LLM Exhibits

Exhibit inventory

Exhibit	Summary available	Original full text available
E-A, E-B, E-C, E-D	Yes	No — pending collection
E-E, E-F, E-G	Yes	Yes — reproduced below
G-A, G-B, G-C, G-D, G-E	Yes	No — pending collection

Source limit: The available Markdown contains a one-page audit and interleaved summaries for E-A through E-G and G-A through G-E, but only three verbatim exhibit texts. The evaluation script, full architectural graph, and underlying interview transcripts are not present in the available Markdown.

One-Page Exhibit Audit

Changes and outcomes across E-A through E-G.

Review mapping: G-A reviews E-A; G-B reviews E-B; G-C reviews E-C; G-D reviews E-E using E-C/E-D as context; G-E reviews E-F and E-G. E-D incorporates selected G-C feedback.

Exhibit Name	Key Criticisms Identified	Adjustment Made	Final Outcome
E-A.md	G-A: Forgetting is not demand; counts overstate priority; distress and dissatisfaction vary.	Clarified motivation and provisional rank; narrowed distress; qualified P5; acknowledged enjoyment.	Five pain points retained; evidence and motivation distinguished.
E-B.md	G-B: Forum demand is inferred; journal undervalued; ranking is not a build order; creative value differs from insight.	Added core-experience caveat, standalone journal value, forum comparison, continuity advantage, and interpretation limits.	Nine features retained; ranking and structure preserved with clearer uncertainty.
E-C.md	G-C: Weak recurring loop; forum overlaps chat; context model unclear; post-save choices crowded.	E-C preserved. E-D adds the loop, contrasting forum views, confirmed context, simpler save state, and success checks.	Design clarified in E-D; all three levels retained.
E-D.md	G-C informed refinement of E-C: prioritize core; integrate patterns; clarify deletion and media scope. No separate direct review.	Finalized seven features; patterns embedded in detail/chat; deletion is a system rule; sleep/video secondary; four tabs justified.	Cohesive final design; sleep and video kept in scope, images first.
E-E.md	G-D: Rigorous is not necessarily pivotal; A version tracking not minimum; B/F omit dependencies; possible node overlap.	Clarified completion; A drops full version tracking; B adds it; F adds filtering; matching/grouping roles separated; script rerun.	Current A/B/F counts: 4/6/6; scores: 50/40/35. Formal articulation points remain unproven.
E-F.md	G-E: Reuse is not free; reflection can disrupt capture; resumed context and search can corrupt base state.	Accepted separate reflection jobs, thread-local goals/permissions, safe retries, and search without pattern mutation. Contracts noted here.	Node sets/costs retained: 2/0/0. Raw exhibit wording unchanged; implementation validation required.
E-G.md	G-E: A-to-B assumes SG1; permissions may leak via memory; transcription can overwrite corrected or deleted records.	Accepted explicit SG1 prerequisite, permission checks on retained context, source checks, stable identities, and separate scheduling.	Node sets/costs retained: 4/0/3 after SG1. Contracts accepted; no further script rerun.

Interleaved Exhibit Summaries

E-A - Interview Pain Points

Reviewed by G-A

Purpose

Identify the problems most closely related to preserving and reflecting on dreams. The ranking uses documented participant frequency, with importance breaking ties, and explicitly treats that order as provisional.

Main findings

Five ranked problems remain: fading dream details (4/6 documented accounts); unanswered personal questions (3/6); unsettling dreams and reassurance needs (3/6); limited or discouraging responses (2/6); and sensitive content that is difficult to disclose (2/6). The proposal reports 6/6 for forgetting, while participant-level evidence supports four.

Important distinctions

Forgetting does not imply demand for a recording app. The opportunity is preserving dreams users care about. P3 does not value recording; P6 selectively shares interesting dreams. The broader unanswered-question problem includes P5, while explicit criticism of search or AI comes from P1 and P2. P4 seeks reassurance without reporting the same continuing effects as P1 and P3.

Scope implication

Relevant, private reflection and timely capture are the clearest opportunities. Continuity, fragmented recording, and pattern discovery are related gaps with less direct evidence. Sleep integration and visuals are proposed benefits rather than established pain points. Enjoyment can motivate use without dissatisfaction.

Review outcome

G-A led to small wording changes about motivation, uncertainty, emotional variation, and P5. The five-point structure and ranking remained intact.

G-A - Review of Pain-Point Identification

Feedback on E-A

Overall assessment

The reviewer finds E-A usable and grounded. It separates participant evidence from proposed solutions, retains a conservative forgetting count, and distinguishes unresolved questions from explicit criticism of AI.

Key criticisms

Four reports of forgetting do not establish four willing users. Frequency should not determine product priority by itself. Emotional-distress language should distinguish lasting effects from a desire to understand nightmares. P5 offers an opportunity for richer responses, but the degree of dissatisfaction is unclear.

Recommended adjustments

Preserve personally meaningful dreams rather than treating all forgetting as a problem. Qualify the ranking and consider motivation and consequences. Narrow the distress statement and avoid claiming the app reduces fear. Recognize that enjoyable conversations and creativity are positive reasons to use the app.

What was adopted

E-A now says the counts describe experiences rather than app demand. Its ranking is provisional, P4 is distinguished from P1/P3, P5 is qualified, and creative enjoyment is acknowledged. These are targeted clarifications, not a rewrite.

Resulting direction

The review supports keeping the original pain-point inventory while improving the interpretation of frequency, user motivation, and emotional support.

E-B - Features and Desired User Value

Reviewed by G-B

Purpose

Translate desired value into nine possible features. Preferences are inferred from needs and behaviors; participants did not rank the features. The proposed advantage is keeping capture, personal context, and later reflection together.

Feature priorities

The order is voice capture; personal reflection; private AI forum; pattern finder; searchable journal; image/video generation; optional recall questions; selective sharing and optional real community; and sleep integration. Close positions are approximate, not a delivery schedule.

User-value connections

Features address preserving meaningful dreams, understanding personal connections, feeling supported, discussing sensitive experiences privately, enjoying unusual experiences, and revisiting history. The forum combines varied perspectives and engagement, while pattern finding makes history useful over time. Neither format is directly validated by the interviews.

Design boundaries

Capture, journal, and reflection form the clearest core. A journal is valuable independently and should precede history-dependent features. Visuals are creative interpretations and can succeed through enjoyment without improved insight. Sleep integration has limited direct support. A real community and AI personas meet different social needs.

Review outcome

G-B prompted stronger caveats, explicit journal value, a forum-versus-chat comparison, and clearer continuity and interpretation claims. The nine-feature structure was retained.

G-B - Review of Feature Priorities

Feedback on E-B

Overall assessment

The reviewer supports the connection between user value and concrete features. Plain English, inferred-demand labels, privacy distinctions, and low priority for sleep integration are strengths.

Key criticisms

The forum bundles several possible benefits into a relatively confident ranking. The journal deserves weight as a complete experience and dependency. A numbered list can imply precision absent from the evidence. The advantage over existing memos, messages, and general AI needs to be explicit.

Recommended adjustments

Compare private conversation with multiple AI personas. Treat capture, journal, and reflection as the core. Distinguish evidence strength from product priority, and show why combining capture, context, and reflection is useful. Keep creative enjoyment separate from claims of psychological insight.

What was adopted

E-B retains its order but calls it approximate and not a build sequence. It adds the journal’s standalone value, forum validation, and the continuity advantage. It explicitly avoids true-meaning claims and accepts creative enjoyment as a valid outcome.

What was not adopted

The suggested restructuring into new priority groups was not applied, preserving the requested document structure. The review informs E-B; it is not feedback on E-C.

E-C - Mobile App Design

Reviewed by G-C; finalized through E-D

Purpose

Design a private mobile journal around capture, revisiting, reflection, and exploration. AI is optional: a user can save a fragment and leave. The original dream account remains separate from generated material.

Functional levels

Basic functionality manages dream records. Advanced functionality provides voice capture, summaries/tags, recall prompts, personal reflection, and sleep integration. Novel functionality supplies the private AI forum, pattern finder, and image/video studio.

Software experience

Four tabs organize Journal, Forum, Patterns, and Settings, with a prominent recording action. Voice/text saving precedes processing. Reflection begins with a user goal, uses permitted context, and allows corrections. Forum threads offer persona viewpoints. Patterns show source dreams and editable matches. Visuals are labeled creative interpretations.

Privacy and reliability

Records and conversations are private by default; sharing previews the selected output. Context used by AI is inspectable. Processing failures preserve source records. Deletion excludes records from retrieval and patterns and removes associated outputs; exports remain outside the app’s control.

Review outcome

G-C found the concept coherent but asked for a clearer recurring loop, sharper forum distinction, clearer personal-context management, and stronger hierarchy. E-C was preserved, and selected improvements were incorporated into E-D.

G-C - Review of Mobile App Design

Feedback on E-C; informs E-D

Overall assessment

The reviewer values optional AI and separation of original records from generated content. The concern is that many reasonable features dilute the central reason to return.

Product criticisms

Connect morning capture to later reflection and eventual patterns. Distinguish the forum through simultaneous contrasting viewpoints. Avoid crowding the post-save state with AI actions. Define how personal notes, inferred context, and past conversations become permitted information.

Architecture and scope suggestions

Embed related dreams within records and reflection, not only a Patterns tab. Treat deletion of derived outputs and indexes as a system requirement. Add simple success indicators. De-emphasize sleep integration and video until core value is clear. Consider reducing permanent tabs.

What was adopted in E-D

The final design states a recurring loop, adds Done/Add details after saving, requires approval for suggested life notes, integrates related records, strengthens the forum, and adds consistency requirements and indicators of value.

Selective decisions

All seven final features and four tabs were retained. Forum supports resuming threads across sessions; sleep and video remain secondary capabilities, with images first. The review guided a new finalized exhibit rather than a direct rewrite of E-C.

E-D - Final Feature Specification

Finalizes E-C using selected G-C feedback

Product story

DreamLog combines capture, personal context, and reflection in one lasting history. The optional loop is capture, save, return, reflect or create, then discover connections. A saved dream remains a complete use of the app.

Final feature set

Seven features span the three levels: dream journal/CRUD; voice capture and lightweight organization; personal reflection; sleep integration; private AI forum; pattern finder; and image/video studio. Privacy, context controls, and selective sharing are shared requirements.

Cohesive design

Journal is the home and dream detail is the common entry point. Forum collects resumable private AI threads. Patterns shows broader history. Post-save is intentionally simple. Confirmed life notes are distinct from AI suggestions. The forum presents contrasting viewpoints, and related dreams appear inside reflection and record detail.

State and evidence

Generated outputs stay separate from original accounts. Edits mark affected outputs as outdated; deletion excludes records and late results from restoring content. Pattern evidence comes from recorded accounts, not generated visuals. Context is controlled and inspectable, while sleep associations remain exploratory.

Delivery and review role

Capture, journal, and reflection come first; forum/images, patterns, sleep, and video extend them. E-D is the design context for G-D’s architectural review, not a separately reviewed scope calculation.

E-E - Base Scopes and Full Script Results

Reviewed by G-D; current text includes the rerun

Purpose

Record the full scope evaluation at target capacity T = 5, including three alternative base scopes and their sequential stretch expansions. Ordinary CRUD and screens are excluded from architectural node counts.

Current base results

A: voice capture, recoverable transcription, and organized output; score 50.0, weight 100, 4 nodes. B: permitted-history reflection with inspectable context and resumption; score 40.0, weight 80, 6 nodes. F: evidenced recurring patterns respecting corrections; score 35.0, weight 70, 6 nodes.

Stretch expansions

A adds C with 2 nodes, then B with 4 further nodes. B adds C and E with zero new nodes. F adds G with zero nodes, then voice-capture A with 3. Second-stage costs are measured after SG1, not only against the original base.

Changes after G-D

A drops full source/output version tracking from its minimum capture path. B adds version tracking and clarifies resumed, inspectable reflection. F adds permission filtering. Matching proposes similarities; persistent grouping stores identities and user corrections; aggregation computes evidence.

Interpretation limits

B and F exceed the target of five nodes. Scores are supplied outputs; no scoring formula is available here. Required-component sets do not prove formal articulation points. The original overlap wording is retained verbatim, while explanations distinguish reuse from effort.

G-D - Review of Architectural Node Classifications

Feedback on E-E; E-C/E-D are design context

Overall assessment

A, B, and F are valid value-delivering workflows, and most nodes can qualify as rigorous features. A rigorous component is not automatically pivotal. Completion conditions and an architecture graph are needed for stronger claims.

Arc A findings

Microphone and APIs qualify through external dependencies. Recoverable jobs qualify when they handle restart, retry, and stale results. Summaries/tags are necessary only for organized output. Full version tracking is not minimum for initial capture and revisiting, though edit-and-regenerate workflows require consistency.

Arc B findings

History retrieval must do more than read one record. Permissions must govern retrieval and retained context before model calls. Provenance must identify actual supplied sources. Conversation state must support correction and resumption, not just message CRUD. Source consistency was missing from the set.

Arc F findings

Specify extraction, matching, persistent groups, and evidence aggregation as distinct responsibilities. Preserve user decisions through recomputation. Add eligibility filtering and maintain trustworthy evidence through edits/deletions. Avoid double counting matching and grouping.

Outcome

Accepted changes produced A/B/F counts of 4/6/6 and a script rerun. Providers and retrieval methods remain implementation choices. Counts should follow genuine dependencies, not be split or merged to meet T; formal graph validation remains outstanding.

E-F - Stretch Goal 1

Reviewed jointly with E-G by G-E

Purpose

List the first expansion for each alternative base. The exhibit reproduces net-new architectural node costs, not estimates of workload or elapsed development time.

Base A to C

After voice capture and organization, add discussion of an unsettling dream with resumable supportive follow-up. New nodes are persistent_conversation_context and ai_reflection_api: cost 2. The limited C workflow uses the selected dream and its thread rather than permitted historical retrieval.

Base B to C

The permitted-history reflection base already supplies conversation and AI capabilities. Supportive follow-up costs zero new component types. Goals, appropriate responses, per-thread context settings, and permission checks still require implementation and validation.

Base F to G

The pattern base already supplies semantic matching, source consistency, and permission filtering. Natural-language retrieval adds zero node types. Query results should remain separate from saved grouping decisions and resolve against current sources.

Review outcome

G-E’s boundary recommendations were accepted as implementation contracts without changing the dictionary. Reflection must not block capture, retries must not duplicate turns, and search must not mutate pattern state. The raw file still uses the script’s architectural-overlap wording.

E-G - Stretch Goal 2

Reviewed jointly with E-F by G-E

Purpose

List the second expansion for each base after its first stretch goal. Sequential prerequisites are essential to interpreting the reported costs.

Base A plus C to B

Add permitted-history reflection, source consistency, retrieval, source linking, and permission filtering: 4 additional nodes. The AI and conversation nodes already came from SG1. Directly adding B to A without SG1 would require 6 nodes.

Base B plus C to E

Sensitive reflection with controlled history reuses the base machinery: zero new nodes. This requires permission enforcement across retrieval, retained summaries, messages, and resumed sessions. Private presentation alone is insufficient.

Base F plus G to A

Add microphone capture, recoverable audio/transcription jobs, and speech-to-text: 3 nodes. Existing extraction organizes accepted transcripts. Stable identities, current-revision checks, and separate scheduling prevent retries or delayed results from corrupting pattern state.

Review outcome

G-E identifies A-to-B as a substantial state expansion, B-to-E as permissions/memory hardening, and F-to-A as a separately recoverable input pipeline. These accepted contracts do not change component counts, and no additional rerun followed the review.

G-E - Review of Stretch-Goal Leverage

Feedback on E-F and E-G; E-C is context

Overall assessment

Shared components provide useful leverage, but zero new nodes do not mean zero effort or risk. The three bases are alternatives; each cannot assume every feature in the full design exists.

SG1 risks

A-to-C can accidentally couple reflection failures to capture or attach late replies to changed records. B-to-C can leak goals or context across threads. F-to-G can let search mutate saved groups or return stale and excluded content. Separate jobs, thread-local state, and read-only search decisions address these risks.

SG2 risks

A-to-B adds source and permission changes across retrieval and conversation memory. B-to-E can reuse excluded content indirectly through cached summaries. F-to-A introduces retries, corrected transcripts, partial indexing, and service outages. Explicit prerequisites and source checks are essential.

Accepted contracts

Capture remains usable when extensions fail. Recheck permissions for new responses, retain per-thread choices, reject obsolete results, use stable request identities, and keep transcription scheduling independent from interactive retrieval. Arc C’s conversation node must include interruption, retry, and source-validity handling.

Final outcome

The arc dictionary and node counts were retained. Overlap should be interpreted as no new component types with additional behavior and validation required. These clarifications are accepted requirements, not evidence that the architecture has already been implemented or tested.

Full Exhibit Texts Available

E-E.md - Full Text

Verbatim exhibit text; Markdown markers retained.

Scope Evaluation Output

Target Capacity (T): 5 Rigorous Features

Top 3 Base Scopes

Arc A: Record a Dream by Voice and Return to Its Transcript and Organized Account

Score: 50.0
Weight: 100
Nodes: 4
Features: microphone_audio_capture, recoverable_audio_and_transcription_jobs, ai_summary_and_tag_extraction, speech_to_text_api

Arc B: Explore a Personal Question Using Permitted History, Inspect Context, and Resume

Score: 40.0
Weight: 80
Nodes: 6
Features: dream_source_and_output_version_tracking, response_source_linking, persistent_conversation_context, dream_and_personal_context_retrieval, context_permission_filtering, ai_reflection_api

Arc F: Discover a Recurring Pattern and Inspect Evidence That Respects Corrections

Score: 35.0
Weight: 70
Nodes: 6
Features: dream_source_and_output_version_tracking, ai_summary_and_tag_extraction, persistent_entity_and_theme_grouping, semantic_dream_matching, pattern_aggregation_and_source_linking, context_permission_filtering

Stretch Goal Expansions

Base: Arc A — Record a Dream by Voice and Return to Its Transcript and Organized Account

Stretch Goal 1: Arc C: Discuss an Unsettling Dream and Resume Supportive Follow-Up
Cost: 2 net-new nodes
New features: persistent_conversation_context, ai_reflection_api
Stretch Goal 2: Arc B: Explore a Personal Question Using Permitted History, Inspect Context, and Resume
Cost: 4 net-new nodes
New features: dream_source_and_output_version_tracking, dream_and_personal_context_retrieval, response_source_linking, context_permission_filtering

Base: Arc B — Explore a Personal Question Using Permitted History, Inspect Context, and Resume

Stretch Goal 1: Arc C: Discuss an Unsettling Dream and Resume Supportive Follow-Up
Cost: 0 net-new nodes (100% architectural overlap)
Stretch Goal 2: Arc E: Reflect on a Sensitive Dream Using Controlled History and Resume Privately
Cost: 0 net-new nodes (100% architectural overlap with Base + SG1)

Base: Arc F — Discover a Recurring Pattern and Inspect Evidence That Respects Corrections

Stretch Goal 1: Arc G: Find an Old Dream through a Natural-Language Query
Cost: 0 net-new nodes (100% architectural overlap)
Stretch Goal 2: Arc A: Record a Dream by Voice and Return to Its Transcript and Organized Account
Cost: 3 net-new nodes
New features: microphone_audio_capture, recoverable_audio_and_transcription_jobs, speech_to_text_api

Concise explanation

Reading note: T = 5 is the target. Base A has 4 nodes; B and F each have 6. Scores are reproduced from the script, whose formula is not supplied. B and F exceed T; they are alternatives to assess, not confirmed capacity-compliant selections.

E-F.md - Full Text

Verbatim exhibit text; Markdown markers retained.

Stretch Goal 1

Base: Arc A — Record a Dream by Voice and Return to Its Transcript and Organized Account

Stretch Goal 1: Arc C: Discuss an Unsettling Dream and Resume Supportive Follow-Up
Cost: 2 net-new nodes
New features: persistent_conversation_context, ai_reflection_api

Base: Arc B — Explore a Personal Question Using Permitted History, Inspect Context, and Resume

Stretch Goal 1: Arc C: Discuss an Unsettling Dream and Resume Supportive Follow-Up
Cost: 0 net-new nodes (100% architectural overlap)

Base: Arc F — Discover a Recurring Pattern and Inspect Evidence That Respects Corrections

Stretch Goal 1: Arc G: Find an Old Dream through a Natural-Language Query
Cost: 0 net-new nodes (100% architectural overlap)

Concise explanation

Reading note: Each base is an alternative scope. Zero net-new nodes means component reuse, not zero work or failure risk. Extension failures must leave the base workflow usable.

E-G.md - Full Text

Verbatim exhibit text; Markdown markers retained.

Stretch Goal 2

Base: Arc A — Record a Dream by Voice and Return to Its Transcript and Organized Account

Stretch Goal 2: Arc B: Explore a Personal Question Using Permitted History, Inspect Context, and Resume
Cost: 4 net-new nodes
New features: dream_source_and_output_version_tracking, dream_and_personal_context_retrieval, response_source_linking, context_permission_filtering

Base: Arc B — Explore a Personal Question Using Permitted History, Inspect Context, and Resume

Stretch Goal 2: Arc E: Reflect on a Sensitive Dream Using Controlled History and Resume Privately
Cost: 0 net-new nodes (100% architectural overlap with Base + SG1)

Base: Arc F — Discover a Recurring Pattern and Inspect Evidence That Respects Corrections

Stretch Goal 2: Arc A: Record a Dream by Voice and Return to Its Transcript and Organized Account
Cost: 3 net-new nodes
New features: microphone_audio_capture, recoverable_audio_and_transcription_jobs, speech_to_text_api

Concise explanation

Reading note: Costs are incremental after Stretch Goal 1. For base A, adding B costs 4 nodes after A + C, or 6 without SG1. Permission and lifecycle requirements are implementation contracts, not additional node counts.

Full Exhibit Texts Pending Collection

E-A — Full text

G-A — Full text

E-B — Full text

G-B — Full text

E-C — Full text

G-C — Full text

E-D — Full text

G-D — Full text

G-E — Full text