Draft arguments extracted for conceptual review. Required evidence holes and proposed peer links remain unratified. These files have not been admitted to the pattern library or used to dispatch work.
Source: walkthrough, SHA256 f141b3044edcd27478c98dec480e73d45bf4fad0f70c238407c8ad457f09003e.
Independent arguments for conceptual review. These are authored draft
extractions, not validated solutions or execution instructions. Required evidence
holes remain explicit. Cross-references are proposed peer connections (@see-also),
not assertions that AIF formulas, historical deployment claims or parameter choices
have been adopted. No @why authority edge is invented.
meaning-before-carrier: aif/belief-state-operational-hypotheses, agent/state-is-hypothesis. These supply revisable belief discipline; neither supplies the seven-status mission mapping.
outcome-meaning-across-models: aif/declare-the-conditioning, aif/term-to-channel-traceability. Reuse conditioning and term-input traceability; this pattern asks for the semantic justification of domain changes.
observation-can-disagree: aif/no-self-certification, aif/structured-observation-vector. Reuse independence and typed observations; do not reproduce their mechanisms or imply that an independence tag alone proves origin.
name-the-uncertainty-to-learn: aif/two-layer-calibration, aif/structure-learning-by-model-reduction. These distinguish evaluation and model change; their historical examples do not settle this parameter family or supply a new EIG formula.
References resolve against futon3/library/; the collection stays beside its blog source until reviewed. Admission into the library graph, edge attestation and runtime use have not occurred.
Preserve Probability Laws Across Representations. This extends correspondence to probability carriers; shared executable code alone is not a probability-law proof.
Distinguish Missing Observations from Observed Zeros. Typed observations and independence are reused; neither chooses sparse-data priors.
Three draft arguments cover prospective selection, exact independent review and the authority of paired labels. They extend the WM-04 design; they are not extractions from the original blog or independently accepted patterns.
Draft arguments for the WM-02 build design, not independent validation or original-blog extraction.
Separate Numeric Admission from Exact Equality is a draft argument for the WM-01 design. It does not add a new tolerance or establish a correspondence theorem.
@flexiarg wm-prediction/meaning-before-carrier @title Explain the Difference Before Choosing the Carrier @style pattern @grade technique @draft codex-28-narrative-extraction-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance p4ng/wm-walkthroughs/01-prediction.md; section: The state is a claim about that situation; sha256: f141b3044edcd27478c98dec480e73d45bf4fad0f70c238407c8ad457f09003e @see-also aif/belief-state-operational-hypotheses agent/state-is-hypothesis ! conclusion: Explain the Difference Before Choosing the Carrier. + context: A predictor is being extended from one kind of work item to another. + IF: The selected mission or ticket has no entry in the existing belief representation. + HOWEVER: Reusing an available status table restores lookup success cheaply, but its distinctions may describe a different problem. + THEN: Contrast two real-world situations the decision must distinguish; explain what each proposed state means and what could contradict it. Choose or revise the carrier only after that account is coherent. Keep a proposed registry mapping explicitly unresolved until its interpretation is justified. + BECAUSE: A model can fail by erasing a relevant distinction even when every mass is normalized and every identifier resolves. + example: Illustrative, not an executed result. A complete registry flag and a reviewer who cannot resume are compatible facts. An encoding that treats the flag as proof of useful handoff misses the problem. + reference-scope: These supply revisable belief discipline; neither supplies the seven-status mission mapping. ? evidence(required): No independent use evidence or attestation is claimed. The blog is the source of this proposed argument, not validation of its response. + NEXT-STEPS: Conceptual review question: What observation could show that the proposed mission-status interpretation is wrong? No implementation or dispatch is authorized by this draft.
@flexiarg wm-prediction/outcome-meaning-across-models @title Carry the Outcome Meaning Through Prediction @style pattern @grade technique @draft codex-28-narrative-extraction-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance p4ng/wm-walkthroughs/01-prediction.md; section: Follow one prediction from beginning to end; sha256: f141b3044edcd27478c98dec480e73d45bf4fad0f70c238407c8ad457f09003e @see-also aif/declare-the-conditioning aif/term-to-channel-traceability ! conclusion: Carry the Outcome Meaning Through Prediction. + context: A design is evaluated through state transitions, observations and preferences. + IF: Different components use document features, mission statuses and terminal dispositions. + HOWEVER: Adapters can make the signatures fit while discarding the very differences that the design should affect. + THEN: Follow one candidate from its interpreted action through predicted state and observation to the preferred outcome. Explain each change of domain and what information it preserves. Refuse to treat a constant bridge as evidence of outcome-dependent prediction. + BECAUSE: Preference comparisons concern the predicted outcome only when the model carries the same meaning through the chain. + example: Illustrative, not an executed result. Producing a handoff file predicts a document, not automatically a successor able to resume responsibly. + reference-scope: Reuse conditioning and term-input traceability; this pattern asks for the semantic justification of domain changes. ? evidence(required): No independent use evidence or attestation is claimed. The blog is the source of this proposed argument, not validation of its response. + NEXT-STEPS: Conceptual review question: Which changed prediction would make the outcome bridge produce a different result, and why? No implementation or dispatch is authorized by this draft.
@flexiarg wm-prediction/observation-can-disagree @title Let the Observation Contradict the Belief @style pattern @grade technique @draft codex-28-narrative-extraction-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance p4ng/wm-walkthroughs/01-prediction.md; section: The state is a claim about that situation; sha256: f141b3044edcd27478c98dec480e73d45bf4fad0f70c238407c8ad457f09003e @see-also aif/no-self-certification aif/structured-observation-vector ! conclusion: Let the Observation Contradict the Belief. + context: A team needs observations to evaluate whether a chosen design helped. + IF: Registry metadata and the agent belief are readily available but an independent result is scarce. + HOWEVER: Using those available values as ground truth manufactures agreement and conceals insufficient evidence. + THEN: Identify what an observer can inspect independently of the prediction and which criterion that inspection addresses. Keep registry facts, beliefs and observed labels separate; retain an insufficient result without turning it into a categorical label. + BECAUSE: An observation can test a prediction only if agreement was not built into its source. + example: Illustrative, not an executed result. A review may confirm that the evidence is inaccessible. Acceptance of that finding does not establish that the mission succeeded. + reference-scope: Reuse independence and typed observations; do not reproduce their mechanisms or imply that an independence tag alone proves origin. ? evidence(required): No independent use evidence or attestation is claimed. The blog is the source of this proposed argument, not validation of its response. + NEXT-STEPS: Conceptual review question: Could the observation procedure return a result that the current belief strongly disfavours? No implementation or dispatch is authorized by this draft.
@flexiarg wm-prediction/name-the-uncertainty-to-learn @title Name What a New Observation Could Teach @style pattern @grade technique @draft codex-28-narrative-extraction-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance p4ng/wm-walkthroughs/01-prediction.md; section: Learning requires another distinction; sha256: f141b3044edcd27478c98dec480e73d45bf4fad0f70c238407c8ad457f09003e @see-also aif/two-layer-calibration aif/structure-learning-by-model-reduction ! conclusion: Name What a New Observation Could Teach. + context: A proposed action is justified by what the machine expects to learn. + IF: The situation is uncertain and the reliability of the observation model is also uncertain. + HOWEVER: A single confidence number hides whether work should investigate the situation or revise the model. + THEN: Name the hidden condition and the separate model parameter in words, then describe an observation that changes each belief. Require compatible priors, likelihoods and counterfactual posteriors before assigning information value; leave the parameter family open if it is not established. + BECAUSE: The reason to gather evidence depends on what uncertainty that evidence can actually reduce. + example: Illustrative, not an executed result. An incomplete handoff and a review procedure that routinely misses omissions are different explanations of a failed resumption. + reference-scope: These distinguish evaluation and model change; their historical examples do not settle this parameter family or supply a new EIG formula. ? evidence(required): No independent use evidence or attestation is claimed. The blog is the source of this proposed argument, not validation of its response. + NEXT-STEPS: Conceptual review question: What would distinguish learning the mission condition from learning the reliability of its test? No implementation or dispatch is authorized by this draft.
@flexiarg wm-prediction/probability-laws-with-runtime-meaning @title Preserve Probability Laws Across Representations @style pattern @grade technique @draft codex-28-checklist-pattern-gap-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance Checklist revision 4 (WM-01), p4ng b2c7d1d; added during Joe's requested indicative-pattern mapping, not observed use. @see-also aif/shared-kernel-predictive-forward-model ! conclusion: Preserve Probability Laws Across Representations. + context: A proved probability carrier is represented by executable numeric code. + IF: A repaired proof or a newly normalized runtime table is offered as model validity. + HOWEVER: Duplicate support, mass outside support, negative precision or unaccounted numeric error can change the law despite familiar type names. + THEN: Name the common support and laws, pin the repaired definitions, and compare the runtime interpretation against them under an explicit numerical criterion. Keep rejection controls for invalid support and mass, and distinguish author validation from independent acceptance. + BECAUSE: Probability calculations only compose as the intended model when their representations preserve its laws. + example: Illustrative only. A table summing to one still fails if it carries hidden mass outside the support used by its consumer. + reference-scope: This extends correspondence to probability carriers; shared executable code alone is not a probability-law proof. ? evidence(required): This is an unreviewed design argument. No implementation, attestation or successful use is claimed. + NEXT-STEPS: Which concrete invalid distribution does each representation reject? Conceptual review only; no build dispatch.
@flexiarg wm-prediction/unobserved-is-not-zero @title Distinguish Missing Observations from Observed Zeros @style pattern @grade technique @draft codex-28-checklist-pattern-gap-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance Checklist revision 4 (WM-04), p4ng b2c7d1d; added during Joe's requested indicative-pattern mapping, not observed use. @see-also aif/structured-observation-vector aif/no-self-certification ! conclusion: Distinguish Missing Observations from Observed Zeros. + context: An observation model is assembled from sparse independent records. + IF: A state row has no eligible observations or an observed outcome has zero frequency. + HOWEVER: Encoding both as zero hides different evidence conditions; automatic smoothing can silently invent a supported model. + THEN: Retain the acquisition denominator, row coverage, observation eligibility and insufficiency findings. Distinguish no eligible observations from an observed zero-frequency cell. Refuse unsupported assembly and identify needed collection; any prior or smoothing requires its own explicit authority. + BECAUSE: An empirical zero is a statement about observations, whereas missing coverage supplies no frequency estimate. + example: Illustrative only. Seven reviewed insufficiency findings supply no seven-row likelihood table. + reference-scope: Typed observations and independence are reused; neither chooses sparse-data priors. ? evidence(required): This is an unreviewed design argument. No implementation, attestation or successful use is claimed. + NEXT-STEPS: Can the consumer distinguish an unobserved row from a sampled row with a zero count? Conceptual review only; no build dispatch.
@flexiarg wm-prediction/choose-observations-before-outcomes @title Choose Observations Before Seeing Their Outcomes @style pattern @grade technique @draft codex-28-wm04-cascade-design-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance WM-04 checklist obligation; p4ng/wm-walkthroughs/cascades/wm-04/README.md. Authored design, not observed use. @see-also wm-qualifying-run/complete-acquisition-universe aif/measurement-window-hygiene ! conclusion: Choose Observations Before Seeing Their Outcomes. + context: An observation model needs production data that can support claims beyond the conveniently retained successes. + IF: An operator can select which closes to inspect and some outcomes are already visible. + HOWEVER: Choosing after seeing the outcome biases the sample even when every individual annotation is honest. + THEN: Freeze the target population, inclusion rule, sampling procedure and stopping rule before revealing outcomes. Register every selected occurrence, including unavailable evidence and exclusions. Separate authority to commission an observer from authority to classify the result; preserve any departures as departures rather than rewriting the plan. + BECAUSE: A prospective selection record makes preferential acquisition detectable and gives missingness a denominator. + example: Illustrative only. A predeclared cohort includes an inaccessible close. Keep that selected occurrence as unavailable; do not replace it with a successful close. + reference-scope: Reuse the cited arguments within their stated scope; no external parameter, historical deployment claim or admission is inherited. ? evidence(required): Independent conceptual review and actual use evidence remain absent. This argument does not establish an implemented acquisition path or a measured model. + NEXT-STEPS: Challenge whether this response could conceal an unsupported label or an omitted occurrence. No implementation dispatch is authorized.
@flexiarg wm-prediction/review-the-exact-observation @title Review the Exact Observation and Its Evidence @style pattern @grade technique @draft codex-28-wm04-cascade-design-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance WM-04 checklist obligation; p4ng/wm-walkthroughs/cascades/wm-04/README.md. Authored design, not observed use. @see-also aif/no-self-certification aif/structured-observation-vector ! conclusion: Review the Exact Observation and Its Evidence. + context: An observer submits a classification for independent acceptance. + IF: The reviewer can approve a statement while the subject, cutoff or evidence changes underneath it. + HOWEVER: An acceptance attached only to a mission name can accidentally approve a different occurrence or later evidence. + THEN: Bind observer authority, rubric revision, exact entity and occurrence, immutable cutoff, admitted evidence bytes and the complete finding into one identified observation. Have a distinct authorized reviewer accept or reject those exact bytes. Corrections create a successor record and fresh review; retain the original. Acceptance of insufficiency is never acceptance of a categorical label. + BECAUSE: Independent review concerns a particular evidential claim, not the reputation of its author or the continued existence of its target. + example: Illustrative only. A reviewer accepts that the pre-cutoff evidence is insufficient. A later successful result cannot retroactively turn that record into a positive label. + reference-scope: Reuse the cited arguments within their stated scope; no external parameter, historical deployment claim or admission is inherited. ? evidence(required): Independent conceptual review and actual use evidence remain absent. This argument does not establish an implemented acquisition path or a measured model. + NEXT-STEPS: Challenge whether this response could conceal an unsupported label or an omitted occurrence. No implementation dispatch is authorized.
@flexiarg wm-prediction/pair-state-and-outcome-without-belief-labels @title Establish Both Sides of an Observation Pair @style pattern @grade technique @draft codex-28-wm04-cascade-design-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance WM-04 checklist obligation; p4ng/wm-walkthroughs/cascades/wm-04/README.md. Authored design, not observed use. @see-also wm-prediction/meaning-before-carrier wm-prediction/observation-can-disagree ! conclusion: Establish Both Sides of an Observation Pair. + context: A conditional observation model is estimated from records labelled by state and observed outcome. + IF: The machine already has a belief vector and the close record already has a disposition. + HOWEVER: Using belief argmax as the state label makes the model certify its own hypothesis; a disposition alone supplies no state-conditioned sample. + THEN: Declare what state and observation mean at the same reference time and scope. Obtain an independently justified state classification and an admitted outcome for the exact occurrence under separately documented criteria. Do not reveal the posterior or close disposition to the state annotator. Join only accepted compatible records; retain unresolved state, outcome or timing as unavailable pairs. If the state cannot be independently classified, surface the estimation problem for model review rather than fabricating labels. + BECAUSE: Counting requires justified conditioning labels as well as outcomes; arithmetic cannot create either authority. + example: Illustrative only. A handoff has a retained completion disposition but its criterion evidence is insufficient. It contributes to the acquisition ledger, not a cell of A. + reference-scope: Reuse the cited arguments within their stated scope; no external parameter, historical deployment claim or admission is inherited. ? evidence(required): Independent conceptual review and actual use evidence remain absent. This argument does not establish an implemented acquisition path or a measured model. + NEXT-STEPS: Challenge whether this response could conceal an unsupported label or an omitted occurrence. No implementation dispatch is authorized.
@flexiarg wm-prediction/belief-keeps-its-lineage @title Keep the Lineage of the Belief Being Consumed @style pattern @grade technique @draft codex-28-wm02-cascade-design-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance WM-02 checklist; p4ng/wm-walkthroughs/cascades/wm-02/README.md. Authored design, not observed use. @see-also aif/belief-state-operational-hypotheses wm-connected-learning/observation-to-next-choice ! conclusion: Keep the Lineage of the Belief Being Consumed. + context: A predictor needs the current posterior for a selected work item. + IF: A normalized row can be read from several snapshots or recreated when a lookup fails. + HOWEVER: The same numbers can represent an admitted prior, a stale posterior or a fabricated replacement; normalization cannot distinguish them. + THEN: Retain the entity domain, prior authority, predecessor snapshot, ordered admitted update identities, model revision and tick boundary with the posterior. Bind the consumer to that exact snapshot. Separate genuine initialization, explicit migration and unavailable carry; never silently reset missing or unreadable belief. + BECAUSE: A prediction is conditioned on a particular history of information, not merely a vector with a unit sum. + example: Illustrative only. A completed mission leaves the live selector but its belief history remains needed for outcome interpretation. A disappearing lookup is not authority to initialize it again. + reference-scope: Reuse the cited arguments; no prior values, state mapping, independence assumption or live correspondence is inherited. ? evidence(required): Independent conceptual review and use evidence remain absent. No successful construction or execution is claimed. + NEXT-STEPS: What retained evidence distinguishes the intended current belief from a plausible substitute? No build dispatch is authorized.
@flexiarg wm-prediction/numeric-admission-is-not-exact-equality @title Separate Numeric Admission from Exact Equality @style pattern @grade technique @draft codex-28-wm01-cascade-design-2026-09-15 @audience War Machine designers and conceptual reviewers @provenance WM-01 checklist; p4ng/wm-walkthroughs/cascades/wm-01/README.md. Authored design, not observed use. @see-also wm-prediction/probability-laws-with-runtime-meaning wm-qualifying-run/certificate-and-world-must-meet ! conclusion: Separate Numeric Admission from Exact Equality. + context: Runtime probability values are connected to a theorem over exact normalized kernels. + IF: The runtime admits a floating-point row whose sum lies within an approved tolerance of one. + HOWEVER: Approximate admission does not make that row exactly normalized or justify downstream error bounds. + THEN: Retain the actual numeric representation, admission rule and conversion path. State the exact relation between admitted runtime values and the formal object, and prove or independently validate the required bounds for each consumer. Refuse unaccounted conversions; do not silently renormalize or widen tolerance to obtain a proof. + BECAUSE: A tolerance answers whether a value may enter an interface, while correspondence explains what mathematical claim remains true of computations using it. + example: Illustrative only. A row totaling slightly below one may pass a reviewed float gate; a certificate cannot then assert exact normalization of those unchanged real-valued masses. + reference-scope: Existing carrier and certificate patterns supply the surrounding obligations; this argument introduces no epsilon, normalization rule or approximate theorem. ? evidence(required): Independent review and implementation evidence remain absent. No exact/float correspondence is established by this draft. + NEXT-STEPS: Which specific theorem applies to the numbers actually retained by the consumer, and under which assumptions? No build dispatch is authorized.