path_t,rule_t,proof_t⟩.

Applied here, the payload carries the measured or inferred quantity, the observer and environment remain explicit, and the path, rule and proof fields identify how the observation was produced. The imported conditions preserve anchor, adjacency, orientation and commit order; they do not substitute a state label for a measured payload. WP-01's historical anchor notation remains in its original edition and is not silently substituted for the current literal .1. [6, §§2, 5]

An observer-anchored record separates enduring subject identity from a particular observation, applicable definition revision, measured or inferred value, and carrier address. The record must identify the original source, acquisition context, response or transformation version, units, uncertainty and any declared omissions. Storing a new observation does not overwrite the subject or erase an earlier observation with the same numeric result. [1–3]

A useful processing sequence is: retain the supplied source; admit a domain-specific transformation; execute it with the declared inputs; observe its output; preserve the derivation and uncertainty; then compare the result under the intended equivalence rule. An exact byte digest establishes carrier identity, not calibration correctness or physical truth.

Consider an astronomical instrument, a sequencer and a virtual camera. Each has a subject, an observation channel and conditions, but those similarities do not license exchanging their response matrices. A sequencer's error process is not a telescope's sky foreground. A virtual camera's configured scene is not an independently observed physical scene. Transfer of the record contract is plausible; transfer of a scientific law requires a structure-preserving mapping and appropriate evidence.

The cosmological follow-on, WP-17, makes this separation concrete by retaining a baseline prediction plus distinct foreground, observer and dynamic residue terms. Its statistical extension is a separate local model. Neither white paper converts a common vocabulary into evidence of a new physical mechanism. [5]

6. Executed editorial checks and proposed experiments

On 18 October 2026, this edition's arithmetic was checked locally in Python 3.12.14 using standard-library exact rational arithmetic. Gaussian elimination checked the single, repeated and independently transformed matrices for n=1, 2 and 3. All nine ranks matched equation (3) and its repeated-view counterpart. Three additional cases checked exact reconstruction, unchanged observations under selected null-space shifts, and the near-dependent error example. The retained receipt records 12 passed cases and zero failures.

The finite rank cases corroborate the implementation of the arithmetic; the general rank argument is the derivation in Section 3. These runs did not exercise a physical instrument, estimate a real foreground, evaluate native relational operators or execute the full T. The original paper's shared series calculation-receipt identifier is retained in the unchanged source copy. Its underlying archive was not rerun or treated as the receipt for these new calculations.

The source's experimental programme becomes actionable through four separate gates:

1. Mixture gate. Generate known source and foreground vectors under declared responses. Test full-rank, rank-deficient and nearly dependent cases. Report reconstruction error, identifiability and sensitivity, including failed decompositions.
2. Calibration gate. Use independently characterised instrument inputs. Specify admissible bias and uncertainty coverage before fitting. Estimate calibration parameters where required, and propagate their uncertainty rather than treating them as known after calibration.
3. Frame gate. Express the same observation in equivalent declared frames, transform all relevant model terms consistently, and compare reconstructed observables. A changed coordinate label must not create a physical effect.
4. Held-out gate. Freeze the fitted model before applying it to a reserved observer or environment. Account for shared errors between training and test data. Evaluate predictive error and uncertainty calibration against a baseline with equivalent information.

These are planned domain tests, not completed results of E2. The outcomes should remain separate: successful record preservation cannot compensate for a failed calibration or unsupported physical inference.

7. Claims, failure conditions and conclusion

O-01: Formalised comparison result. One blended source/foreground view is underdetermined under the declared linear assumptions. The null-space derivation and exact example establish this bounded claim.

O-02: Conditional recovery result. A second response restores structural identifiability when the stacked design has full column rank. Practical recovery also depends on conditioning, uncertainty and response knowledge.

O-03: Architectural boundary. Observer metadata makes a calculation inspectable; it does not by itself prove a physical theory or complete native-algebra implementation.

A source-resolved claim fails when the data cannot distinguish the proposed components. A residue interpretation fails if it merely absorbs known foreground without an independent discriminant. A physical mapping fails if predictions change solely because equivalent coordinate labels were exchanged. An implementation fails its representation contract when it collapses unavailable observation into numerical zero, changes .1 into a number, or identifies distinct ordered records solely by their equal projections.

Observer anchoring is thus a concrete discipline for preserving the conditions of a claim. Its conventional mixture example shows how different observation operators alter what can be inferred. Its broader research value depends on a domain making the translation explicit and demonstrating that the retained distinctions improve reproducibility, recovery or independently tested prediction.

References and edition custody

[1] A. Keddeh. *Observer-Anchored Measurement Algebra*, KEX-WP-05, 19 July 2026, §§1–2, component/state table and falsifiers. Source document. The supplied text extraction is retained unchanged alongside this edition.

[2] Aboudy Keddeh. Current author interpretation directions, retained in *Observation and Identity Review*, R3, current at 18 October 2026. Owner-supplied clarification; exact source identity is recorded in this edition's metadata.

[3] A. Keddeh. *Observer Relational Mathematics*, Publication Extension R1, 10 October 2026, §§1–8. Source document.

[4] JCGM. *International Vocabulary of Metrology*, JCGM 200:2012, §2.9, measurement result and §2.48, measurement model. Accessed 18 October 2026.

[5] A. Keddeh. *Observer, Foreground, and Residue in Cosmological Inference*, KEX-WP-17, 19 July 2026. Source document.

[6] A. Keddeh. *KEX Active-State Foundations*, KEX-WP-01, 19 July 2026, §2 equation (3) and §5. Source document.

[7] A. Keddeh. *KEX/BRAINK Monograph Draft XXXVII*, §XXV.8, “Mathematical Working: Observer-State Translation.” Source document. Full readable source retrieved 18 October 2026; relevant source passage retained with this edition.

This edition is accompanied by its exact source extraction, source digest, change log, source metadata and new calculation receipt. The master and monograph citations supplied by the July paper remain in its preserved original; E2 does not represent those entire works as freshly verified.

© 2026 Aboudy Keddeh | Publisher: Keddeh Systems | KEX-WP-05 | E2 | Internal research
Licence profile: KEDDEH-RESERVED-1.0. All rights reserved. No licence is granted by this document. Separate written permission governs reuse, redistribution and implementation, subject to applicable law. Third-party material retains its own rights. Authorship, copyright ownership and legal publishing capacity are separate records; no assignment of rights, patent, trademark or software licence is asserted.