8.7 KiB
Presentation Story-Flow Audit
Audit date: 2026-07-13. This review uses the complete
rehearsal matrix, the dated
rehearsal log, the
presentation runbook, the
Q&A runbook, and the locally available private narrative
notes in random shit/. The matrix covers all 13 scenes; the log records
targeted replay/fallback checks and does not establish a completed live
end-to-end run.
Scene Audit
Scene 1 — Thesis
- Audience takeaway: The AI-agent goal narrows to an implemented workflow substrate.
- Visible proof: Title and substrate beats state the motivation and the typed lifecycle contribution.
- Missing or confusing: The external-planner boundary must be said before the audience asks where the agent is.
- Next action: keep
Scene 2 — The Problem
- Audience takeaway: Direct actions do not become reusable automation without durable contracts.
- Visible proof: The contrast and missing-contracts beats name schemas, persistence, traces, and recovery.
- Missing or confusing: The causal bridge from “AI lowers the barrier” to “the platform is still necessary” is spoken, not shown.
- Next action: keep
Scene 3 — Positioning and Related Systems
- Audience takeaway: lda.chat occupies a typed, provider-neutral substrate position rather than replacing adjacent systems.
- Visible proof: The landscape and lda.chat position map distinguish tool loops, scripts, hosted automation, agent graphs, and MCP.
- Missing or confusing: The comparison can feel like a catalog unless the axis and intended external-agent user are stated once.
- Next action: keep
Scene 4 — Planner and Runtime
- Audience takeaway: An external planner proposes while the runtime validates, executes, records, and resumes.
- Visible proof: Planner, runtime, and boundary beats make the ownership split visible.
- Missing or confusing: The audience needs the sentence that this is the thesis boundary, not a claim of a new planner.
- Next action: keep
Scene 5 — Workflow Lifecycle
- Audience takeaway: Draft, Artifact, Deployment, and Run are distinct records with different responsibilities.
- Visible proof: The lifecycle strip shows mutable authoring, immutable artifact, source binding, and persisted execution.
- Missing or confusing: The jump from vocabulary to architecture can be abstract if Deployment is not tied to concrete sources.
- Next action: keep
Scene 6 — Architecture Zoom
- Audience takeaway: The same public surface reaches API, runtime/providers, stores, and the execution kernel.
- Visible proof: Client, API, and runtime focus beats provide a semantic zoom with package and operation evidence.
- Missing or confusing: Three nested levels risk becoming a component tour unless each level answers what responsibility moves inward. NodeUse remains an optional deep dive.
- Next action: keep
Scene 7 — Agent Request
- Audience takeaway: A thin external-agent surface can request work without pretending to execute it.
- Visible proof: Send reveals the prepared conversation and discovery tool calls with no run claim.
- Missing or confusing: The chat surface could be mistaken for the thesis product unless the substrate remains the stated center.
- Next action: keep
Scene 8 — Prepared Workflow Lifecycle
- Audience takeaway: The request becomes six concise authoring stages: Discover, Draft, Diagnose, Repair, Artifact, and Deployment.
- Visible proof: Draft revision 2 is valid with
analyze.ok -> __end__; the transcript explicitly labels a prepared fault injection, not an accidental discovery, usingwf draft remove-route lda_report_workflow --revision 2 --step analyze --outcome ok, then shows invalid revision 3 withmissing_outcome_edgeand repairs it with the existingwf draft set-route lda_report_workflow --revision 3 --step analyze --outcome ok --to __end__command to valid revision 4. - Missing or confusing: Deployment binds and validates three local sources but does not execute; that boundary needs an explicit spoken pause.
- Next action: keep
Scene 9 — Run From Deployment
- Audience takeaway: Execution starts from a ready deployment through a public run operation.
- Visible proof: Input, operation, and graph beats show selected inputs,
workflow.runs.start, a run ID, and the typed boundary. - Missing or confusing: The live end-to-end path was blocked, so live output must not be implied from the replay-backed operation view.
- Next action: keep
Scene 10 — Typed Human Boundary
- Audience takeaway: A persisted run pauses at a typed issue-review decision with explicit submitted and revision-requested outcomes.
- Visible proof: Interrupt payload, selected issue, comment field, and both decision controls are visible in replay.
- Missing or confusing: The revision replay uses
run_recorded_lda_report_revision, so “same run” wording is false for that branch. - Next action: factual fix
Scene 11 — Resume, Output, Evidence
- Audience takeaway: A decision leads to inspectable output and trace evidence for the workflow run.
- Visible proof: Resume, output, and trace beats show status, report/issue result, interrupt continuation, and terminal frames.
- Missing or confusing: Submitted replay continuity is demonstrated; live continuity and revision same-run continuity are not yet established.
- Next action: factual fix
Scene 12 — Evaluation
- Audience takeaway: The 36 trials are bounded engineering evidence about operability and UX failure modes.
- Visible proof: Cohort, validity, and findings beats separate audited evidence from benchmark-style claims.
- Missing or confusing: The evaluation arrives after demo evidence, so the presenter must state that failures motivate the product-surface argument.
- Next action: keep
Scene 13 — Limits and Conclusion
- Audience takeaway: The contribution is a useful substrate, not a production agent, scheduler, or broad benchmark.
- Visible proof: Limits, future, conclusion, and questions beats distinguish implemented core from future layers.
- Missing or confusing: Q&A must not begin before the contribution sentence and limitations have landed.
- Next action: keep
Flow Findings
Transitions
- Scene 1 -> 2 works: the goal becomes the missing reusable-automation contracts.
- Scene 5 -> 6 works if Deployment is the handoff: lifecycle vocabulary becomes the architecture that owns it.
- Scene 6 -> 7 works: architecture hands off to a bounded prepared request surface.
- Scene 7 -> 8 works and is not a duplicate: request/discovery becomes staged lifecycle evidence.
- Scene 8 -> 9 is the critical boundary: the prepared route-removal fault injection is repaired before Deployment; Deployment prepares, and Scene 9 alone starts execution. Keep the explicit no-run wording.
- Scene 9 -> 10 -> 11 works as one climax in replay, but the revision branch has a factual run-identity defect and the live path is blocked.
- Scene 12 -> 13 works: evaluation limits become the contribution and future-work close.
Order, Duplicated Beats, And Q&A
The 13-scene order is coherent: motivation, positioning, boundary, vocabulary, architecture, request, authoring, then demo proof, evaluation, and limits. No duplicated demo beat was found. Scene 7 introduces the request, Scene 8 owns authoring and deployment, Scene 9 owns run activation, Scene 10 owns the decision, and Scene 11 owns resume/output/trace. The Q&A branch belongs after Scene 13's Questions beat; opening a prepared branch earlier would interrupt the argument and make a discussion answer look like core evidence.
Defects
- Factual: The revision-requested replay has a separate run ID while the surrounding same-run wording suggests continuity. Correct the recording or label that branch explicitly as a separate prepared recording.
- Factual: The live health boundary passed, but live Scene 9 -> 11 was blocked. Do not present live output or trace as rehearsed evidence.
- Visual: No visual defect was recorded in the available rehearsal log. The
matrix remains the acceptance checklist for both
1280x720and1024x768.
Rehearsal And Factual Follow-ups
These are outside the story-flow acceptance pass and must be resolved or clearly labeled before claiming stronger evidence:
- Complete and record a stateful live run through approval, resume, output, and trace.
- Decide whether the revision branch should preserve the submitted run identity or remain a separately labeled prepared recording.
Product Follow-ups
These are missing capabilities, not rehearsal fixes, and must not be presented as current proof:
- Add a factual input-file browser that distinguishes declared, selected, read, and produced files.
- Consider live agent execution, richer trace frames, and real source reads as future product layers, not rehearsal fixes.