Milestone 12 of 12

Public case study and competency dossier

Make the work inspectable without exposing the production layer.

Inspectable proof

sanitized demofailure evidencecompetency dossier

Package the work as evidence a hiring team can inspect. Required route: public/private audit (60 min), sanitized demo (90 min), case-study narrative (80 min), dossier and Git evidence (65 min), narrated defense (65 min), selected reading (60 min). Total: 7 hours. Optional external review and polish are excluded.

Entry diagnostic

Sort proposed assets into public, private, or omit. For every public claim, point to a runnable artifact; for every private item, explain the production need and access boundary.

Self-check

Can an architecture diagram expose customer filenames or storage locations even when no image is shown?

Why this matters in CineQC

A certificate says a curriculum was completed. CineQC Hybrid shows the work: data contracts, deterministic and learned signals, tests, benchmark decisions, integration boundaries, failures, and communication. The proof is valuable only when it is reproducible, privacy-safe, and modest about scope.

Visual model — two-layer proof

Narrated system overview · 35 seconds

Complete hybrid system map with fallback, review, report, and proof routes. If the video cannot play, this image and the transcript preserve the complete teaching claim.

Read accessible transcript · Download English captions

Build a clear public/private split. The public layer contains a reproducible sanitized demo, lawful generated/public fixtures, setup instructions, fixed prediction samples, benchmark narrative, model card, system card, dataset card, architecture diagram, failure gallery, tests, selected Git evidence, and concise narrated defense. The private production layer contains customer media, internal storage paths, production configuration, secrets, raw logs, full evaluation, and proprietary CineQC implementation details.

Publish an audit table with artifact, layer, reason, sanitization, license/provenance, and verification. A filename changed to “sample-01” is not sanitized if pixels, text, metadata, or commit history still identify the source.

Self-check

What makes a benchmark narrative defensible: the decision, protocol, result, failure evidence, and limitation—or only the best F1?

Case-study spine and competency mapping

Write the case study as production problem → constraints → selection rule → system → evaluation → integration → failures → limitations → next decision. Recompute displayed results from saved predictions. The model card describes model purpose and evaluation; the system card describes heuristics, model, fallback, review, privacy, and operation; the dataset card describes lawful sources, labels, splits, exclusions, and limitations.

Map each job-relevant competency to inspectable evidence: Python and debugging to focused commits/tests; visual computing to Media Probe and Artifact Lab; data discipline to label guide and dataset card; PyTorch to reproducible training; evaluation to saved predictions and benchmark; APIs to FastAPI contracts; privacy/product judgment to the layer audit and review routing; communication to the architecture walkthrough and narrated defense. Avoid proficiency percentages and invented certification language.

Sandbox lab

Assemble the CineQC Hybrid public case study plus clearly separated private production layer and inspectable competency dossier. Start from a clean clone or export, install pinned dependencies, run the sanitized demo, rebuild metrics, exercise API success/fallback/failure, and compare output checksums with the release manifest. Capture only public-safe screens.

Build a failure gallery with false positive, false negative, corrupted-input, timeout/fallback, and out-of-scope examples. Each card shows safe input evidence, expected/observed behavior, production consequence, diagnosis, present mitigation, and next experiment. Link claims to stable repository paths, release tags, small reviewable commits, and test commands as Git evidence.

Failure clinic

Remove copied private thumbnails, local usernames/paths, API keys, internal hostnames, raw OCR text, unverifiable metrics, cherry-picked successes, hidden fallback, and claims broader than the fixed tests. Inspect repository history and generated files, not only the current folder. If safe provenance cannot be demonstrated, replace the example with controlled generated media or omit it.

CineQC transfer

Create a 90–150 second narrated defense in the DVP motion language: dark editorial field, amber/blue signal paths, restrained kinetic type, readable evidence cards, and no decorative claim inflation. Show the problem, hybrid architecture, one benchmark decision, one failure, privacy boundary, and where a reviewer can reproduce the result. Captions and transcript are required; optional polish does not extend the course.

Self-check

Which competency mapping is stronger: “advanced ML — 90%” or “evaluation — recompute script, saved predictions, fixed-set checksum, and benchmark commit”?

Evidence gate

Release the CineQC Hybrid case study, sanitized demo, benchmark narrative, model/system and dataset cards, architecture diagram, failure gallery, Git evidence, public/private audit, competency mapping, setup/reproduction record, and concise narrated defense. Keep the private production layer clearly named but inaccessible. An optional external review may improve clarity after release; it is not required, not counted in seven hours, and does not become a credential.

One-page summary

• Public proof is lawful, sanitized, reproducible, and inspectable.
• Private media, secrets, configuration, logs, and proprietary details stay private.
• Cards constrain data, model, system, and claim boundaries.
• The case study shows decisions and failures, not only wins.
• Competencies map to files, commands, tests, commits, and artifacts.
• A short DVP-language defense makes the evidence navigable.
• External review and polish are optional and excluded from required time.