confidence-engine/docs/current-handoff.md

Current Handoff — Confidence Engine

Role: Concise operational snapshot for resuming work today. Not a historical diary. The design evolution archive index at docs/design-evolution/README.md provides progressive loading of experiment history; load the relevant chapter only when a specific historical question requires it.

Repository checkpoint

• Branch: feature/investigation-report-v0.55
• HEAD: 37a9a12 — route design evolution provenance through archive index
• Working tree: will be clean after v0.59a commit

v0.5 initial reconstruction relationship contract

• v0.5 first-class initial reconstruction relationship contract implemented.
• Reconstruction relationships use explicit fromId/toId references.
• Declared relationships project deterministically to existing SituationGraph edges.
• Invalid relationship references are skipped rather than inferred or repaired.
• SituationGraph schema unchanged; v0.5 is the production default.
• Targeted deterministic suite passed; live v0.5 semantic/reliability validation remains pending.
• 502 structured-output reliability remains a separate unresolved issue.

Current product architecture

Three distinct routes, not a single page:

/                              → Portfolio (notebook index)
/investigations/case-1          → Investigation (working case/pages)
/investigations/case-1/report   → Investigation Report (readable derived summary)

Portfolio = investigator notebook index. Shows the single canonical persisted investigation card with actions: View report, Continue investigation, Restart investigation. Below the card: + Create new investigation (portfolio-level, not inside the card).

Investigation = working case. Contains ScenarioForm + ReasoningWorkspace. Handles graph reasoning, focused investigation turns, Done/Re-open semantics, Current Understanding synthesis. No Report presentation — that is owned by the dedicated Report route.

Report = derived artefact. Renders persisted investigationReport snapshot. Generation is on-demand, triggered by the Report page itself (not ReasoningWorkspace). Exactly one /api/cases/overview call on first visit; zero calls on subsequent visits. The Report is not canonical reasoning evidence — it is a derived summary for review/export/use.

Current working product journey

Portfolio (/)
  → "Continue investigation"
    → Investigation page (/investigations/case-1)
      → Focused question asked → user answers → Done for now
      → Current Understanding synthesizes
      → If zero Open Questions: "Review current understanding" appears
        OR → "Review current understanding"
          → Report page (/investigations/case-1/report)
            → Generates via /api/cases/overview (once)
            → Persists investigationReport
            → Subsequent visits render persisted snapshot (zero calls)

Restart flow

Portfolio card Restart investigation → confirmation dialog (title: "Restart this investigation?") → destructive second confirmation → clearInvestigation() (canonical localStorage clear seam). No direct storage-key manipulation.

Current reasoning / ownership invariants

Evidence discipline:

RAW USER EVIDENCE
≠ MODEL-DERIVED CONTRIBUTION SEMANTICS
≠ CURRENT CANONICAL FINDING
≠ IMMUTABLE SOURCE OBSERVATION
≠ USER DISPOSITION / AUTHORITY
≠ TURN CONTEXT / PROVENANCE

Finding dispositions: null (eligible working premise), agree (user-endorsed), not_relevant (excluded from eligible reasoning, provenance retained), corrected Finding. User disposition never directly mutates authoritative graph state.

Evidence distinctions preserved by the persistence layer. Contributions preserve: question/context, verbatim answer, model observations, uncertainties, assumptions, relationships, follow-ups, target/provenance, sequence/order. Findings preserve: canonical proposition, sourceObservation, contributionId, originatingTargetNodeId, userDisposition.

Focused investigation presentation ownership (v0.52)

FocusedQuestionBody derives a thread-local subset (targetNodeId || originatingTargetNodeId) for every presentation surface. Previously answered content does NOT bleed from one question to another. Verified manually and by targeted Vitest.

Empty Done + Re-open semantics (v0.53)

• Empty Done is valid: parks/resolves the question locally, does NOT invoke episode processing, does NOT produce no_episodic_content 400.
• Re-open returns the question to Open Questions and removes from doneForNowIds.
• Older stale development localStorage states (pre-v0.53 shape) may be discarded during current dev phase. No migration required.

v0.59a — Investigation revision provenance rule

• Investigation has a semantic revision (investigationRevision).
• Meaningful persisted Investigation changes advance it by exactly 1.
• Persistence activity itself (autosave, save) does NOT advance revision.
• Report generation records generatedFromRevision: investigationRevision.
• Report generation advances revision by 0.
• Equal revisions mean Report reflects current Investigation.
• Different revisions mean the Investigation has changed since Report generation.
• Existing Report remains available (not invalidated).
• Report update/regeneration remains manual (user-triggered).
• Restart clears Investigation + Report via clearInvestigation() + setInvestigationRevision(0).
• Report history/comparison remains deferred beyond MVP.

v0.59b — Report freshness UI on the Report page

• Report page shows Current (matching revisions) or Update available (differing revisions).
• Manual Update report action on Report page triggers regeneration via /api/cases/overview.
• Duplicate prevention guard during in-flight update.
• No automatic regeneration.
• Explanation copy: "The investigation has changed since this report was generated."

v0.59c — Report freshness state surfaced on the Portfolio

• Portfolio surfaces Current / Update available alongside existing View report link.
• Derives freshness solely from revision provenance (investigationReport.generatedFromRevision === investigationRevision).
• Zero model calls during Portfolio render.
• No mutation of Investigation.
• No Update report action on the Portfolio — manual Report updating remains owned by the Report page.
• If no Report, neither freshness state is shown.

Semantic transitions that advance revision:

Episode Done (with content)          → +1
Re-open (resolved → unknown)         → +1
Finding proposition correction       → +1
Eligible → not_relevant              → +1
not_relevant → eligible (restore)    → +1
First meaningful change              → =1 (from initial 0)

Transitions that do NOT advance revision:

Autosave                             → 0
Report generation                    → 0
Hydration                            → 0 (reads value)
Empty proposition correction (no-op) → 0
null → null disposition (no-op)      → 0

Zero Open Questions milestone (v0.51)

When all unknowns are resolved and clarified questions exist: "You've now worked through all of the questions we surfaced. Would you like to see an overview of what we understand so far?" with "Review current understanding" button. This occupies the former Open Questions position. The invitation is a milestone, not a readiness/completion judgement.

Current Understanding refresh invariants

Reconstruct CU when canonical meaning or eligible evidence set changes — NOT when investigation/question status changes alone. Re-open ≠ change what we understand; Finding correction / Not Relevant / completed episode = change what we understand.

Persistence and Report ownership

• Canonical persistence owner: lib/storage/providers/local-storage.js (saveInvestigation / loadInvestigation). All routes read from the same snapshot.
• Report generation owner: Report page only (NOT ReasoningWorkspace, NOT Investigation page).
• localStorage key: confidence-engine-investigation (single canonical key — multi-investigation not yet implemented).
• Temporary identity: case-1. True multi-investigation persistence/identity is future work.
• Portfolio client hydration: Portfolio page uses 'use client' — initial pre-hydration empty state ≠ absence of persisted data. Always wait for hydrated semantic controls before classifying state.

MVP boundaries (v0.59b)

Implemented in MVP:

• visible Report freshness state (Current / Update available)
• explanation that the Investigation has changed since Report generation
• manual Update report action (user-triggered)
• no automatic regeneration
• duplicate prevention guard during in-flight update

Deferred beyond MVP:

• Report history
• retaining multiple Reports
• Report comparison
• modelling/preview comparison between Investigation revisions

Current development / verification constraints

• Canonical dev server at http://localhost:3000. Never start/stop/restart/probe it. If unavailable → BLOCKED and stop.
• Playwright MCP: use Run Playwright code with semantic locators for known controls. Snapshot refs ([ref=...]) are observational only.
• For async/hydration states: use waitFor({ state: 'visible', timeout }) — not arbitrary sleeps.
• If a prescribed semantic locator cannot find its control → STOP. No fallback to CSS/XPath/DOM traversal.
• Live freeze: once Playwright verification begins, no production file edits until evidence is captured and classified.
• Tests are instruments, not product truth. At first deterministic failure: classify PRODUCT vs APPARATUS failure, then stop. Do not enter test-harness repair loops.
• Mocked boundary ownership: if a lower-layer function is mocked, test the value crossing the seam — do not require the mock to reproduce real implementation.

Current limitations / genuinely open boundaries

Not yet implemented:

• Multi-investigation portfolio (search/tag/archive/group)
• Durable investigation identities beyond case-1
• Export/copy of Reports to Jira or external document
• Portfolio expansion beyond one canonical investigation
• Report history / comparison

Known boundaries:

• Current Understanding and Investigation Report are architecturally distinct artefacts. Plausible interpretations in the Report remain explicitly interpretive, not evidence.
• The ≤5 processing bound observed during development was an experimental apparatus constraint, NOT a product requirement. Six Open Questions surfacing is legitimate product output, not a formulation defect.
• Evidence discipline: what proves useful in live experimentation must be captured at provenance level, not as test diary entries that become operational constraints.

v0.60 — Multi-Investigation Storage Contract Decision

Completed on branch feature/multi-investigation-storage-v0.60. Bounded architecture/contract decision task following v0.60 audit. Zero production changes. Documentation-only decision checkpoint.

1. Contract ownership

Decision: lib/storage/investigation-storage.js is the application-facing persistence boundary. It defines the contract whose semantics (not just signatures) are immutable across provider implementations. Consumers import only from this module.

lib/storage/providers/local-storage.js remains one concrete provider implementation behind the contract, not an application import target.

Invariant: No consumer imports providers/local-storage directly.

2. Durable Investigation identity

Decision: Canonical Investigation carries a durable id field.

Property	Value
ID allocation owner	Application (not storage contract)
Allocation timing	On create new, before any persistence call
Mutable after creation	NO
Route [id] represents same identity	YES — direct round-trip: app sets → route reads → storage uses

3. Minimum contract operations — resolved in v0.60b

Four semantic operations. createEmptyInvestigation() does not belong to the storage contract (resolved below).

#	Operation	Input	Output	Product use
1	listInvestigations()	—	InvestigationSummary[] (lightweight)	Portfolio renders index/list
2	loadInvestigation(id)	durable ID string	Investigation | null	Investigation/Report page hydration
3	saveInvestigation(investigation)	full snapshot (with id already present)	void	Autosave, Report generation, all writes
4	restartInvestigation(id)	durable Investigation ID	void (reasoning state cleared)	"Restart this investigation"

Explicitly rejected: search, pagination, sorting, sync, merge, transactions, history, report versions, sharing, permissions, accounts. No existing product behaviour requires them.

---

3b. Investigation creation ownership — v0.60b resolution

Contract correction decision. The v0.60a contract contained a contradiction: Section 2 stated "ID allocation owner: Application" while Section 3 listed createEmptyInvestigation() as a storage operation producing { id: string }. These are incompatible models (application-owned identity vs. storage-owned creation). They cannot both be true.

Evidence from current product:

1. Create new behaviour today: Portfolio's "+ Create new investigation" is a <Link href="/investigations/case-1"> — a static navigation link with zero ID allocation and zero storage call. It navigates to the Investigation route which calls loadInvestigation(), receives null (no persisted data), and shows an empty ScenarioForm state. No empty Investigation is persisted on click.

2. First meaningful persistence event: scenario-form.jsx line 681 — after submitScenarioForStartCase returns successfully. The user has selected a scenario, submitted it, and the application receives a structured result. Only then does saveInvestigation() persist with real data (scenario + situationGraph + investigationRevision: 1). This is a domain-driven persistence boundary, not an identity-driven one.

3. No product requirement for empty-persisted Investigations: There is no current product behaviour that creates, displays, or expects empty persisted Investigations in Portfolio. The concept of persisting an abandoned start as a visible Portfolio item was asserted in v0.60a without product evidence.

4. No product requirement for storage to allocate identity: Current code uses hardcoded INVESTIGATION_ID = "case-1" in application code with no storage involvement in ID generation or allocation.

Decision: MODEL A wins — application-owned identity, deferred persistence.

• ID allocation owner: Application layer (not storage contract).
• When allocated: On user click of "+ Create new investigation", before navigation.
• First persistence: When the user produces meaningful Investigation state (scenario submitted) and saveInvestigation(investigation) is called — not on create-new click.
• Empty abandoned Investigation persisted: NO. No durable Investigation exists until meaningful state triggers save.
• Storage allocates identity: NO.

createEmptyInvestigation() does not belong in the storage contract. It never did — it was inconsistent with Section 2's "Application (not storage)" decision. The creation flow lives entirely at the application layer: app allocates durable ID → navigates to /investigations/{id} → user produces state → saveInvestigation(investigation) persists.

Create new investigation flow (corrected):

1. User clicks "+ Create new investigation" on Portfolio.
2. Application allocates a durable immutable ID (library/algorithm deferred).
3. Application navigates to /investigations/{id}.
4. User fills out ScenarioForm → submits → first meaningful saveInvestigation(investigation) persists the Investigation under its already-known ID. The Investigation becomes listable in Portfolio at that point.

Investigation properties (unchanged from v0.60a):

• Carries durable id: YES
• ID immutable: YES
• ID exists before first persistence: YES

Restart decision retained: YES — identity preserved, reasoning/report state cleared.**

4. Portfolio listing returns lightweight metadata

listInvestigations() returns summaries, not full snapshots. Minimum information:

• id — navigation target
• scenario (or title framing) — card text
• updatedAt — freshness display
• investigationRevision — revision metadata
• Report existence and generatedFromRevision — "Current" / "Update available" indicator

Full Investigation loads separately via loadInvestigation(id) on navigation.

5. Restart = reset reasoning state, NOT remove

Restart clears reasoning state (findings, questions, report) within the Investigation but preserves the Investigation's durable identity and container. The Investigation remains visible in Portfolio listing as a named item with cleared state.

A separate "Delete investigation" UI is not required by current MVP. Contract-level removeInvestigation(id) is acceptable for future use but has no product owner yet.

6. Create new investigation flow

User clicks "+ Create new investigation" → application allocates durable immutable ID (library/algorithm deferred) → navigates to /investigations/{id} → Investigation page loads, loadInvestigation(id) returns null (no persisted state yet) → ScenarioForm renders in empty-start mode → user fills scenario and submits → first meaningful state is captured → saveInvestigation(investigation) persists the full snapshot under its already-known ID → Investigation becomes listable in Portfolio. No empty Investigation is persisted at creation time.

7. localStorage ≠ contract

The storage contract accepts/returns domain-level Investigation objects keyed by id. How each provider maps those to its storage mechanism is implementation detail. Contract does not dictate key scheme, row schema, or collection structure.

8. Synchronization separate

Storage contract = persistence only. Not synchronization. No sync-specific fields (syncStatus, remoteId, dirtyFlags, lastSyncedAt) belong in the Investigation shape during this increment. Decisions here do not prevent future sync — the Investigation object carries its own durable ID sufficient for identity resolution.

v0.60c — Identity-Aware Investigation Storage Implementation

First production implementation increment of v0.60. Bounded capability: two independently addressable Investigations can be persisted and loaded through the storage contract.

What was implemented

• loadInvestigation(id) — accepts optional durable ID; selects by that identity when provided; returns null for unknown IDs
• saveInvestigation(investigation, id) — accepts optional durable ID; persists under provider-chosen key derived from id; does NOT allocate or replace the supplied ID
• clearInvestigation(id) — accepts optional durable ID; removes specific investigation by identity when provided

Representation chosen (internal to localStorage provider)

Each Investigation is stored as a separate top-level localStorage key:

confidence-engine-investigation:<durable-id>

e.g. confidence-engine-investigation:inv-abc123

This was the smallest sufficient representation because:

• Direct key lookup provides O(1) per-investigation access without needing an index
• No generic repository layer required — each key is independently addressable by its durable ID
• The two-Investigation proof requires independent storage and retrieval, which this achieves with zero indexes or aggregation

Representation is not exposed to callers — the contract does not reveal localStorage key structure.

Two-Investigation proof results (deterministic test)

Invariant	Result
saveInvestigation(A) persists A under A.id	✅ PASS
saveInvestigation(B) persists B independently under B.id	✅ PASS
saving B does not overwrite A	✅ PASS
loadInvestigation(A.id) returns A	✅ PASS
loadInvestigation(B.id) returns B	✅ PASS
loadInvestigation(unknownId) returns null	✅ PASS
saveInvestigation() does not invent/change supplied ID	✅ PASS

Singleton compatibility seam

• Existing consumers call loadInvestigation() and clearInvestigation() without an id argument — these continue via the legacy singleton path (canonical key + sessionStorage fallback)
• No caller was refactored in this increment
• The compatibility seam is: functions accept optional second parameter; when absent, behaviour matches pre-v0.60c singleton semantics
• case-1 is NOT introduced as canonical identity — it exists only in existing consumer route URLs and legacy data

Legacy migration

• No legacy singleton → new durable ID migration policy was invented
• Existing case-1 localStorage data continues to be served by the backward-compatible path
• Migration of existing consumers to the new identity-aware calls is deferred to a later increment

Production files changed

File	Purpose
lib/storage/providers/local-storage.js	Identity-aware storage contract implementation
tests/storage/investigation-storage.test.js	6 new targeted tests for v0.60c invariants

Next restart point

Migrate existing application callers to the identity-aware contract signatures:

1. [id]/page.jsx — pass route [id] to loadInvestigation(id)
2. [id]/report/page.jsx — pass route [id] to loadInvestigation(id) and saveInvestigation(..., id)
3. scenario-form.jsx — pass investigation ID through save calls
4. app/page.jsx — migrate Portfolio to use listInvestigations() (next increment)

v0.60d — Canonical Save Identity Contract

Problem: investigation-storage.js was a bare re-export (export { ... } from "./providers/local-storage.js"). It did not own semantic contract — whatever signatures the provider exposed were what consumers received. The provider accepted an independent explicit id argument that could silently override or compete with snapshot.id.

Decision: investigation-storage.js now owns the canonical identity contract. It wraps the provider with explicit semantics:

• Canonical save identity comes solely from investigation.id;
• investigation-storage.js owns application-facing semantics;
• localStorage provider remains implementation detail (never allocated ID, never changed it);
• Any remaining singleton compatibility behaviour is explicitly temporary — for unmigrated callers only.

Implementation:

Module	Role
lib/storage/investigation-storage.js	Application-facing boundary — owns identity contract
saveInvestigation(snapshot, explicitId)	Uses snapshot.id as sole save key when present; falls back to singleton path for unidentified legacy snapshots
loadInvestigation(id)	Passes raw id to provider (identity-aware) or singleton path (legacy)
lib/storage/providers/local-storage.js	Concrete localStorage implementation — unchanged, representation remains private

Test: 7 new deterministic tests in tests/storage/investigation-storage.test.js under describe block "v0.60d canonical identity" prove:

1. Identified snapshot persists under its own id key (not CANONICAL_KEY)
2. Load by same id round-trips correctly
3. Provider does not invent or replace the supplied ID
4. Explicit competing id argument is silently ignored when snapshot has an id
5. Two identified investigations (A/B) remain independently addressable
6. Unknown ID returns null
7. Unidentified legacy snapshots still fall back to CANONICAL_KEY

Status: UI/routes are not migrated. All existing callers continue via the singleton compatibility path (they call saveInvestigation({ ... }) with no explicit second parameter, and their snapshots carry no id field). No production caller was modified in this increment.

v0.60e — Route-Owned Investigation Identity

Purpose: Migrate the Investigation page route to own durable investigation identity via its route [id] segment, passing that ID through to ScenarioForm for hydration and persistence without migrating legacy singleton data or changing Portfolio/Report behaviour.

What was implemented

File	Change
app/investigations/[id]/page.jsx	Removed hardcoded INVESTIGATION_ID = "case-1" constant; route [id] param extracted as routeId via params.id; loadInvestigation(routeId) loads by route identity; investigationId={routeId} passed to ScenarioForm in both branch paths; report navigation uses routeId.
components/scenario-form.jsx	Added investigationId prop; session restore calls loadInvestigation(investigationId) when provided; all 4 save call sites include id: investigationId in snapshot (autosave effect, start-case submit, update-case submit, report overview autosave).

Contract crossings verified by deterministic test

• Owning test files: tests/storage/scenario-form-persistence.test.js (11 tests) + tests/storage/investigation-storage.test.js (23 tests, including v0.60d canonical identity section)
• All 34 tests pass on first run; no reruns required

Verified behaviour

• Route ID is passed from [id]/page.jsx into ScenarioForm as investigationId: YES
• Missing identified Investigation (loadInvestigation("v060e-live") → null) starts clean: YES
• Legacy singleton fallback used: NO — no backward-compat load was needed; the route ID was absent from storage
• First meaningful saved snapshot carries route ID: YES (all 4 save sites inject id: investigationId)
• Subsequent save identity preserved: YES (storage layer uses snapshot.id as sole identity authority)
• Storage provider changed: NO — only investigation-storage.js wrapper, already-proved in v0.60d

Portfolio

Not migrated. app/page.jsx retains hardcoded INVESTIGATION_ID = "case-1". Create New navigation to a durable-ID route is later work.

Report

Not migrated. app/investigations/[id]/report/page.jsx untouched. Remains a later caller migration increment.

Restart

Untouched in this increment. All clearInvestigation() calls remain without an id argument (legacy singleton path). If identity-aware restart is needed, the next increment should pass investigationId through those clear calls.

v0.60f — Allocate Investigation Identity on Create New

Purpose: Replace the Portfolio's "+ Create new investigation" static link (/investigations/case-1) with an application-owned durable ID allocation that navigates to /investigations/{id} without persisting any empty Investigation.

What was implemented

File	Change
app/page.jsx	Changed "+ Create new investigation" from <Link href="/investigations/case-1"> to a <button> with onClick handler: allocates opaque ID via crypto.randomUUID(), navigates via useRouter().push(/investigations/${id}). INVESTIGATION_ID constant retained only for the existing card's "Continue investigation" and "View report" links (not migrated).
tests/ui/investigation-overview-ui.test.jsx	Added file-level vi.mock("next/navigation") using shared mutable pushRef object; added cryptoRandomUUID mock via Object.defineProperty(global, "crypto"). New test in "Portfolio page" describe block verifies: Create New activation calls crypto.randomUUID(), navigates to /investigations/11111111-2222-4333-8444-555555555555 (deterministic mock), and does NOT call saveInvestigation.

Contract crossings verified by deterministic test

• Create New allocates fresh ID: YES (crypto.randomUUID() called once)
• Navigation uses generated ID: YES (router.push("/investigations/11111111-..."))
• No saveInvestigation during creation: YES (storage mock was not invoked)
• Control is a <button>, not a <Link>: YES (semantically correct — href cannot be static when ID is allocated at activation time)

Verified behaviour

• Create New navigates to /investigations/{UUID}: YES
• Generated ID does not equal case-1: YES
• Resulting Investigation route starts clean (no persisted data): YES (route calls loadInvestigation(id) → null)
• Clean scenario-start state visible on arrival: YES (ScenarioForm textbox present, empty)
• Empty Investigation persisted on Create New click: NO (only allocation + navigation — no persistence)

Portfolio scope bounded

• Card "Continue investigation" / "View report" links still use INVESTIGATION_ID = "case-1": unchanged (not migrated in this increment)
• No portfolio listing changes, search, sorting, pagination, or delete: not implemented
• Report route untouched: not migrated

Production files changed

File	Purpose
app/page.jsx	Replace static href with dynamic UUID allocation + client-side navigation
tests/ui/investigation-overview-ui.test.jsx	Add useRouter mock (shared mutable reference) + Create New deterministic test

Next restart point

Next increment: migrate Portfolio card links ("Continue investigation", "View report") to use the first existing Investigation's durable ID — or migrate Portfolio to listInvestigations() with the four-operation storage contract. Do not proceed until both are addressed.

v0.60g1 — List Investigation Summaries

Purpose: The previous v0.60g stalled attempt was discarded. This increment cleanly implements only the listInvestigations() storage contract: durable-ID listing returning lightweight summaries, excluding legacy/unrelated state.

What was implemented

File	Change
lib/storage/providers/local-storage.js	Added listInvestigations() — enumerates provider records by prefix match, parses each record, projects lightweight summary (explicitly excludes situationGraph, findings, investigationReport), sorts by updatedAt descending
lib/storage/investigation-storage.js	Imported and re-exported listInvestigations as the application-facing public API
tests/storage/investigation-storage.test.js	8 new deterministic tests covering all listing invariants; fixed MockStorageMap to implement .length and .key(i) for WebStorage API compatibility

Listing contract details

• Operation: listInvestigations() — no arguments, returns InvestigationSummary[]
• Summary fields (explicitly projected): id, scenario, updatedAt, investigationRevision, reportExists, reportGeneratedFromRevision
• Excluded fields: situationGraph, findings, full investigationReport, reasoning history, Open Questions, graph nodes
• Records included: only keys matching INVESTIGATION_PREFIX (durable-ID entries)
• Records excluded: legacy singleton (confidence-engine-investigation), sessionStorage state, unrelated keys, malformed entries
• Ordering: updatedAt descending (most recent first); deterministic fallback by id for equal timestamps
• Malformed handling: skip silently — one malformed entry never blocks valid records

Deterministic test evidence (8 tests, all pass on first run)

Invariant	Result
Two durable-ID Investigations coexist → 2 summaries	✅ PASS
Each summary carries correct durable ID	✅ PASS
Lightweight projection: portfolio fields present, payload fields absent	✅ PASS
Legacy singleton excluded from listing	✅ PASS
Unrelated localStorage key excluded	✅ PASS
Independent update preserves other Investigation	✅ PASS
Deterministic ordering by updatedAt descending	✅ PASS
Malformed durable-ID entry skipped, valid records still listed	✅ PASS

Portfolio scope bounded

• Portfolio has NOT been migrated to consume listInvestigations() — that is v0.60g2
• app/page.jsx unchanged from v0.60f
• No rendering or interactive behaviour changes

Production files changed

File	Purpose
lib/storage/providers/local-storage.js	Provider implements record enumeration + lightweight projection
lib/storage/investigation-storage.js	Application-facing wrapper re-exports listing operation
tests/storage/investigation-storage.test.js	8 new listing contract tests; MockStorageMap WebStorage API fix

Next restart point

v0.60g2: Migrate Portfolio to consume listInvestigations() for collection rendering — replace the hardcoded singleton card with a rendered list of Investigation summaries.

Smallest next increment: implement the four-operation storage contract in lib/storage/investigation-storage.js as a re-export of a provider-backed interface whose signatures accept/return domain Investigation objects keyed by durable ID — without committing to any specific localStorage or database representation. This means defining the exported function signatures and the Investigation shape that flows through them, while deferring key scheme, row schema, and collection structure to a later implementation decision.

v0.60g2 — Render Investigation Portfolio

Purpose: Migrate Portfolio to consume listInvestigations() for collection rendering — replace the hardcoded singleton card with a rendered list of Investigation summaries.

What was implemented

File	Change
app/page.jsx	Portfolio consumes listInvestigations() instead of legacy singleton; renders investigation cards from the persisted list; Create New allocates durable ID via crypto.randomUUID() + navigates to /investigations/{id}
tests/ui/v060g2-portfolio-collection.test.jsx	Dedicated UI test covering: two distinct Investigations persist independently → Portfolio lists both → cards show correct IDs/scenarios/freshness → Create New allocates ID and navigates

Deterministic evidence

• Test file: tests/ui/v060g2-portfolio-collection.test.jsx
• First run result: 24/24 PASS (no reruns)
• Build: PASS

Live verification

Rob manually verified: created a second genuine durable-ID Investigation → returned to Portfolio → two distinct persisted Investigation cards visible → Create New visible → no case-1 card presentation.

Known defect discovered during live verification

A separate pre-existing Report identity defect was discovered during live verification of v0.60g2. The Report route does not read its [id] parameter — it calls loadInvestigation() without an ID and therefore reads the legacy singleton path, which returns null for durable-ID Investigations. This is classified as:

PRIMARY CLASSIFICATION: A — REPORT IDENTITY DEFECT

This defect will be addressed in a separate increment (v0.60h). It is NOT caused by empty-Done semantics.

Next restart point

v0.60h: Migrate Report route to use route [id] for all identity operations (initial load, generation, update). Preserve existing v0.58/v0.59 Report lifecycle and freshness semantics. This is an identity migration only — no Report redesign.

v0.60h — Report Route Identity Migration

Purpose: Fix the Report route's two unscoped loadInvestigation() calls so it reads by route [id] instead of the legacy singleton path. Preserve all existing v0.58/v0.59 Report lifecycle and freshness semantics.

What was implemented

File	Change
app/investigations/[id]/report/page.jsx	Extract routeId from params.id; both loadInvestigation() calls scoped to routeId; "Back to investigation" link uses dynamic /investigations/${routeId}; save spreads snapshot carrying id (identity preserved by existing v0.60d contract)
tests/ui/investigation-overview-ui.test.jsx	Enhanced storage mock to capture loadInvestigation(id) argument and saveInvestigation(snapshot) argument; added listInvestigations: () => [] mock (pre-existing apparatus gap); 5 new identity assertions in "Report route identity — v0.60h" describe block

Deterministic evidence

• Owning test file: tests/ui/investigation-overview-ui.test.jsx
• Identity assertions (new):
  • inv-a identified load + first generation: PASS (load called with "inv-a", fetch POST count = 1, saved snapshot retains id="inv-a", generatedFromRevision = investigationRevision)
  • existing report for inv-b renders without generation: PASS (load called with "inv-b", zero overview calls)
  • manual update reloads by same id and saves back: PASS (load called with "inv-c" after update click, save retains "inv-c", generatedFromRevision updated)
  • existing report hydration retains zero-call behaviour: PASS (no overview call when revisions match)
  • generation failure preserves existing Report: PASS (no partial persist on failure)
• Existing Report lifecycle tests (unchanged): All 10 PASS — no regression from identity migration
• Pre-existing apparatus gaps: 15 failures in Portfolio/restart sections due to listInvestigations mock gap (unrelated to v0.60h; pre-existing from v0.60g2)

Identity mechanism

• Route params prop (params.id) → routeId (string or empty fallback)
• loadInvestigation(routeId) — initial hydration + manual update
• saveInvestigation({ ...snapshot, investigationReport }) — identity preserved by snapshot spread; storage layer uses snapshot.id per v0.60d contract
• "Back to investigation" link → /investigations/${routeId} (dynamic)

Product files changed

File	Scope
app/investigations/[id]/report/page.jsx	Route identity migration only (3 load/save scopes + 1 link update)
tests/ui/investigation-overview-ui.test.jsx	Mock enhancement (load/save argument capture, listInvestigations stub) + 5 identity assertions

NOT changed

• Report generation logic
• Overview API semantics
• Empty-Done semantics
• Findings workflow
• Storage contract (investigation-storage.js, local-storage.js)
• Portfolio/ScenarioForm/SituationGraph

---

v0.60h-a — Clean deterministic verification of Report identity migration

Purpose: Isolate the contaminated v0.60h deterministic test evidence into a dedicated file, proving that Report identity assertions pass cleanly without unrelated Portfolio/restart apparatus failures.

Why this exists: The original v0.60h deterministic run targeted tests/ui/investigation-overview-ui.test.jsx (mixed file) and produced 15/15 unrelated Portfolio/restart failures due to shared mutable mocks — not Report defects. Under experiment discipline, this should have been classified as APPARATUS FAILURE with the v0.60h implementation left intact (which it was).

What was done: Created tests/ui/v060h-report-identity.test.jsx — a dedicated file containing only Report page assertions, isolated from Portfolio/restart test apparatus.

Item	Value
Source mixed test inspected	tests/ui/investigation-overview-ui.test.jsx (Report route identity describe block at line 497)
Dedicated Report test	tests/ui/v060h-report-identity.test.jsx
Report assertions isolated	A (route identity on load), B (existing report no generation), C (first-gen with findings=[]), D (manual update), E (generation failure), F (update failure) — 6 tests total
Portfolio rendered/imported	NO
Restart behaviour included	NO
Storage provider imported	NO
localStorage used	NO (only as implementation detail of mock setup, not in assertions)
Module-cache manipulation used	NO

Isolation apparatus

Category	What is included	What is excluded
Mocks	loadInvestigation(id), saveInvestigation(snapshot), listInvestigations(), clearInvestigation()	None beyond storage contract
Fixture	Minimal identified Investigation with id: "inv-report-a", valid situationGraph, empty findings	No Findings, no re-open workflow
Route id	Fixed: "inv-report-a" (not dynamic UUIDs, not case-1)	No random discovery, no shared mutable refs

Deterministic evidence

Invariant	Result
First run isolated command	npx vitest run tests/ui/v060h-report-identity.test.jsx
6/6 Report assertions PASS on first run	YES
No reruns required	YES
Production files changed	NO
Build run	NO
Playwright / model / server activity	NONE

Assertions proven

Category	Assertion	Result
A	loadInvestigation("inv-report-a") on initial load	PASS
B	Existing report renders without new overview POST	PASS
C	First generation: exactly one POST, snapshot id === "inv-report-a", generatedFromRevision === 4	PASS
D	Manual update reloads same id, saves back with updated revision	PASS
E	Generation failure preserves existing (no partial save)	PASS
F	Update failure preserves existing report	PASS

Classification: PASS

v0.60h production implementation now has uncontaminated deterministic evidence for durable Report identity and the established Report lifecycle.

• v0.60h production commit remains 2af5971 (untouched)
• Original v0.60h deterministic evidence was contaminated by unrelated mixed-file Portfolio/restart failures
• v0.60h-a introduced dedicated Report-only deterministic test
• Exact isolated command npx vitest run tests/ui/v060h-report-identity.test.jsx passed on its first run (6/6)
• Report identity/lifecycle is now cleanly verified independently of Portfolio/restart apparatus

NOT proven by this verification

• All UI tests pass
• Portfolio/restart mixed apparatus is repaired
• The entire Vitest suite passes
• Empty-Done semantics were retested end-to-end

---

v0.60j — Semantic Restart within Investigation Container

Purpose: Implement restartInvestigation(id) that preserves the durable Investigation container (id, scenario) while clearing all reasoning/report state. Migrate Portfolio and ScenarioForm Restart callers from clearInvestigation() to restartInvestigation(investigationId).

What was implemented

File	Change
lib/storage/providers/local-storage.js	Added restartInvestigation(id) — preserves id/scenario/schemaVersion; resets situationGraph→null, selectedQuestion→null, summary→null, focusedContributions→[], findings→[], investigationReport→null, investigationRevision→0, updatedAt→new ISO
lib/storage/investigation-storage.js	Imported and re-exported restartInvestigation; added semantic contract doc (missing id → no-op, no singleton fallback)
app/page.jsx	Portfolio card Restart: replaced clearInvestigation(summary.id) with restartInvestigation(summary.id) — preserves container, clears reasoning
components/scenario-form.jsx	Three callers migrated: ReasoningWorkspace onRestart, ContinueLaterBanner onRestart, "Start new investigation" button — all call restartInvestigation(investigationId) instead of clearInvestigation()

Deterministic evidence

• Storage tests: 44/44 PASS (investigation-storage.test.js)
• UI contract tests: 7/7 PASS (v060j-restart-contract.test.jsx)
• Exact command: npx vitest run tests/storage/investigation-storage.test.js tests/ui/v060j-restart-contract.test.jsx
• First run result: 51/51 PASS, no reruns

Semantic contract of restartInvestigation(id)

Preserved (container-level): id, scenario, schemaVersion Reset: situationGraph→null, selectedQuestion→null, summary→null, focusedContributions→[], findings→[], investigationReport→null, investigationRevision→0, updatedAt→new timestamp

Portfolio caller migrated

• Portfolio Restart confirmation dialog → restartInvestigation(summary.id) ✓
• No longer calls clearInvestigation() ✓
• Card id passed correctly (inv-a, inv-b verified by UI test)

ScenarioForm three callers migrated

Path	Location	Contract	Observability
1 — ReasoningWorkspace onRestart	line ~949	restartInvestigation(investigationId)	Source-inspected (PATH 1 NOT DIRECTLY OBSERVABLE IN BOUNDED APPARATUS)
2 — ContinueLaterBanner onRestart	line ~971	restartInvestigation(investigationId)	Directly exercised by UI test
3 — "Start new investigation" button	line ~979	restartInvestigation(investigationId)	Directly exercised by UI test

UI apparatus

• ScenarioForm direct caller assertions added: YES (2 directly observable via ContinueLaterBanner and StartNew buttons)
• ReasoningWorkspace onRestart confirmed by bounded source inspection
• No localStorage used, no provider imported, no module-cache manipulation
• All mocks at application-facing boundary only

Live verification (accepted)

• same Portfolio card remained
• same durable ID remained
• same scenario remained
• second Investigation unaffected
• old reasoning state absent
• clean Analyse state visible
• no model call caused by Restart

v0.60 is structurally complete

No remaining current production identity migration is known. v0.60 is structurally complete.

The established multi-Investigation architecture:

Portfolio
→ durable-ID Investigation collection

Create New
→ allocate crypto.randomUUID()
→ navigate to /investigations/{id}
→ no empty Investigation persisted

Investigation route
→ route [id] owns Investigation identity

ScenarioForm
→ hydrates by durable ID
→ saves identified snapshots

Report route
→ loads/saves by route [id]

Restart
→ preserves Investigation container/id/scenario
→ clears reasoning/Report state
→ leaves other Investigations untouched

v0.60 result: STRUCTURALLY COMPLETE.

Acceptance evidence

• case-1 production dependencies: 0
• current production unscoped load dependencies: 0
• current production unidentified save dependencies: 0
• current user-facing clear/delete Restart callers: 0
• Portfolio navigation: durable-ID owned
• Report navigation/persistence: durable-ID owned
• legacy singleton compatibility: still present internally, unused by current product, non-blocking deferred cleanup

Investigation culmination semantics

Report is the point-in-time culmination of an Investigation.

The flow:

Open Questions investigated
→ user chooses Done for now
→ question parked
→ graph reconsidered
→ Current Understanding regenerated
→ user chooses what to investigate next
→ eventually zero Open Questions
→ milestone invitation
→ "Review current understanding"
→ /investigations/{id}/report

There is NO active intermediate overview/review surface between "Review current understanding" and Report. The control performs direct route navigation.

Report represents: Situation, What we understand, What remains plausible when applicable. It does NOT represent a decision recommendation, confidence score, readiness judgement, or proof that the user should act.

The user retains ownership of whether the understanding is sufficient, whether to act, whether to return to the Investigation, whether to re-open and investigate further.

Report provenance remains: generatedFromRevision against investigationRevision, so a Report can later become "Update available" without ceasing to be the valid point-in-time Report generated from its earlier Investigation revision.

Zero Open Questions ("You've now worked through all of the questions we surfaced") means: the currently surfaced Open Questions have been worked through. It does NOT mean: the user's decision is complete, sufficient confidence, or ready to act. The Engine facilitates the culmination. The user owns what that culmination means for their decision/action.

Report is the established Investigation culmination. No further product boundary is selected.

v0.61 — Repeated-Run Initial Decomposition Stability

Objective: How stable is the initial semantic decomposition of the same fixed scenario across repeated runs using the same model, promptVersion and production API path?

Status: Experiments 1–3 produced evidence that different visible/semantic decomposition shapes arise for the same fixed scenario. The active next step is a formalised repeated-same-input experiment to quantify variation in materially important uncertainty coverage, question structure and interpretation.

Canonical production seam

The existing route /api/cases/start is directly suitable for curl/Postman/Claude experimentation:

┌───────────┐    POST /api/cases/start     ┌──────────────┐
│ Scenario   │ ───────────────────────────► │ startCase()  │
│ text (req) │   { scenario, promptVersion? } │              │
│             │                              │ analyseScenario │
│             │                              │ buildInitialGraph │
│             │                              │ selectUnknown  │
└────────────┘                              └──────────────┘

No browser state required. No Investigation ID required by the route itself. The route accepts scenario string directly and invokes the full production reasoning path (analyseScenario → buildInitialGraph → determineGraphBackedQuestion).

Direct curl/Postman contract (Rob)

Method: POST URL: http://localhost:3000/api/cases/start Content-Type: application/json

Request body schema:

{
  "scenario": "<your scenario text here>",
  "promptVersion": "v0.2"
}

• scenario (required): string, 1–10000 characters
• promptVersion (optional): "v0.1" or "v0.2" (defaults to "v0.2")

Response shape (success):

{
  "success": true,
  "summary": "<reconstruction summary>",
  "situationGraph": { /* full graph with nodes/edges/reasoningState */ },
  "selectedQuestion": { "id": "...", "question": "...", "reasoningPattern": "..." },
  "diagnostics": { /* decompositionApplied, questionComplexityAssessment, etc. */ },
  "assessment": { "phase": "...", "progress": "..." },
  "modelName": "qwen-claude:latest",
  "responseDurationMs": 3210,
  "validationStatus": "valid",
  "promptVersion": "v0.2"
}

Response shape (failure):

{
  "success": false,
  "error": "<message>",
  "statusCode": 400|500|502,
  "validationErrors": [...],
  "analysisErrors": [...],
  "diagnostics": {...}
}

Claude apparatus

Needed: YES — a thin CJS helper exists for repeated controlled experiments.

Path: scripts/start-case-experiment-helper.cjs Command:

node scripts/start-case-experiment-helper.cjs "<scenario text>"

or

node scripts/start-case-experiment-helper.cjs --file scenario.json

Input: scenario string (positional arg or JSON file with { "scenario": "..." }) Output: structured JSON to stdout (success fields + endToEndElapsedMs) Retries: NO — single call, no retry logic Canonical production logic duplicated: NO — imports startCase from lib/graph/orchestrator.js, exercises identical code path

Deterministic zero-live-call verification

Check	Source	Result
Input reaches startCase seam	tests/app/api/cases-start-route.test.js:15	PASS (mocked analyseScenario verified)
Output passed through correctly	tests/start-case-summary.test.js:81	PASS (exact summary field round-trip)
Execution failure returns 400/5xx	tests/app/api/cases-start-route.test.js:55,79	PASS
Malformed JSON returns 500	tests/app/api/cases-start-route.test.js:100	PASS
No retry occurs	source inspection (single await)	CONFIRMED
Schema validation present	lib/graph/schema.js:202-205	Zod enforced

Live model calls: ZERO Build: NOT required (scripts/test only, no production code changes) Playwright: NOT used (decomposition-only experiments do not require browser instrumentation)

Browser state investigation

Question	Answer
Does the route mutate persistence?	NO — persistence is handled by ScenarioForm caller AFTER receiving result
Does it require an Investigation ID?	NO — route accepts scenario text directly; id comes from UI caller's state
Does it depend on browser localStorage?	NO — pure HTTP JSON exchange
Does it require any browser-only state?	NO

Decision: direct curl/Postman suitable = YES

Why: The /api/cases/start route is a thin layer (19 lines) over startCase() that validates input via Zod, calls the production function, and returns structured results. No browser state, no persistence side effects, no unrelated mutations. Identical behaviour to what ScenarioForm exercises in production.

Playwright posture for v0.61 experiments

• Playwright NOT default for decomposition-only semantic experiments (apparatus reaches production reasoning path via direct import or HTTP)
• Playwright REMAINS required when the experiment concerns visible/browser behaviour, UI state transitions, or localStorage hydration

Repeated-same-input experiment — active design

Input (fixed scenario):

I run a small manufacturing business. Customer complaints have risen by 35% over the last six months, while production volume increased by 40%.

Most complaints mention late delivery or minor product defects, but our complaint categories changed when we introduced a new CRM tagging system three months ago. During the same period we also changed one supplier and introduced a weekend production shift.

I am deciding whether to spend about £120,000 on automated quality inspection now or wait until we understand whether there is actually a quality problem.

I do not yet know the complaint rate per unit, whether defect rates differ by shift or supplier, or whether the new tagging system changed what gets counted as a complaint.

Execution: same /api/cases/start route · same promptVersion = v0.2 · same configured model · multiple independent runs · zero production changes.

What is compared across runs (semantic structure, not wording):

• Open Question count range
• unknown-node count range
• assumption-node count range
• material uncertainty channels present/absent: normalised complaint incidence · CRM measurement/comparability · late-delivery vs product-defect distinction · supplier/shift/source ambiguity · intervention fit
• redundant questions
• speculative subdivisions
• premature interpretations
• action implications / steering / prioritisation

Raw-response policy: responses only need to survive long enough for within-experiment comparison. Temporary files or in-memory capture are acceptable. Immutable per-run fixtures are NOT required.

Run count: to be fixed in the next experiment prompt.

Active v0.61 question:

For the same fixed scenario under the same current production configuration, how much does the initial decomposition vary in materially important uncertainty coverage, question structure and interpretation?

Corresponding next evidence question:

Across repeated independent decompositions of the same scenario, which material uncertainty channels are consistently preserved, intermittently omitted, or replaced by speculative/redundant structure?

What is NOT the current v0.61 objective: whether assumption-kind nodes render under Possible Interpretations · whether intervention-fit is user-selectable in the UI · whether experiments can be projected into persisted Investigations · whether raw API responses should become UI fixtures · downstream question answering · Current Understanding progression after answers · persistence · localStorage · new scenarios · prompt changes · schema changes · model changes

v0.61.1 — Start-Case Experiment Helper Apparatus Verification

Status: PASSED (18/18, first run, zero reruns)

Purpose: Prove scripts/start-case-experiment-helper.cjs works as a standalone Node command with the same configured environment as production, without requiring live model calls or browser state.

What was verified

Check	Source	Result
A — Positional input reaches startCase seam	tests/scripts/start-case-experiment-helper.test.js:76	PASS
B — File input reads JSON fixture	tests/scripts/start-case-experiment-helper.test.js:89	PASS
B — Malformed file fails before startCase	tests/scripts/start-case-experiment-helper.test.js:224	PASS
C — .env.local loads without dotenv dependency	tests/scripts/start-case-experiment-helper.test.js:116	PASS
D — Exit code 0 on success	tests/scripts/start-case-experiment-helper.test.js:131	PASS
D — stdout is valid JSON with required fields	tests/scripts/start-case-experiment-helper.test.js:136,143	PASS
D — endToEndElapsedMs non-negative	tests/scripts/start-case-experiment-helper.test.js:152	PASS
E — Failure produces non-zero exit code	tests/scripts/start-case-experiment-helper.test.js:169	PASS
E — Failure output is valid JSON	tests/scripts/start-case-experiment-helper.test.js:180	PASS
F — startCase called exactly once on success	tests/scripts/start-case-experiment-helper.test.js:194	PASS
F — No retry on failure	tests/scripts/start-case-experiment-helper.test.js:205	PASS
G — Malformed --file fails before startCase	tests/scripts/start-case-experiment-helper.test.js:224	PASS
G — --file without path fails before startCase	tests/scripts/start-case-experiment-helper.test.js:235	PASS
Integrity — no diagnostic leakage to stdout	tests/scripts/start-case-experiment-helper.test.js:248	PASS

Apparatus fixes applied during verification

1. Inline .env.local parser — replaced require("dotenv") (MODULE_NOT_FOUND) with built-in fs + path loader
2. Mock injection seam — START_CASE_EXPERIMENT_HELPER_MOCK=1 enables deterministic test doubles without live calls
3. Structured failure output — plain text console.error → valid JSON on stdout
4. execFile harness fix — capture stdout/stderr regardless of execFile error state

Zero-live-call verification

• Live model calls: ZERO
• Build required: NO (scripts/test only, no production code changes)
• Exact command: npx vitest run tests/scripts/start-case-experiment-helper.test.js

NOT proven

• Actual semantic quality of decomposition output
• Behaviour with real LLM endpoints
• Performance at scale
• All UI integration tests pass

v0.61.2 — Direct-Helper Production-Seam Proof

Status: APPARATUS FAILURE (import seam blocked) — but apparatus itself verified.

Environment loading result

Item	Value
Existing compatible loader available	YES
Package/function	@next/env → loadEnvConfig(projectDir, configPath?, logger, debug)
Previous bespoke parser	Hand-written .env.local KEY=VALUE parser (replaced)
Final mechanism	require("@next/env").loadEnvConfig(__dirname, undefined, { logOutput: "none" }, false)
lib/config.js remains configuration authority	YES — interprets OLLAMA_BASE_URL and OLLAMA_MODEL
New dependency installed	NO — @next/env transitively available through next ^14.2.0

Real production import result

Item	Value
Process	plain Node dynamic import() from standalone .cjs helper
Actual orchestrator imported	NO — module resolution barrier
Actual startCase export resolved	N/A (import fails before export resolution)
startCase invoked during import-only proof	NO — mode structurally stops before any invocation
Provider invoked	NO
Network used	NO

Failure reason: The orchestrator's transitive dependency chain includes lib/graph/apply-proposal.js which imports from @/lib/llm/provider. This @/ path alias is a Next.js compiler convention (configured in jsconfig.json as "@/*": ["./*"]). Plain Node has no resolver for @/ aliases — it attempts to resolve @/lib as a bare package name and fails.

Evidence:

Cannot find module '/Users/.../lib/graph/orchestrator.js' imported from
/Users/.../scripts/start-case-experiment-helper.cjs
(cause: Cannot find package '@/lib' imported from apply-proposal.js)

Two files in the import chain use @/:

• lib/graph/apply-proposal.js → import { getProvider } from "@/lib/llm/provider"
• lib/graph/focused-investigation.js → (same alias pattern)

Import-only structured output (failure path)

{
  "success": false,
  "mode": "import-only",
  "startCaseResolved": false,
  "failureReason": "Module resolution failed: Cannot find module '@/lib/llm/provider' imported from /path/to/lib/graph/apply-proposal.js"
}

Deterministic evidence

Item	Value
Test file	tests/scripts/start-case-experiment-helper.test.js
Exact command	npx vitest run tests/scripts/start-case-experiment-helper.test.js
First-run result	25/25 PASS (first run, zero reruns)
Tests passed	25 (18 existing apparatus + 7 new import-only)
Tests failed	0
Reruns	0

What was proven by this task

1. Standard environment loader established: @next/env loadEnvConfig replaces bespoke parser — no new dependency needed
2. Import-only mode implemented: Helper supports START_CASE_EXPERIMENT_HELPER_IMPORT_ONLY=1 for deterministic seam verification
3. Standalone production import incompatible: Plain Node cannot load lib/graph/orchestrator.js due to Next.js @/ alias chain in transitive dependencies
4. No production files modified: Helper and test changes only — zero impact on reasoning path
5. Zero live calls: Model calls: 0, HTTP /api/cases/start: 0, curl: 0, Playwright: NO

Classification

APPARATUS FAILURE — standalone plain Node cannot import the real production orchestrator due to Next.js @/ module alias convention.

This is documented evidence of a tooling seam gap: the helper architecture (standalone .cjs CLI) is incompatible with the repository's ESM path-alias module system without:

• Adding a bundler/loader (tsx, esbuild, bundler)
• Modifying production imports to use relative paths
• Running through Next.js tooling

The @/ alias is a legitimate architectural choice that should not be changed. The gap means the helper must use mock mode for deterministic verification or run under Next.js-aware tooling.

Existing helper status (pre-v0.61.2)

Check	Result
Positional input reaches startCase seam	PASS (mock mode)
File input reads JSON fixture	PASS
.env.local loads without dotenv dependency	PASS (now via @next/env)
Exit code 0 on success	PASS
stdout is valid JSON with required fields	PASS
endToEndElapsedMs non-negative	PASS
Failure produces non-zero exit code	PASS
Failure output is valid JSON	PASS
startCase called exactly once on success	PASS
No retry on failure	PASS
Malformed --file fails before startCase	PASS
--file without path fails before startCase	PASS
Output structure integrity (stderr/stdout)	PASS

Live execution

• Model calls: 0
• /api/cases/start calls: 0
• curl calls: 0
• Playwright: NO

Files changed

• scripts/start-case-experiment-helper.cjs — environment loader replaced, import-only mode added
• tests/scripts/start-case-experiment-helper.test.js — 7 new import-only assertions (describe block H)

---

v0.61 Experiment 2 — Initial Decomposition Coverage (Scenario Reuse)

Status: B — USEFUL BUT MATERIAL UNCERTAINTY LOST

Execution

Item	Value
Branch	feature/initial-decomposition-v0.61
Starting HEAD	6425540 test(confidence-engine): prove direct helper production seam
Route	POST /api/cases/start
Requests made	1
Successful results	1
Retries	0
Execution errors	NONE
Model	qwen-claude:latest
Response duration	77,984 ms
Validation status	valid
Prompt version	v0.2

Observed Decomposition

Summary: "A manufacturing business owner is evaluating whether to invest £120k in automated quality inspection amid rising complaint counts and recent operational changes, but lacks rate-based and categorical data to determine if a genuine quality or delivery problem exists."

Graph topology: 14 nodes (1 state, 3 observation, 3 metric, 2 relationship, 3 unknown, 2 assumption), 6 edges.

Unknown investigative paths:

1. nvb2z51 — Current and historical complaint rate per unit produced is unknown
2. nn7tkbd — Whether the new CRM system changed the threshold or definition of a valid complaint compared to the old system
3. n9fhzkz — Defect and delivery failure rates specifically attributed to the new supplier or weekend shift are unknown

Other materially relevant graph content:

• nqpcnaa (relationship): Complaint volume rose less than production volume (35% vs 40%), suggesting the complaint rate per unit may have decreased or remained stable
• netiuwi (relationship): Complaint categories changed three months ago due to CRM updates, decoupling historical complaint data from current classification methods
• npf0jdh (assumption): No genuine quality degradation exists
• n1kszew (assumption): A true quality or delivery issue exists, likely triggered by new supplier or weekend shift

Note on selected question: The automatic unknown selection process produced a decomposition-rejection in this run. The selectedQuestion is null; the system generated a reconstruction question but rejected it as not authoritative, deferring to graph-backed pipeline. Three unresolved unknowns remain at initial decomposition — no auto-selection/prioritisation occurred beyond tie-breaking via scoring heuristics.

Semantic Reference Evaluation (A–F)

A — facts and explicit unknowns: CLEARLY PRESERVED Evidence: All six material facts present as distinct nodes (complaints +35%, production +40%, late delivery/minor defects, CRM tagging change, supplier change, weekend shift). £120k decision node present. Three explicit unknowns preserved. No invented facts.

B — normalization uncertainty: CLEARLY PRESERVED Evidence: Node nqpcnaa explicitly captures the 35% vs 40% divergence as a relationship node with status "supported" and confidence "medium". Notes that complaint rate per unit may have decreased or remained stable despite higher absolute numbers. This is distinct from arithmetic resolution — it frames the question without resolving it.

C — measurement/comparability: CLEARLY PRESERVED Evidence: Node netiuwi explicitly captures CRM categories changed decoupling historical data; node nn7tkbd explicitly frames whether the new CRM system changed the threshold or definition of a valid complaint. Both are selectable unknown investigative paths. The distinction between "measurement instrument changed" and "rate normalization needed" is partially compressed into shared graph space, but both concerns are explicitly available as user-investigable uncertainties — not folded away.

D — nature of the problem: PARTIALLY PRESERVED Evidence: "Late delivery" and "minor product defects" are bundled into a single observation node (nqjt07e). The distinction between a delivery/logistics problem and a quality/defect problem is not preserved at sufficient granularity to guide targeted investigation.

E — causal/source ambiguity: CLEARLY PRESERVED Evidence: Supplier change, weekend shift, CRM/tagging, and production volume changes all present in the graph without any edge asserting causality between them and the complaint increase. Node nwg8ma4 bundles these as known metric changes; assumptions (npf0jdh, n1kszew) frame competing hypotheses without resolution.

F — intervention fit: PARTIALLY PRESERVED / MATERIAL GAP Evidence: £120k automation cost node exists (nubgwjr). The proposed intervention is represented as part of the situation, but no graph content explicitly exposes whether that intervention fits the actual problem being investigated (e.g., automated inspection may address some defects but not late delivery, CRM/tagging effects, supplier effects, shift-management effects). This gap between "there might be a quality problem" and "£120k automated inspection is the right response" was not surfaced as an explicit unknown in Experiment 2.

Failure Patterns

Pattern	Result	Evidence
F1 — premature conclusion	NO	Both competing assumptions presented; no resolution asserted
F2 — invented evidence	NO	All facts trace to scenario text
F3 — material qualification lost	YES — late delivery/defect distinction (D) compressed; intervention-fit gap (F) not surfaced as explicit unknown
F4 — steering/prioritisation	NO	No "first/most important/primary/priority" language in graph or summary
F5 — action recommendation	NO	Summary uses neutral framing ("evaluating whether to invest"); no buy/wait directive
F6 — redundant Open Questions	NO	Three unknowns are materially distinct
F7 — premature arithmetic resolution	NO	35% vs 40% divergence framed as open question, not resolved
F8 — speculative proliferation	NO	Two assumptions are direct negations of each other (minimal viable pair)
F9 — unsupported action implication	NO	No explicit recommendation to buy or wait; "waiting" only implied by null selectedQuestion

Classification: B — USEFUL BUT MATERIAL UNCERTAINTY LOST

Reason: The decomposition is useful and restrained — all facts are preserved, no steering, no premature conclusions. Three materially distinct unknowns at initial decomposition. However, dimensions D (nature of the problem) and F (intervention fit) are partially lost through compression: the late-delivery/defect distinction collapses into generic "complaints", and the £120k intervention's potential mismatch with delivery problems is not surfaced as an explicit uncertainty channel. This is a material compression, though less severe than Experiment 1 where CRM comparability and intervention fit were fully compressed.

Comparison with Clean Experiment 1

The two clean runs surfaced somewhat different subsets of the scenario's uncertainty.

Normalized across both runs: Both clean experiments clearly preserved normalization uncertainty. Experiment 1 preserved it in explicit Open Question framing (absoluteness of counts without denominators); Experiment 2 preserved it via graph node nqpcnaa (35% vs 40% rate divergence). Neither run compressed the need for denominator data.

CRM comparability across runs: Experiment 1 preserved CRM change in framing but did not expose it as clearly as a selectable investigative path. Experiment 2 explicitly surfaced CRM threshold/definition comparability as an Open Question. The distinction is one of explicitness of exposure, not presence vs absence.

Repeated signal across both runs — intervention fit: Across both clean runs, intervention fit remained unsurfaced as a distinct investigative path. The proposed £120k automated-inspection action was represented as part of the situation in both, but neither decomposition explicitly exposed whether that intervention fits the actual problem being investigated.

Experiment 2 additionally shows partial compression of late-delivery vs product-defect distinction across runs.

What two observations support: Both clean experiments preserved normalization and CRM measurement concerns (at different levels of explicitness). The repeated signal is intervention fit — a gap in both runs.

What two observations do NOT establish: Systematic instability, model inconsistency, reliable coverage, improvement, regression, or reproducible patterns. Two observations of the same current system behaviour are insufficient to establish either direction.

Git

Documentation commit: docs(confidence-engine): record v0.61 decomposition experiment 2 Working tree: clean after commit

Next Evidence Question

One neutral evidence question only:

When a scenario contains a proposed action whose usefulness depends on what kind of problem actually exists, does the initial decomposition preserve intervention fit as a distinct uncertainty without steering toward or against the action?

---

v0.61 Experiment 3 — Intervention Fit (Scenario Reuse)

Status: B — INTERVENTION FIT ONLY IMPLICIT

Execution

Item	Value
Branch	feature/initial-decomposition-v0.61
Starting HEAD	95d9965 docs(confidence-engine): correct v0.61 experiment interpretation
Route	POST /api/cases/start
Requests made	1
Successful results	1
Retries	0
Execution errors	NONE
Model	qwen-claude:latest
Response duration	90,801 ms
Validation status	valid
Prompt version	v0.2

Observed Decomposition

Summary: "A manufacturing business owner is evaluating whether to invest in automated quality inspection based on a 35% rise in customer complaints, while production volume increased by 40%, recent operational changes altered tracking methods, and critical defect/volume metrics remain unknown."

Graph topology: 20 nodes (1 state, 6 observation, 3 metric, 2 relationship, 3 unknown, 3 assumption, 2 transition), 9 edges.

Unknown investigative paths:

1. nvwsfko — Actual complaint rate normalized by production volume
2. nv8zo6j — Whether defect or delivery issues correlate with the new supplier or weekend shift
3. n3ftrpn — How much of the complaint increase is attributable to changed CRM counting rules rather than actual performance

Relevant graph content for intervention-fit check:

• n93k83a (metric): "Proposed automated quality inspection system costing approximately £120,000"
• nyi88el (assumption): "Complaints are driven largely by late delivery logistics rather than manufacturing defects, meaning a £120k quality investment addresses the wrong bottleneck." — conditional hypothesis explicitly framing the problem-type-vs-intervention-fit uncertainty
• ne49eqk (assumption): Competing hypothesis that investment is premature due to volume/CRM artifact
• nywshmd (assumption): Opposing hypothesis that genuine issues justify inspection investment

Intervention-Fit Check

Check	Result	Evidence
Proposed £120k action represented	YES	Node n93k83a — "Proposed automated quality inspection system costing approximately £120,000"
Late-delivery problem represented	YES	Node noiiqc5 and node nyi88el explicitly reference late delivery logistics
Product-defect problem represented	YES	Node noiiqc5 references minor product defects; node nyi88el references manufacturing defects
Difference between delivery and defect mechanisms preserved	YES	Assumption nyi88el explicitly distinguishes "late delivery logistics" from "manufacturing defects" and frames the £120k investment as potentially addressing the wrong bottleneck
Any explicit uncertainty linking problem type to automation suitability	YES	Assumption nyi88el: "Complaints are driven largely by late delivery logistics rather than manufacturing defects, meaning a £120k quality investment addresses the wrong bottleneck."
Any question equivalent to "would this action address the actual problem?"	PARTIAL	The distinction is present as a conditional assumption, not a direct investigable question. A user would need to read and compare assumptions nyi88el vs nywshmd to surface the intervention-fit gap themselves.
Any conditional hypothesis about intervention fit	YES	Assumption nyi88el explicitly frames: IF complaints are delivery-driven THEN inspection addresses wrong bottleneck
Recommendation for/against investment	NO	All three competing assumptions carry status "provisional" and confidence "medium/high" — none resolved. The assumption is a conditional hypothesis, not a recommendation.
Priority language	NO	No "first/most important/primary/priority" language in graph or summary

Guardrail Evaluation

Guardrail	Result
G1 invented evidence	NO
G2 unsupported causal conclusion	NO
G3 premature arithmetic conclusion	NO
G4 action recommendation	NO
G5 steering/prioritisation	NO
G6 speculative proliferation	NO (only 3 assumptions — minimal viable set: investment-justified, investment-premature, wrong-intervention)

Classification: B — INTERVENTION FIT ONLY IMPLICIT

Why: The £120k action and both problem types (late delivery / product defects) are explicitly represented. Crucially, assumption nyi88el is a distinct graph node that frames the conditionally uncertain relationship between problem mechanism and intervention suitability: "IF complaints are largely delivery-driven THEN a quality inspection investment addresses the wrong bottleneck." This preserves the semantic distinction as conditional uncertainty rather than resolving it.

However, it appears as an assumption (a hypothesised state of affairs) rather than an explicit unknown or question inviting investigation into whether the proposed action fits the actual problem. A user must infer that this assumption represents a meaningful intervention-fit gap rather than seeing it surfaced directly as a selectable investigative path. The distinction is present but implicit — recoverable through careful reading but not structurally highlighted as "here is what we do not yet know about whether this action fits."

Comparison with Experiments 1 and 2

	Experiment 1	Experiment 2	Experiment 3
Intervention fit status	Lost	Partially preserved (implicit)	Partially preserved (conditional assumption)
£120k action represented	Yes	Yes	Yes (node n93k83a)
Problem-type distinction surfaced	No	Partially (D — compressed)	Yes (assumption nyi88el)
Explicit conditional fit hypothesis	No	No	Yes (assumption nyi88el)
Three competing assumptions presented	—	Competing assumptions on quality existence	Competing assumptions: investment justified / premature / wrong bottleneck

Experiment 3 exposes the distinction more explicitly than Experiments 1 and 2 — assumption nyi88el directly ties the mechanism (delivery vs defects) to the proposed action's potential mismatch. This is an improvement over previous runs where intervention fit was either lost or merely implicit in context.

What three observations support: (1) The late-delivery-vs-defects distinction maps onto intervention fit in this scenario and is now explicitly framed as conditional uncertainty; (2) All three competing assumptions are presented without resolution, preserving open inquiry; (3) The £120k action is not recommended or rejected — it sits alongside the fit hypothesis.

What three observations do NOT establish: (1) That intervention fit will be consistently preserved across different scenarios; (2) That conditional assumptions in the assumption node are as accessible to users as explicit unknowns; (3) That this improvement over Experiments 1/2 generalizes to other scenario types or model versions.

Next Evidence Question

Does surfacing intervention-fit uncertainty as a conditional assumption node provide equivalent user accessibility to surfacing it as an explicit unknown — and if not, what structural change would preserve the distinction without steering?

---

v0.61 Experiment 4 — Intervention-Fit Accessibility in User-Facing UI

Status: BLOCKED — exact Experiment 3 persisted fixture unavailable

PAUSED / NOT CURRENT v0.61 OBJECTIVE

The intervention-fit accessibility / UI-projection line of inquiry from this experiment is NOT the active next boundary for v0.61. The active direction is repeated-same-input initial-decomposition stability (see the "active design" section at the top of §v0.61).

Valid source-level findings recorded in this experiment remain as historical context:

• unknown nodes feed Open Questions
• assumption nodes can feed Possible Interpretations
• assumption-kind meaning and user-selectable unknown-kind meaning are structurally different surfaces

Any valid findings from this thread are preserved but the projection/accessibility line of inquiry is marked PAUSED / NOT CURRENT v0.61 OBJECTIVE.

Execution

Item	Value
Branch	feature/initial-decomposition-v0.61
Starting HEAD	42da768 docs(confidence-engine): record v0.61 intervention-fit experiment
Source trace completed: YES — bounded to reasoning-workspace.jsx lines 1850–2216
Browser navigation: Two older persisted Investigations examined (IDs from Portfolio list) after required Experiment 3 fixture could not be identified
Model/API calls: ZERO
State mutations: ZERO
Execution errors: NONE

Protocol breach — why Experiment 3 fixture was unavailable

Experiment 3 invoked POST /api/cases/start directly. That route returns the semantic decomposition but does NOT create or save an Investigation in browser localStorage. Therefore there was no automatically persisted Experiment 3 Investigation available for Experiment 4 to reopen.

The Experiment 4 protocol required: "Find the Experiment 3 Investigation ID from existing Experiment 3 evidence only. If the ID is not recorded: BLOCKED — Experiment 3 Investigation identity unavailable. STOP."

Instead, when no definitive Experiment 3 ID was available, two older persisted Investigations with identical scenario text were inspected:

• 53862863-...
• d1b44afb-...

Neither was proven to contain Experiment 3's actual graph. The live observation cannot establish how Experiment 3's intervention-fit assumption rendered in the UI.

Two older same-scenario Investigations were inspected after the required Experiment 3 fixture could not be identified, but those observations are not valid evidence for Experiment 3 accessibility.

SOURCE-PROVED EVIDENCE (from source trace)

These findings derive from bounded source inspection only. They do NOT depend on any live Investigation fixture.

Open Questions: Derive from (graph?.nodes || []).filter((n) => n.kind === "unknown" && !resolvedIds.has(n.id)) — graph nodes with kind === "unknown", filtered for resolved status. Rendered as selectable <button> controls that can begin focused investigation via startFocused(n.id).

Assumptions / Possible Interpretations: Provisional assumption nodes (kind === "assumption") satisfy the rendering filter and are rendered under "Possible Interpretations" when they meet the criteria. They are informational, non-interactive <div> content (status: "(informational, not investigable)"). They are NOT Open Question buttons and do not directly call the focused-investigation path.

Assumption → selectable Open Question: No code path converts an assumption node into a selectable investigative path without another semantic/model update. The filter separates assumptions (kind === "assumption") from unknowns (kind === "unknown") with different rendering roles and no cross-conversion mechanism.

Structural conclusion (SOURCE-PROVED): Assumption-kind meaning and user-selectable unknown-kind meaning are structurally different surfaces in the current product. No normal product path equivalent to "assumption → selectable Open Question" exists without later reasoning/state change.

Assumptions rendered: YES — in two locations in reasoning-workspace.jsx:

1. Lines 2029–2048: Possible Interpretations section within initial-proposed-findings div, conditional on possibleInterpretations.length > 0
2. Lines 2187–2216: Persistent "Possible Interpretations" in workspace grid (row 4, full-width), conditional on postAnalyseStatus !== "success" AND interpretationNodes.length > 0

Filter: n.kind === "assumption" && n.status !== "resolved" && !resolvedIds.has(n.id)

User-facing location: Both render under <h2>Possible Interpretations</h2> heading. Cards styled with blue-100 border, blue-50/40 background, labeled "Plausible interpretation" in blue text.

LIVE FIXTURE — invalid substituted observations

The two inspected Investigations are older same-scenario cases, NOT the Experiment 3 result. Their contents cannot establish:

• Whether Experiment 3's nyi88el assumption was visible in any rendered surface
• Whether it would appear under Possible Interpretations if Experiment 3's exact graph were loaded
• Whether it would be selectable as an Open Question
• Whether the intervention-fit meaning would be visible in Current Understanding
• What accessibility classification (B, C, or D) applies to the actual Experiment 3 result

Observation of Investigation 1 (53862863-...):

• Current Understanding: "A manufacturing business owner is weighing a £120k automated quality inspection investment against the need to analyze complaint trends amid concurrent changes in production scale, supplier, shift scheduling, and tracking methodology."
• Open Questions: 2 (complaint rate per unit; defect rates by supplier/shift)

Observation of Investigation 2 (d1b44afb-...):

• Current Understanding: "A manufacturing business observes a 35% rise in complaints alongside a 40% production increase, coinciding with CRM, supplier, and scheduling changes, prompting a decision on a £120k quality inspection investment."
• Open Questions: 7 (CRM tagging threshold; defect rates by supplier/shift; complaint rate per unit overall; complaint rate by customer segments; complaint rate by production shift; complaint rate by process stages; complaint rate by critical defect categories)

These are factual descriptions of what was observed in those Investigations. They do not prove anything about Experiment 3's specific graph or its UI rendering.

WHAT SOURCE EVIDENCE DOES NOT PROVE

Source inspection alone did NOT prove:

• Whether Experiment 3's specific nyi88el assumption would satisfy the rendering filters;
• Whether nyi88el would appear in Possible Interpretations;
• Whether the intervention-fit meaning would be visible in Current Understanding;
• Whether the user would see the exact intervention-fit distinction anywhere;
• Whether the correct accessibility classification would be B, C, or D.

Do not infer runtime data from source structure.

IMPORTANT UNCERTAINTY TO PRESERVE

Experiment 3 established: intervention-fit meaning existed in its returned graph as assumption nyi88el.

The source trace establishes: qualifying assumptions may render as Possible Interpretations.

Therefore it is entirely possible that the exact Experiment 3 result, if presented through the UI unchanged, would be:

• visible but non-investigable (rendered as a Possible Interpretation informational card)

rather than:

• not visible at all.

Experiment 4 did not resolve this. The uncertainty is recorded; it is not chosen between.

THREE-LAYER STATUS

Question	Answer
Represented in Experiment 3 canonical API result	YES
Would qualifying assumption-kind meaning be potentially renderable (by source)	YES
Would it be directly selectable/investigable as an Open Question (by current source structure)	NO
Was Experiment 3's specific assumption actually observed in rendered UI	NO
Was its actual visibility established	NO

PRODUCT-PRINCIPLE CHECK

• Engine selected what to investigate: NO (no steering in decomposition output)
• Engine prioritised a question: NO (zero prioritisation language observed)
• User retained choice among surfaced Open Questions: YES — for whatever Open Questions the specific Investigation happens to contain, they are selectable buttons. However, the full set of uncertainty from Experiment 3 was not demonstrated as present in any examined Investigation.

NEXT EVIDENCE QUESTION (HISTORICAL — NOT CURRENT)

Given an exact captured /api/cases/start result containing an intervention-fit assumption, what is the smallest valid way to observe how that exact canonical graph projects into the existing Investigation UI without making another semantic/model call?

This question remains recorded as historical apparatus evidence. It is NOT selected as the active next direction for v0.61. The active next direction is the repeated-same-input decomposition stability experiment described at the top of §v0.61.

GIT

Documentation commit: pending Working tree: pending

CHANGES

• Production changed: NO
• Tests changed: NO
• Prompt/schema changed: NO
• Canonical Investigation state changed: NO

---

v0.61 Experiment 5 — Repeated Same-Input Initial Decomposition Stability

Status: INCOMPLETE — experiment protocol required five successful runs; obtained four. One run terminated by validation failure.

Execution

Item	Value
Branch	feature/initial-decomposition-v0.61
Starting HEAD	65c5ded — docs(confidence-engine): restore v0.61 decomposition objective
Route	POST /api/cases/start
Requests made	5
Successful results	4
Failed requests	1 (Run 5 — validation failure)
Retries	0
Model	qwen-claude:latest
Prompt version	v0.2

Response Duration Range

• Min: 86,826 ms (Run 1)
• Max: 101,915 ms (Run 3)
• Successful range: 86,826–101,915 ms

Structure Range

	Nodes	Unknowns	Assumptions
Run 1	16	3	3
Run 2	20	7	2
Run 3	21	3	3
Run 4	14	3	2

Unknown count: 3–7 Assumption count: 2–3 Total node count: 14–21

Semantic Channel Evaluation Table

Run	Unknowns	Assumptions	A Normalisation	B CRM comparability	C Delivery vs defects	D Supplier/shift/source	E Intervention fit	Redundancy/speculation	Steering/action implication
1	3	3	PRESENT	PRESENT	PRESENT	PRESENT	ABSENT	None detected	None detected
2	7	2	PRESENT	PRESENT	PARTIAL	PRESENT	ABSENT	Speculative subdivisions (customer segment, QC team capability, responsibility distribution, batch-size/equipment) not from source text; multiple rate-normalisation unknowns redundant	None detected
3	3	3	PRESENT	PRESENT	PARTIAL	PRESENT	PRESENT	None detected	None detected
4	3	2	PRESENT	PRESENT	PARTIAL	PRESENT	ABSENT	None detected	None detected
5	—	—	FAIL	FAIL	FAIL	FAIL	FAIL	FAIL	FAIL

Run 5 Failure Details

• Error: Scenario analysis failed
• Analysis errors: reconstruction: Required, Required, Required
• Classification: UNCLASSIFIED — EVIDENCE INSUFFICIENT. The preserved record contains only a high-level error message and schema constraint names. It does not preserve the raw /api/cases/start response (or the model-produced content that triggered validation) needed to distinguish whether the production API violated its contract (PRODUCT FAILURE) or the observation mechanism failed to process an otherwise legitimate result (APPARATUS FAILURE). The exact failure boundary for Run 5 was not preserved sufficiently to classify retrospectively.
• This run is excluded from semantic frequency counts.

A–E Frequency Counts (Runs 1–4 only)

A Normalisation: 4/4 present, 0/4 partial, 0/4 absent
B CRM comparability: 4/4 present, 0/4 partial, 0/4 absent
C Delivery vs defects: 1/4 present, 3/4 partial, 0/4 absent
D Supplier/shift/source: 4/4 present, 0/4 partial, 0/4 absent
E Intervention fit: 1/4 present, 0/4 partial, 3/4 absent

Material Structural Variation

Material structural variation observed between runs. Run 2 produced 7 unknowns (vs 3 in all other runs) with multiple speculative subdivisions not grounded in the source text (customer segment variation, quality control team capability, responsibility distribution, batch-size/equipment utilization). Node count ranged from 14 to 21. Different selected questions across runs: Run 1 had no selected question; Runs 2–4 each selected a different unknown as the focused investigation target.

Redundancy / Speculation Findings

Run 2 introduced speculative subdivisions: unknowns about "customer segment variation," "quality control team capability shifts," "responsibility distribution impact," and "batch-size/equipment utilization correlation" — none of which are mentioned or implied in the source scenario text. Run 2 also produced redundant rate-normalisation unknowns (two separate unknown nodes covering essentially the same concept). Runs 1, 3, and 4 showed no redundancy or speculation.

Steering / Action Implication Findings

No run contained steering language ("first/most important/primary/priority"), unsupported causal claims between operational changes and complaint increase, premature arithmetic conclusions (the 35% vs 40% divergence was consistently framed as open rather than resolved), or action recommendations. The £120k intervention was represented neutrally in all runs without directional pressure.

Interpretation Answers

1. Which semantic channels were preserved in all four successful runs? A (Normalisation), B (CRM comparability), D (Supplier/shift/source) — consistently present across all four successful runs.

2. Which channels varied between PRESENT / PARTIAL / ABSENT? C (Late-delivery vs product-defect distinction): 1 PRESENT, 3 PARTIAL. The distinction was fully preserved only in Run 1; compressed in Runs 2–4 into generic treatment without explicitly separating delivery failures from manufacturing defects. E (Intervention fit): 1 PRESENT, 3 ABSENT. Only Run 3 surfaced intervention-fit uncertainty explicitly through an assumption node linking late-delivery logistics to the potential irrelevance of quality inspection.

3. Did question/unknown structure materially vary across runs? YES. Node count ranged from 14 to 21. Unknown count varied from 3 to 7. Run 2's graph contained four speculative unknowns not grounded in the scenario text, while Runs 1, 3, and 4 remained restrained at 3 unknowns each. Three different selected questions were produced across Runs 2–4.

4. Did any run replace useful uncertainty with redundant/speculative structure? Yes — Run 2 introduced four speculative unknowns (customer segment, QC team capability, responsibility distribution, batch-size/equipment) not present in the source text, plus redundant rate-normalisation coverage. Runs 1, 3, and 4 did not introduce speculation.

5. Did any run introduce steering, unsupported causality, premature conclusions, or an action recommendation? No. None of the four successful runs introduced steering language, unsupported causal claims between operational changes and outcomes, premature arithmetic resolution, or action recommendations.

6. What does this four-run experiment support? Four semantic channels (normalisation uncertainty, CRM measurement comparability, supplier/shift/source ambiguity, and basic complaint/incidence framing) are represented to some degree across the observed runs — with A/B/D fully present in all four, and C at partial representation in three of four. No invented evidence was observed in any run. The initial decomposition preserved factual capture under the observed conditions.

7. What does this experiment explicitly NOT establish? This experiment does NOT establish: reliability at higher run counts, systematic stability versus transient variation, generalization to other scenario types, the model's behavior with different prompt versions, or that Run 2's speculative divergence is a reproducible failure mode rather than an isolated artifact. One failed run (Run 5) prevents counting it as a successful observation.

Conclusion

The initial decomposition represents core factual uncertainty channels (normalisation, CRM comparability, supplier/shift ambiguity) to varying degrees across fresh runs — with A/B/D fully present and C partially represented in all four observed runs. The primary instability observed is in two areas: (1) intervention-fit mapping — when and whether the gap between problem mechanism and proposed action surfaces; and (2) speculative subdivision risk — the model can introduce unsupported unknown categories on some runs (observed in 1 of 4 successful runs). Node count variance (14–21) confirms that graph topology, not just node content, varies materially.

---

v0.4 — Strengthened Initial Reconstruction Contract (Implementation)

Status: COMPLETE — deterministic tests pass, one live smoke accepted

Starting checkpoint

• Branch: feature/initial-decomposition-v0.61
• HEAD: 37245a8 — fix(confidence-engine): expose reconstruction failure evidence

What was implemented

File	Change
prompts/reconstruct-v0.4.md	New prompt file — provenance-preserving decomposition contract with explicit stop boundary, supplied-relationship preservation, interpretation separation, normalisation guidance
lib/reconstruction/prompt.js	Added v0.4 to PROMPT_VERSIONS; set DEFAULT_PROMPT_VERSION = "v0.4"; added buildV4Prompt loader; added v0.4 case to switch/default
tests/v03-reasoning.test.js	Updated existing default-version assertions to accept v0.4; added 15 new targeted v0.4 prompt contract tests (loading, output-format preservation, provenance stop contract, relationship preservation, interpretation separation, normalisation, next-question discipline)

Architecture present

Principle	Present
Provenance-preserving decomposition	✅ — rule 1: each child's distinguishing semantic content must be directly supported by meaning supplied; plausible/likely/domain-typical not sufficient
Relationship preservation	✅ — "Preserve supplied relationships" section (rules 36-37)
Explicit stop boundary	✅ — "Explicit stop boundary for decomposition" section
Interpretation separation	✅ — rule 6: interpretation never smuggled into decomposition; provisional status required

Deterministic verification

Item	Value
Command	npx vitest run tests/v03-reasoning.test.js
Result	89/89 passed on first run, zero reruns

Live smoke

Item	Value
Route	POST /api/cases/start (no promptVersion — uses default v0.4)
HTTP status	200
Validation status	valid
Prompt version in diagnostics	v0.4
Supplied meaning preserved	✅ complaints +35%, production +40%, complaint-rate-per-unit uncertainty, late delivery, minor defects, supplier change, weekend shift, CRM/counting uncertainty, £120k inspection decision
Provenance boundary enforced	✅ no decomposition into unsupported CRM mechanisms (segments, training, classification, logging delays)
Relationship boundary preserved	✅ unresolved dependency between automated-inspection appropriateness and complaint-kind preserved as unknown investigative path
Interpretation separation	✅ model-generated interpretations remain provisional in plausibleInterpretations, not smuggled into decomposition structure

Acceptance: PASS

v0.4 initial reconstruction is now production default. This does NOT yet establish repeatability or cross-scenario consistency.

---

v0.61 active restart point

The current handoff direction is the repeated-same-input decomposition stability experiment described in the active design section at the top of §v0.61.

Next task: execute a formalised repeated-run experiment using the fixed manufacturing scenario via POST /api/cases/start, comparing semantic structure across runs.

Previous active direction (drifted, now corrected): intervention-fit accessibility in UI projection · assumption-kind node rendering · persistence/localStorage for experiment fixtures · downstream question-answering. These are NOT current v0.61 objectives.

---

Consult docs/design-evolution/README.md for progressive loading of product reasoning and provenance chronology; load the relevant chapter only when a specific historical question requires it.

The current handoff captures all operational facts needed to resume today. For historical decisions, experiment evidence, or methodology evolution — consult the design evolution archive index or task-context packs as appropriate.

Provenance pointers

Need	Read
Product evolution v0.51–v0.58	docs/design-evolution/README.md (progressive loading)
Methodology / Return-to-Origin axioms	docs/current-working-principles.md §0 (A1–A12)
Architecture guardrails	.claude/architecture-guardrails.md
Task routing by work type	docs/task-context-packs.md
Broader architectural intent	docs/architectural-principles.md
Experiment history (specific)	docs/design-evolution/README.md → relevant chapter