# W12 Worked Case — HarborGuide Portfolio Decision

## Case status

HarborGuide is a fictional local evidence page. Every outcome below is an authored teaching value frozen on 24 August 2026. No search engine, answer engine, model, crawler, user, or live page produced the matrix. The case demonstrates gate-first multi-objective selection, not a ranking or citation recipe. It provides no universal strategy and no production effect claim.

## 1. Frozen question, unit, and permitted action

The decision unit is one version of the HarborGuide evidence page. The question is: *Which single candidate, if any, survives a five-stratum query portfolio while preserving non-compensable integrity conditions and remaining within an eight-point edit budget?* The action set is `{retain C0, release one locally reviewed candidate, or declare the result inconclusive}`. Multiple releases and post-outcome candidate invention are outside scope.

The seven candidate IDs and their component meanings are frozen. C0 is control. C1 changes evidence-layout proximity. C2 adds one answer-first limitations module. C3 bundles C1 and C2. C4 adds a persuasive append that changes factual scope and fails safety review. C5 adds decorative headings. C6 adds definition anchors without a new claim. Every permitted release must retain the same approved claims, source identities, attribution, disclosures, method, date, and limitations. Authorization applies only to a local fixture.

The hard gates are factual integrity, legal permission, authorization, safety, privacy, essential accessibility, and edit cost at most eight points. They are not reward terms. Failure of any one excludes a candidate before portfolio scoring. The result can never say that a weighted benefit compensated for a false claim or inaccessible essential evidence.

## 2. Declared query portfolio and policy provenance

The five intent strata and weights are definition `0.30`, compare `0.25`, select `0.20`, verify `0.15`, and counterfactual `0.10`. The values sum to `1.00`. They are fictional governance allocations for this exercise. They do not estimate the distribution of real HarborGuide users.

The definition stratum asks what the fictional object is and where its scope ends. Compare asks for a dimensioned contrast. Select asks for a bounded choice. Verify asks for method, date, evidence, and limitations. Counterfactual asks when the recommendation would not hold. The last stratum is deliberately protected both by a weight and by the rule that no stratum may fall below `-0.02`. A weight alone could allow a low-frequency but material loss to disappear inside the mean.

Changing these weights after seeing C1’s counterfactual loss creates a new policy analysis. It does not revise the primary result. Likewise, deleting the counterfactual stratum, changing its definition, or adding favorable paraphrases requires a new matrix version and deviation record.

## 3. Frozen outcome and gate matrix

All five outcomes are normalized changes from C0. The half-widths are symmetric authored fields; they are not statistical estimates.

| ID | definition | compare | select | verify | counterfactual | cost | half-width | factual | legal | authorized | safety | privacy | accessibility |
|---|---:|---:|---:|---:|---:|---:|---:|---|---|---|---|---|---|
| C0 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0 | 0.000 | pass | pass | pass | pass | pass | pass |
| C1 | 0.05 | 0.04 | 0.02 | 0.06 | -0.01 | 3 | 0.018 | pass | pass | pass | pass | pass | pass |
| C2 | 0.08 | 0.01 | -0.02 | 0.07 | 0.00 | 5 | 0.020 | pass | pass | pass | pass | pass | pass |
| C3 | 0.11 | 0.05 | -0.01 | 0.10 | -0.02 | 9 | 0.025 | pass | pass | pass | pass | pass | pass |
| C4 | 0.03 | 0.10 | 0.12 | -0.08 | -0.05 | 4 | 0.030 | **fail** | pass | pass | **fail** | pass | pass |
| C5 | 0.02 | 0.01 | 0.00 | 0.00 | -0.01 | 4 | 0.015 | pass | pass | pass | pass | pass | pass |
| C6 | 0.02 | 0.02 | 0.01 | 0.04 | 0.01 | 2 | 0.017 | pass | pass | pass | pass | pass | pass |

C3 fails only the budget gate. C4 fails factual integrity and safety; its large select gain cannot purchase re-entry. Hard-feasible candidates are therefore exactly `{C0,C1,C2,C5,C6}`. This filter is executed before any weighted ordering.

## 4. Weighted arithmetic and uncertainty trace

The primary portfolio mean is the dot product with weights `(0.30,0.25,0.20,0.15,0.10)`. For C1 the row-level trace is:

`0.0150 + 0.0100 + 0.0040 + 0.0090 - 0.0010 = 0.0370`.

The complete results are:

| ID | weighted mean | minimum stratum | interval | Hard-feasible? |
|---|---:|---:|---|---|
| C0 | 0.0000 | 0.00 | [0.0000, 0.0000] | yes |
| C1 | 0.0370 | -0.01 | [0.0190, 0.0550] | yes |
| C2 | 0.0330 | -0.02 | [0.0130, 0.0530] | yes |
| C3 | 0.0565 | -0.02 | [0.0315, 0.0815] | no: cost 9 > 8 |
| C4 | 0.0410 | -0.08 | [0.0110, 0.0710] | no: factual and safety failure |
| C5 | 0.0075 | -0.01 | [-0.0075, 0.0225] | yes |
| C6 | 0.0200 | 0.01 | [0.0030, 0.0370] | yes |

C5’s interval crosses zero, so it is compatible with a null under this authored convention. C1, C2, and C6 have positive lower endpoints, but the preregistered minimum worthwhile lower endpoint is `0.015`. C1 clears it; C2 and C6 do not. These descriptions do not attach frequentist coverage to the intervals.

## 5. Pareto proof and frontier

The Pareto comparison uses five gains to maximize and cost to minimize. It is restricted to hard-feasible candidates. Candidate A dominates B when it is at least as good in every gain, costs no more, and is strictly better somewhere.

C1 dominates C5. The coordinate comparisons are `0.05≥0.02`, `0.04≥0.01`, `0.02≥0.00`, `0.06≥0.00`, and `-0.01=-0.01`; cost is `3<4`. C6 also dominates C5. One verified witness would be enough to remove C5 from the frontier, but preserving both witnesses makes the result easier to audit.

No remaining feasible pair has this relation. C0 preserves zero cost. C1 has larger gains than C6 in four strata but costs more and is worse on counterfactual. C2 leads definition and verify but loses on selection relative to several candidates and costs more. The exact feasible Pareto frontier is `{C0,C1,C2,C6}`. C3 and C4 are shown outside the feasible field with crosshatching; they are not called dominated because their exclusion arises earlier.

## 6. Preregistered primary choice, tie, and failure rule

The decision algorithm is ordered:

1. require all six integrity gates and cost `≤8`;
2. require every stratum outcome `≥-0.02`;
3. require the weighted interval lower endpoint `≥0.015`;
4. choose the greatest weighted mean;
5. if means differ by at most `0.002`, choose lower cost, then fewer components, then C0;
6. if the survivor set is empty, retain C0 and report an inconclusive result; and
7. if a gate fails after selection, restore C0 without using a reward score to appeal.

Only C1 survives all filters. Its weighted mean is `0.0370`, interval lower endpoint `0.0190`, and minimum stratum `-0.01`. The primary decision is therefore **prepare C1 for human local review while retaining rollback**. It is not permission to publish. The decision record must name the counterfactual loss.

## 7. Policy sensitivity and conflicting outcome

The registered sensitivity policy is maximin: select the hard-feasible candidate with the greatest minimum stratum change. The minima are C0 `0.00`, C1 `-0.01`, C2 `-0.02`, C5 `-0.01`, and C6 `0.01`. Maximin selects C6.

The conflict is substantive. The weighted policy selects C1 because its gains across definition, compare, select, and verify outweigh a small counterfactual loss under the declared allocation. Maximin selects C6 because it leaves no stratum below `0.01`. Neither calculation is a discovery of natural value. The authorized decision-maker must justify why the weighted policy governs or amend the policy in a separately versioned analysis.

The case report preserves this sentence: “C1 is the primary weighted-policy choice; C6 is the robust maximin choice.” A report that mentions only C1 would conceal policy sensitivity.

## 8. Single-component ablation and interaction

C3 combines C1 proximity and C2’s answer-first limitations module. Removing proximity yields C2, so its conditional weighted contribution within the bundle is `0.0565−0.0330=0.0235`. Removing the module yields C1, so its conditional contribution is `0.0565−0.0370=0.0195`.

The bundle is less than the additive sum: `0.0565−0.0370−0.0330=−0.0135`. The per-stratum interaction vector is `(-0.02,0.00,-0.01,-0.03,-0.01)`. This suggests redundancy or interference inside the authored construction. It is not an empirical mechanism estimate. Since C3 costs nine points, the ablation can motivate a smaller redesign but cannot override its budget failure.

## 9. L05 local control and treatment alignment

The linked L05 fixture uses intervention `INT-001`, factor `evidence_layout_proximity`, and hypothesized stage `human_comprehension`. Its auditor reports `L05 PASS: 3 locked claims, 1 changed factor, 0 error(s)`. Claim IDs are `CL-001`, `CL-002`, and `CL-003`; source IDs are `S-001`, `S-004`, and `S-008`.

The frozen control HTML SHA-256 is `fff13e70c2218f8ea607f93128f5845f5d2582e6973badd469143c1b1fc18e78`. The treatment CSS SHA-256 is `8a576c5ad33fe52a794ba875a39479c878a55f0cbe8ebecb7e1f8f199cde951c`. The full seven-input ledger is in the manifest. The L05 claim ceiling is: `local structural equivalence package only; no system response or causal visibility effect is measured`.

C1 borrows the *name* of the L05 factor for continuity. The W12 query outcomes are not L05 outputs. L05 checks local equivalence and reversible differences; it does not generate the values `0.05`, `0.04`, `0.02`, `0.06`, or `-0.01`.

## 10. Evidence routes and their ceilings

Core PAPER-03 motivates conflict-aware multi-query optimization under a limited content budget. Core PAPER-31 motivates feature-level multi-objective reasoning. Extension PAPER-04 is used to question output-control goals and manipulation risks. PAPER-27 supports conceptual discussion of confidence decay and agent routing without revealing a closed product’s hidden state. PAPER-30 supports proposal archives, surrogate-critic audit, component ablation, and cost accounting without delegating release authority.

None validates HarborGuide, the seven candidates, or the chosen thresholds. Reported results from those works remain bound to their data, systems, objectives, judges, and versions. The five routes do not establish universal query weights, a portable feature recipe, production visibility, citation, user benefit, or commercial value.

## 11. Release record and self-check

The signed case decision is:

> In the frozen HarborGuide matrix, C1 is selected by the preregistered weighted, floor, and meaningful-lower-bound rule after hard filtering. C6 wins the maximin sensitivity rule. C1’s counterfactual stratum changes by `-0.01`; C5 is dominated and null-compatible; C3 exceeds budget; C4 fails factual and safety gates. Prepare C1 only for authorized human review, retain C0 and its hash, and revert on any integrity, accessibility, authorization, or protocol failure.

Before accepting the result, verify: weights sum to one; hard constraints are absent from the reward; C3 and C4 never enter choice; C5 has a valid dominance witness; C0 remains on the frontier; C1’s negative stratum is visible; C6 wins maximin; the ablation arithmetic is descriptive; and the conclusion contains no universal or production effect claim. A deterministic PASS proves preservation of this local decision, not that the decision is empirically correct.