An authored choice, saved as editable vectors.
PulsePatch Calm helps an author prepare a static or hold-only alternative from a short supported Lottie. The author supplies the intention, inspects the actual proposed images and approves the states to export. The deliverable is an editable vector artifact with an approved-state record and a report.
This report separates release 218 software changes and their engineering checks from the recorded engine 216 development workflow and the closed study 215. The accepted film, original downloads and study results retain their historical identities. Updating report or interface code is not a new model execution, holdout or human evaluation.
What the author does
- Bring the original and a brief. State what the viewer needs to see. Supported input is standalone 2D vector Lottie, 1–4 seconds, 1–12 layers, dimensions up to 512 × 512 pixels, up to 1 MiB and 1–120 fps.
- Inspect the proposed states and their order. The model proposes one to three states. The author may revise them; a revision renders new images, invalidates approval and is marked as not revalidated by the model.
- Approve the current choice. The exact current proposal hash must be approved. The model cannot supply the author's approval.
- Keep the editable output. A successful package contains
calm.json,approved-states.json,report.jsonandREADME.txt. The original is retained separately.
The result is a static pose or a finite sequence of held states, with editable vector layers and shapes. This does not imply an external-editor integration, viewer-controlled steps or an indefinite pause. Choosing and editing states manually remains a valid baseline; a time or quality advantage over that workflow has not been measured.
Alternatives and workflow fit
Existing tools already support held frames, playback controls and editable animation authoring. PulsePatch’s candidate value is the constrained proposal → author approval → deterministic editable export workflow, with rendered images and CV evidence. Whether that process improves on competent alternatives remains unmeasured.
| Route | Documented capability or baseline | When it may fit |
|---|---|---|
| After Effects + Bodymovin | Hold interpolation and Freeze Frame; export an authored composition as Lottie. | An experienced author with the original project wants direct control. |
| LottieFiles Creator | General animation authoring and Lottie/dotLottie export. | A broader or already-established design workflow; no equivalent-feature absence is claimed. |
| Pause/freeze a player | lottie-web pause, goToAndStop and freeze controls. | A playback state is sufficient and a separate edited artifact adds little value. |
| Rive | Native .riv runtime assets, animation and state machines. | The team wants that interactive runtime and accepts its deployment format. |
| Equipped multimodal chat | A proposed baseline with the same brief, images, editable input and permitted render/edit tools. | Exploration or an author with reliable editing/checking tools. This baseline was not run here. |
| PulsePatch | A narrow supported Lottie, an author-approved 1–3-state proposal and editable held/static output with checks. | A candidate when this constrained workflow fits; relative effort, quality and adoption remain unmeasured. |
A documentation review on 27 October 2026 used official Adobe, LottieFiles, W3C and maintainer documentation; no competitor account was tested. The host still needs an appropriate reduced-motion/playback policy and user evaluation. No exclusivity, time saving, ROI, accessibility, superiority or willingness-to-pay result follows from this comparison.
Primary sources: https://helpx.adobe.com/after-effects/desktop/animate-in-after-effects/animation-keyframes/keyframe-interpolation.html ; https://docs.lottiefiles.com/en/creator ; https://github.com/airbnb/lottie-web/blob/master/README.md ; https://github.com/rive-app/help-center/blob/master/runtimes/advanced_topics/format.md ; https://www.w3.org/TR/mediaqueries-5/#prefers-reduced-motion .
How images, model decisions and checks fit together
The React interface talks to an asynchronous Python service on AWS. Chromium/lottie-web renders actual composition and isolated-layer images. OpenCV measures sampled appearance, change and visible layer contribution. Claude receives the brief, selected actual images and bounded context through server-owned Anthropic Files IDs. Claude is an external Anthropic service; this project does not host it on AWS.
The model proposes requirements tied to quotations from the brief, then links states to those requirements. The complete validated goal is bound to the source and brief. Coverage is a structural check: it cannot prove that the goal captured every instruction or that a frame communicates it. A same-model review sees the full brief and selected images without the proposer's labels or reasons. Review can still miss errors.
The retained pipeline samples at most 32 rendered times and offers at most 12 candidates. Stable, transitional and unknown are sampled physical classifications; changes between samples can be missed. Adaptive observation, proposal and review share at most four physical model decisions. Budget, transport, physical-validation and decision-limit stops are technical outcomes, not semantic abstentions.
One concrete feedback trace
| Recorded engine 216 step | What actually happened |
|---|---|
| Initial proposed confirmation | The model described the state at 2.48 seconds as stable. |
| Sampled physical check | CV classified it as transitional and rejected the stability claim. Visibility was true; this was not a visibility failure. |
| Later decision | The next proposal received that feedback and selected the stable confirmation at 2.81240625 seconds. |
| Review and approval | A third model call reviewed the selected states as supported. The root AI operator inspected and approved the current proposal without manual revision. |
| Saved artifact | The actual downloaded editable export passed all eight mandatory checks. |
This is a selected development trace, not a controlled comparison showing that feedback caused a semantic advantage. The narrow engine 216 recovery for known, settled zero/multiple-tool responses was not exercised by this run. It does not add decisions, retry uncertain dispatches, remove charge holds or relax export checks.
What is verified, and which version it belongs to
| Identity | Evidence boundary |
|---|---|
| Software release 218 | Current release 218 uses source commit 9a76b31e6e7376d7212b730b5bff1e6c52a926b0 and image sha256:15a6b38ad2b1280dc250541ea5f899a17e23a6003560d38afb7f563dc6cb216d. Anonymous verification matched all 14 served frontend files and health identity API 210 / engine 218. The actual client index hash is 6c0b30fe8273a16bc350128fba8b52c1acb3cf10ce4c1ec6ad3e28cdb7734bca. See the safe verification record. This is software/deployment evidence, not a new model run. |
| Recorded engine 216 demonstration | Run real-public216-20261026T202035515Z, job dc0913fcf9f7410f81e4eb5b74ee3db7. One Analyze action; no retry or manual revision. Three approved states at 0, 1.49995 and 2.81240625 seconds. Outcome SEQUENCE_READY. |
| Recorded 216 accounting | Three settled model calls, USD 0.130228. Proposal wait 90.581 seconds; export wait 26.749 seconds. The separate AI-operator approval checkpoint is excluded. These are one-run observations, not human time-saving or future-cost estimates. |
| Closed study 215 | Six paired synthetic briefs and twelve original arm attempts on frozen engine 215. The study, all outcomes and its sealed factual reference retain their 215 identity. |
Release 218 changes and engineering verification
The actual 218 source ZIP is 2,776,483 bytes with 206 members; SHA-256 cc33eca56b6e67a747bc8ed2023945384968753978f6aef3f56db60c9afaaeb1. Curated maps bind 44 runtime files, 14 compiled frontend files and 55 frontend inputs. A fresh rebuild from this exact source ZIP used its unchanged standard Dockerfile and normal dependency-registry access, matched all 44 runtime and 14 frontend fingerprints, rendered three own-fixture PNGs and passed six pure coverage tests. The rebuilt image remains distinct from the deployed image. The archive-bound verification record separates this build/render evidence from hosted-model answers and OCI-byte identity, neither of which is promised.
Architecture and service boundaries
Select a supported Lottie locally; send it and the brief when analysis starts. Inspect proposals and download accepted output.
HTTPS: input, proposal state and downloads
The container serves the interface and orchestrates asynchronous jobs, author approval and budget admission.
Render source and proposed states
Produces actual composition and isolated-layer PNG images.
Actual rendered pixels
Measures sampled appearance, change and visible layer contribution. Python separately validates inferred requirement mapping and set coverage.
Only after current author approval
Reopens and freshly renders the exact saved file. An accepted result can be downloaded.
Persistent job/artifact storage and a shared SQLite budget ledger are mounted on AWS. Provider credentials remain server-side. Artifact removal is operator-managed.
Python service exchanges selected images and bounded context for semantic proposals and review
External to this AWS deployment, accessed through Anthropic’s API. Supplies model decisions; cannot provide the author’s approval.
Bounded perception → decision → action
At most four model decisions total, shared by observation, proposal and same-model review. Budget/accounting gates can stop earlier.
The author supplies the intention, rather than the model inventing a task.
Chromium/lottie-web renders; OpenCV 5 measures sampled appearance, change and visibility.
Actual images + bounded observations
Requests observations or proposes brief-quoted requirements and candidate states using the supplied visual evidence.
Observation request → AWS render/measure → next model decision
Returns to step 3 within the same four-slot limit. Rendering itself does not consume a model slot.
Proposal only
CV checks visibility/stability; Python validates requirement mapping and set coverage.
↶ Negative CV feedback → a later decision at step 3
Only if an admitted decision slot and budget remain. Coverage feedback is structural, not a CV measurement. No unlimited retry or retry of an uncertain charge.
Physically and structurally supported proposal
Sees the full brief and actual selected images without the proposer’s labels. This consumes the same four-decision allowance.
Supported result; clarification or technical stop gives no accepted export
Model review cannot replace this approval.
Manual revision → AWS render/check → new author approval
Editing invalidates approval. The revised choice does not inherit model review or trigger a new model review automatically.
Current hash approved
Freshly render the exact saved artifact; accept only if every required gate passes.
Editable Lottie, approved-state record, report and integration instructions.
| Release 218 change | Meaning and verification boundary |
|---|---|
| Optional workflow guide and clearer intake | The Guide is off by default and explains the current workflow without starting analysis. Its progress reflects completed steps. Supported input, output and operator-managed retention are explained before analysis; file selection is local. Removal requires contacting the creator. No deletion endpoint or automatic expiry is promised. |
| Explicit software identity | Health keeps compatibility fields version: "210" and api_version: "210", separately reporting engine_version: 218 and the actual compiled client index hash. This distinguishes the current software from the recorded engine 216 demonstration. |
| Complete bounded exposure evidence | The eight export checks retain their criteria. The new check.evidence uses schema pulsepatch-calm-export-exposure-1 and retains the validated finite sample times, source-state indices, minimum exposure and same-renderer comparisons, at most nine samples. This is mechanical evidence, not proof of meaning. |
| Accurate proposal rejection reasons | New public diagnostics distinguish visibility, stability and coverage. Coverage validates inferred mapping and set coverage; it is not a physical CV measurement. The model-facing projection retains the established 216 wire contract; the underlying physical checks and model prompt are unchanged. Historical 216 traces retain their original generic labels. |
| English report and player instructions | English requests receive readable English report labels and details; Polish remains available. The exported README explains explicit play with loop: false and autoplay: false. A finite held sequence, the host's loop setting and viewer-controlled stepping remain separate concepts. |
Engineering test scope and QA setup history
Frontend checks passed 70 unit tests and 29 browser tests using offline fixtures; the production build produced 14 files. They cover the optional guide, truthful completion, keyboard and focus, retained work, history, reduced motion and 390-pixel reflow. CSS zoom 2 and a separate 720-pixel reflow viewport were checked; native browser-toolbar zoom and a physical phone were not tested.
A public desktop-browser check also opened and closed the Guide, chose an existing example and goal, checked completed-step counts, used Escape and switched PL/EN while preserving the source, brief and URL state. Retention was visible and the console was clear. It activated no Analyze or approval control and produced no new model run or export.
A bounded public load observation recorded desktop first contentful paint at 376 ms and DOMContentLoaded at 257.1 ms on this machine. A separate 390-pixel viewport load also checked the Guide entry, English metadata and layout. Each loaded nine GET resources with no failed requests, console errors, page errors or horizontal overflow. The fresh contexts shared one Edge 156 process and network state, with no controlled cold cache. These observations do not establish device comparisons, a performance guarantee, Lighthouse, full Core Web Vitals or field performance.
Four native-renderer tests passed on the bound image, with zero errors or skips and no hosted-model calls. They cover asymmetric easing and nonzero input time, ordered holds, a Bezier morph with local colour features and an occlusion-feedback fixture. This is deterministic engineering coverage, not a new generative comparison or human-benefit result.
Two earlier QA setups failed and remain preserved: the first used a read-only HOME and stopped before raster output; the second made HOME writable but omitted production --init, then encountered Chromium crashes after repeated launches. The final run kept the image, tests and resource limits, added production docker-init and passed all four tests in 39.318 seconds. It recorded no zombie processes, PID-limit events or OOM events, then removed its container. These setup attempts are separate from product/model outcomes and closed study 215.
Release 218 has no new generative model run or accepted semantic demonstration in this packet. The adapter/provider retains its established 216 lineage. The accepted film, original downloads, recorded costs, failures and closed study 215 remain unchanged.
The recorded engine 216 source commit was 51144a1205d1c84f7ecae1098495e4ad0a1f08a3, image sha256:13a203fab57e61fc920589a834a337d666e9899aa26a5e0a15d1eb65d0f6aeb6, with 44 runtime and 14 compiled frontend fingerprints verified. Those are the historical demonstration’s bindings; they must not be presented as a release 218 image.
The actual saved demo calm.json has SHA-256 4f33ebe088d1ade53a9c4246cfb4bf3fc1de480e14c605b6d2f110b069fd6845. Its report has SHA-256 0f4bf936e18e62ebe462f0211f66d4fd3f76b6bc4bbee5bc99e74229cb42a3e2. See the public UI receipt, actual exported Lottie, actual report and actual package.
Eight checks establish an export contract
The mandatory checks cover current author approval, complete supported pose profile, equivalence to source poses, absence of continuous motion, state order and exposure, visible selected-layer contribution, editable structure and exact saved-export integrity. The exporter reopens and freshly renders the exact saved calm.json. A failed or missing check cannot produce an accepted download.
The preserved, actual downloadable engine 216 report truncates one detail field, export_states_order_exposure, to 1,000 characters. Its passing check status is recorded, but that string is not a complete raw comparison or exposure schedule. A held-state file also depends on its host player: configure loop: false when a single non-looping play is required. The exported file alone does not override the host’s loop setting.
These checks concern the approved visual export contract. Human understanding, medical comfort and accessibility compliance were not measured. The supported profile excludes 3D, hierarchy/precompositions, masks/mattes, effects, non-normal blending, text/fonts, external images, expressions, spatial-Bezier paths, split positions and unknown animated properties; native rendering must succeed.
The complete study result includes its limits
| Closed study 215 | Adaptive | Strong one-pass + CV |
|---|---|---|
| Complete correct first and final outcomes | 6/6: 4 approved exports + 2 appropriate author questions | 4/6 |
| Semantic executions | 6/6 | 5/6 |
| Actual approved exports | 4; all passed eight checks and mandatory gates | 3; all passed eight checks and mandatory gates |
| Physical model requests | 12 | 6 |
| Known settled model cost | USD 0.411234 | USD 0.164320 |
| Atomic AI-operator actions | 10 | 9 |
The registered material-agent-advantage criterion failed. It required six semantic executions in both arms; one-pass had five. Its technical failure remains in the six-attempt denominator. An unsupported visible-fact premise in another one-pass review question remains an error. Six correct adaptive outcomes do not mean six exports.
The sources, factual reference, operation and assessment were AI-based. Concealment was limited; the operator knew the arms. These were small synthetic cases, not a human study, full double blinding, population reliability estimate or product/jury rating. First and final factual outcomes were identical. No new hosted-model comparison was run for the assessment or presentation change. The earlier 213 comparison tied at 4/6 each; different cases do not establish a causal improvement from 213 to 215.
The comparison did not match model-call counts or spend between arms. The adaptive arm used more calls and cost more, so the observed outcomes do not establish a material advantage under an equal budget. No human benefit was measured.
The 18 study calls cost USD 0.575554 known settled, with no new unknown-charge hold. Historical reserves remained; infrastructure and artist time were not measured. Cheap last-frame and uniformly spaced three-pose baselines retained complete visible information in 2/4 feasible ready-reference cases, but had no semantic decisions, author approvals or product-ready outputs. This does not establish cost efficiency or a benefit over manual authoring.
Full factual summary JSON and all six paired cases retain unchanged study 215 bytes. Their 216 filenames identify the evidence packet, not a newer study. The complete evidence archive includes all twelve original attempts, seven approved export triplets, 108 declared study artifacts and all 1,668 frame records, stored losslessly through a complete frame index and 272 unique PNG objects.
Intended buyer, offer and costs
Our intended buyers are small animation studios and product design teams maintaining libraries of short Lottie instructions. A potential offer to validate is per-asset credit packs for a repeatable workflow of author-approved, editable variants. This is a business hypothesis; pricing and willingness to pay have not been tested. Main costs are model requests, rendering, hosting and the author’s review time. Author time savings and unit economics have not been measured.
Film, provenance and privacy
The main film is 2 min 25 sec. Campus and designer scenes are disclosed fiction, generated with Veo and ImageGen references; Melissa is an AI voice through Epidemic Voices. The student’s request for less motion is fictional, not a real quotation or study observation. Actual UI, approval and export come from the recorded engine 216 run operated by the root AI operator. The film demonstrates no noticeboard/OCR, campus-system integration or recipient-understanding result. Exact transcript · Exact captions.
Main-film caption provenance: the supplied narration contains all 211 script words. Caption timing uses coarse non-ML blocks based on measured pauses and text order, not ASR or word timestamps. It should not be read as word-level alignment; the final media receipt defines the delivered timing and review scope.
Current production credits and clearance: AI voice Melissa via Epidemic Voices, with ElevenLabs shown as provider; music A Lovely Sunday by Franz Gordon, courtesy of Epidemic Sound. The existing source music and exact pilot voice are documented, and the project host and intended channel were checked in the native safelist. Licensed audio is used only in local synchronized editing, not supplied to video/image AI tools or distributed as a standalone asset. Source-clearance actions and final-film review are recorded separately; the final film receipt defines the review scope. Raw music and standalone voice files are not part of this public packet.
Music source credit: https://www.epidemicsound.com/music/tracks/54ee69b4-f1c6-45d7-bdc6-c911e2cab27c/ . Generated voice and fictional human scenes must remain disclosed in the finished film.
The genuine UI recording remains distinct from the capture defect: the recorder produced three empty standalone PNG extracts. Those original empty files were preserved. Existing same-job PNG bytes were later recovered read-only and separately attested, without rerendering or model calls; the recovery does not make the original extraction valid. The earlier failed public 215 take is retained separately. The successful 216 take is not a fresh holdout.
Selected actual images and bounded context go to external Anthropic Files/Messages. The service verifies returned metadata, requests one-hour provider-file expiry and attempts deletion. These do not guarantee complete provider erasure or define application retention. Application jobs and artifacts persist until operator removal, with no self-service deletion endpoint. Credentials stay server-side; existing accounting history and unknown-charge holds remain.
Comparative usefulness, market demand and a business model remain unvalidated hypotheses. Created by Damian Karolewski with AI coding assistants. Questions about the project or stored material can go through the author’s Devpost profile. Original source, documentation and artwork rights remain reserved; inherited dependency notices retain their separate terms.