- Add reconstruct-v0.5-d-deconstruct-no21.md prompt (removes rule 21)
from original deconstruct-first prompt: 'Do not create artificial
uncertainties to increase their number.'
- Extend start-case-experiment-helper.cjs with --condition D and
--treatment no21 CLI flags for treatment selector apparatus
- Add programmatic treatment validation (rejects invalid treatments
cross-condition, validates only 'no21' is a valid treatment value)
- Condition D output uses freeformText field instead of reconstruction
JSON — avoids production JSON schema enforcement on freeform text
- Mock double supports freeform mode: returns deterministic raw text
instead of structured importantUnknowns when isFreeform=true
- Add 23 targeted tests: J (Condition D prompt selection),
J.1 (treatment validation), J.2 (prompt isolation verification)
- Fix existing B.1 assertions to use matchObject for execution field
Iteration 5 of bounded iteration loop. Two live calls per condition
using fixed HS-03C input (new-service-launch scenario). Model:
qwen-claude:latest via Ollama.
Condition A (question-required): 2 importantUnknowns with broad
descriptions + narrowing questions, 8 nodes, valid validation.
Condition B (question-free): 1 importantUnknown with structurally
narrowed combined-timeline description, 7 nodes, partial validation.
Neither condition preserved the original four-domain structure of
HS-03C. Narrowing occurred through different mechanisms in each
condition — hypothesis unconfirmed for this scenario shape.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Adds experimental apparatus for testing whether removing the question field
from importantUnknowns in reconstruction output improves semantic fidelity.
New files:
- lib/reconstruction/schema-variant.js (conditionB schema + production passthrough)
- prompts/reconstruct-v0.5-b-noreq.md (alt prompt, identical line count to prod)
Changed files:
- lib/graph/orchestrator.js (pass generationSchema through dependencies)
- lib/analysis.js (accept opts.generationSchema fallback to original schema)
- scripts/start-case-experiment-helper.cjs (--condition B/A flag support)
- tests/scripts/start-case-experiment-helper.test.js (8 new deterministic tests)
43 targeted deterministic tests PASS. Production callers use original prompt/schema unchanged.
No live model calls. No full suite evidence available.
Co-Authored-By: Claude Code <noreply@anthropic.com>
HS-03C experiment with explicitly raised cross-area dependencies showed the model preserved relationships between recruitment, software integration and customer commitments while retaining the broader feasibility uncertainty. Result: PARTIAL semantic fidelity. Evidence at .evidence-temp/headless-semantic/HS-03C-focused-E9.json.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Extended deterministic test to write serialized evidence record to
OS temp directory, read back from disk, parse JSON, and assert deep
structural equality. Verified: five structured arrays preserved with
correct lengths; exact inputs unchanged; identity fields intact; null
metadata fields survive round-trip. Temp file cleaned in finally block.
Zero live model calls. Zero browser calls. No production code or
canonical-helper modifications. Experiment 4 original structured
output remains unavailable.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Signed-off-by: robbond <rob@rdbsolutions.co.uk>
Apparatus-level pure serialization function that preserves all five
focused-deconstruction arrays (observations, uncertainties, assumptions,
relationships, possibleFollowUpQuestions) plus exact inputs and execution
metadata from the existing checkpoint — without modifying production code.
Deterministic verification: syntax check pass, 79/79 assertions pass
(including 28 new evidence-ser tests), zero live calls.
Experiment 4's missing structured output cannot be reconstructed (never
persisted by the apparatus). No new semantic evidence produced.
Controlled synthetic experiment (Iteration 4): fixed recruitment scenario with
synthetic answer showed partial preservation of the broader feasibility uncertainty.
The model correctly preserved timing estimates as observations but narrowed the
investigation to hiring pipeline logistics rather than cross-area launch feasibility.
Add focusedInvestigationInput {targetDescription, question, answer} to
composeDeepInvestigationCheckpoint() return object. These are the actual
invocation values: targetDescription from situationGraph node, question
is previousQuestion (NOT node.openQuestion), answer is function param.
Deterministic verification: syntax check passed; source-inspection test
assertions E10/E11 added and pass (0 failures). Call-count ownership:
up to 3 live generateReconstruction calls (focusedDeconstruct +
reconsiderCompletedEpisode + synthesizeCurrentUnderstanding).
No provider, graph, or production behavior changed.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Changed the facilitator-view-adapter to prefer node.description over
node.openQuestion for unresolved unknown items. This is a reversible
presentation change only — graph identity, reasoning-state ownership,
openQuestion generation, and resolved-unknown presentation are unchanged.
Owning test assertions updated to reflect new display contract:
- open-question-presentation.test.js: assertion target shifts from
question text to description text
- facilitator-view-adapter.test.js: deduplication tests use distinct
descriptions so nodes remain distinguishable after the change
Documentation: hypothesis experiment record and handoff return note.
Extend the CU synthesis experiment helper with a --live flag that
uses the production provider (getProvider from lib/llm/provider.js)
for live model invocation. Preserves deterministic mode unchanged.
Adds 10 new deterministic tests verifying live-mode wiring via
injected mock provider: fixture pass-through, single invocation,
schema correctness, immutability, persistence safety, and regression
of existing deterministic behaviour. Zero network calls in all tests.
Co-Authored-By: Claude Code <noreply@anthropic.com>
Remove the 140-line Iteration 16 test section (commit 5d4e7b5) that
represented manually supplied checkpoint values as evidence of
target-scoped contribution behaviour. The section used extractLiveExperimentCheckpoint
to construct mock result objects simulating multi-episode state transitions
but could not observe contributions — the deterministic seam has no
contributions field in its return shape.
Preserves: all pre-existing checkpoint tests, Iteration 15 source-level
assertions, synthetic CU synthesis tests, and existing helper imports.
Add Iteration 16 test extension that documents the behavioral evidence
boundary for target-scoped contribution filtering:
- Observable via extractLiveExperimentCheckpoint:
* Node A pre-episode graph state (graphCheckpoint.nodeCount)
* B's fresh deconstruction trigger (providerGenerateCalls increments)
* Graph/findings continuity (reasoningState turnCount advanced)
- Blocked from deterministic seam (no contributions field in checkpoint):
* Historical contribution retention for non-current targets
* B contribution associated with targetNodeId
* Both A and B surviving in postEpisodeContributions
* B prep scope receiving B-specific contribution
Uses mock result objects to simulate the two-target sequence
(n_enterprise_customer_signing -> n_product_launch_decision) and
explicitly documents which assertions can and cannot be verified.
Historical contributions for a different Open Question node no longer
suppress fresh focused deconstruction of the current target. Only
contributions whose targetNodeId matches the current target are counted
towards suppressing redundant deconstruction.
Adds deriveInvestigationLearning to episode-preparation.js: pure function that maps resolvedNodeIds + Contributions + Findings to investigation learning propositions, preserving node.label question text and contribution-provenance ordering.
Owning tests prove: correct provenance, cross-question isolation, eligibility rule (null/agree eligible; not_quite/not_relevant excluded), unresolved exclusion, no-eligible-material omission, text preservation, and provenance authority over originatingTargetNodeId shortcut.
- Zod validator: plausibleInterpretations accepts empty string (zero interpretations) while remaining required and string-typed
- Prompt instruction: SECTION B now explicitly requests "" instead of absence prose when no plausible inputs exist
- Report generation/update paths derive hasPlausibleInterpretations from actual synthesized data rather than hard-coding true
- Tests: updated validation contract, added hasPlausibleInterpretations derivation assertions, updated seam tests to use derived values
Controlled A/B (zero-Findings scenario, Qwen): unresolved uncertainty exposure
did not contaminate supported understanding or plausible interpretation.
No production changes. No report semantic changes.
robbond
