jRay/SPEC.md

jRay — software specification

Status: alpha. Core read path ships; the schema bump, the exchange client and the audio signature do not.

This is a software spec: its job is to implement the system spec, which owns everything spanning more than one repo. Requirements here trace up to an SR-nnn or a PR-nnn; the prose below is the detail. The authoritative ID list with status lives in docs/requirements.md.

jRay is the Jellyfin plugin: it consumes presence data, displays it in the player, and owns the truth-file format that the other two components produce and exchange.

---

0. Requirements

IDs are JR-nnn, zero-padded and permanent — a withdrawn requirement keeps its number, because renumbering is what produces orphan TRACES tags (system spec §6).

Group	IDs	Where addressed
Truth-file format	JR-001 … JR-007	§1
Sources and precedence	JR-008 … JR-011	§2
Read API	JR-012 … JR-014	§3
Work discovery, policy, coverage	JR-015 … JR-019	§4
Player overlay	JR-020 … JR-024	§5
Manifest exchange client	JR-025 … JR-037	§6
Egress and privacy	JR-038 … JR-041	§7
Audio signature	JR-042 … JR-045	§8
Human-in-the-loop association	JR-046	§9

JR-038 … JR-041 exist because PR-005 had no software row anywhere. The system spec notes that "leak nothing about what the user owns" is preserved structurally — by SR-004 and GR-005 both being prohibitions — and that a goal held only by prohibitions needs watching. jRay is the component that actually opens a socket, so it is the right place for that goal to become checkable.

---

1. Truth-file format — JR-001 … JR-007

JR-001 — This document is normative for the format

The truth file is produced by scene-actor-extraction, read by this plugin, and transformed into a Jmanifest by the exchange client. Three repos touch it, so exactly one must define it, and the system spec §1 assigns that to jRay. Other repos reference this section rather than restating the schema.

The truth file is not the Jmanifest. It carries installation-local fields (movie, jellyfin_id) that the exchange strips, and lacks the portable identity block the exchange adds. See §6.

Current: Models/TruthFile.cs, schema_version: 2 — the schema below is implemented, with Models/TruthScene.cs, Models/TruthExtraction.cs and Models/TruthCut.cs.

Gap: the ownership claim is still not enforced — nothing checks that the other two repos reference this section rather than restating the schema. JR-001's verification tier is static for that reason, and no such check exists yet.

JR-002 — schema_version: 2

{
  "schema_version": 2,
  "movie": "/data/movies/Movie.mkv",
  "extraction": {
    "sample_fps": 5,
    "extinction_sec": 12,
    "gallery_size": 1820,
    "gallery_scope": "global",
    "pipeline_version": "scene-actor-extraction 0.4.1"
  },
  "cut": {
    "runtime_sec": 6420.5,
    "audio_signature": "v1:v7fA3k…"
  },
  "actors": [
    {
      "name": "Steve Buscemi",
      "imdb_id": "nm0000114",
      "tmdb_id": "884",
      "jellyfin_id": "abc123-guid",
      "scenes": [
        { "start": 191.6, "end": 209.2, "belief": 0.98, "route": "live" },
        { "start": 438.2, "end": 465.6, "belief": 0.81, "route": "deferred" }
      ]
    }
  ]
}

Field notes:

• schema_version — system-level (SR-003), incremented once per breaking change and referenced by the same number in all three repos.
• movie — absolute path at extraction time, informational only. Stripped on contribution (JR-034).
• extraction.* — provenance. sample_fps, gallery_size and pipeline_version move here from the top level so the truth file and the Jmanifest's extraction block have the same shape, rather than differing for no reason.
• extraction.extinction_sec — replaces anneal_sec, which is deleted, not retained as a vestigial 0. It is the re-acquisition timeout that shapes window extent, so it is what a consumer needs in order to interpret a window.
• extraction.gallery_scope — "global" or "limited". The strongest single quality signal when two manifests compete for one cut.
• cut.runtime_sec — the decoded duration the timings came from. Required for contribution; the primary alignment guard.
• cut.audio_signature — optional, v1:-prefixed. See §8.
• actors[].scenes[] — objects, not float pairs. start/end in seconds, inclusive, sorted. belief is the accumulated posterior that justified the claim; route is "live", "deferred" or "pooled" (extraction AR-017).

Belief is an attribute, not part of identity. Two servers that validated the same upload must agree on its content_id, and belief is a producer-side estimate that may legitimately differ between pipeline versions for identical timings. It replicates the way audio_signature does — see the server spec §9a.

Current: implemented. scenes are TruthScene objects carrying belief and route; extraction.* and cut.* are their own types; anneal_sec and the top-level sample_fps are deleted, not zeroed. ManifestConverter carries belief, route and both blocks through from a fetched manifest rather than flattening them. Gap: none — this shipped with extraction IR-002 and the server's jmanifest_version: 2 as the one coordinated SR-003 bump.

JR-003 — Unknown schema_version is refused, never guessed

Decision: flag day. The plugin accepts schema_version: 2 and rejects everything else, on every path — sidecar read, managed PUT, and fetched manifest. There is no transitional dual-accept.

All three components are pre-release and move together, and the alternative carries a cost that outlasts the transition: a v1 read path is the one nobody exercises, so it is the one that rots, and it would have to be carried through every subsequent change to the reader.

The consequence must be stated plainly rather than discovered: existing v1 sidecar files on disk stop being read at the bump, and stay dark until the library is re-extracted. The plugin logs this per item, naming the file and the version found, rather than silently reporting no coverage — an item that looks un-extracted when it was merely stale is the failure mode that wastes a user's compute.

Current: implemented. Services/TruthSchema.cs holds the single SupportedVersion = 2 and is consulted on all four paths — sidecar read, managed store load, managed PUT, and the converter that writes a fetched manifest. Rejections name the file and the version found, per the paragraph above.

Managed truth is checked on load as well as on PUT, because data written by an earlier plugin build is already on disk and never passed today's PUT.

One consequence is worth stating: a v1 file usually fails to deserialise before its version can be read, since scenes was a float pair and is now an object. So both the parse-failure path and the version-gate path must name the file — a rejection the user cannot attribute to a stale sidecar is the failure mode JR-003 exists to prevent. Gap: none.

JR-004 — A window is a scene-membership claim

This is SR-002, and it binds this plugin harder than it binds anything else, because jRay is where the claim reaches a human.

An actor who turns away, is occluded, or is off-camera while the shot cuts to whoever they are speaking to is still present. Two windows mean a genuine departure and return, not a break in detection. Gaps shorter than extinction_sec were absorbed upstream and are claimed as presence.

The plugin therefore never reinterprets, merges, splits, or trims windows. It stores and serves what it was given. The one permitted transformation is the timebase offset of JR-030, which shifts every window uniformly and so preserves the claim.

Current: satisfied. PresenceLookup is now the unit that decides presence, so the semantics live in one tagged place instead of being implied by a LINQ predicate in the controller. UT-021 pins that adjacent windows such as [0,10] and [10,20] are not merged, and UT-022 that a truth file round-trips byte-identical. Gap: none.

JR-005 — Query semantics, and how presence is presented

An actor is present at t if any window satisfies start <= t <= end.

The presentation must not assert instantaneous visibility. SR-002 is explicit that a consumer must never interpret window boundaries as "the face was detected here", and the overlay is the exact place that misreading would be made user-visible. "On screen now" is a claim the data does not support; "in this scene" is the claim it does.

This is a wording requirement, not a hedge — it is the difference between the product being right and being a worse version of a frame-by-frame detector.

Current: satisfied, in logic and in wording. Bounds are inclusive at both ends (UT-016/017), a zero-length window is a real sighting rather than a degenerate one to discard (UT-018), and overlapping windows resolve (UT-019).

The wording was the larger half. The overlay now carries an "In this scene" heading — previously it rendered a bare list, which asserted nothing but also told the viewer nothing, and a viewer's default reading of a paused frame is "these people are on screen". The model type ActorAtTime became ActorInScene, and README.md no longer contains the word "on screen" anywhere; it stated the forbidden reading outright in seven places, including the opening sentence. Gap: none.

JR-006 — Numerous windows

SR-002 warns that windows may be numerous and consumers must not assume a handful of long ones. Track-extent presence with a short extinction_sec produces many short windows per actor, and the previous design's few long ones were an artefact of the over-claiming that was removed.

The read path must therefore treat per-actor windows as a sorted sequence to be searched, not a short list to be scanned, and the jray?t= response must stay small regardless of how many windows an actor has.

Current: satisfied, and now measured rather than assumed. UT-023 builds 50 actors × 1000 windows and asserts the jray?t= result is bounded by actor count, never window count — which is what keeps the response small however finely presence is sliced.

The lookup is a full scan, deliberately. An early exit on start > t would exploit the sortedness the format requires, but it would silently under-report the moment one producer emitted windows out of order — a correctness risk traded for a saving that does not register at this scale. UT-020 pins that unsorted input still resolves. WindowsAreSorted exists as a diagnostic for surfacing such a producer bug, not as something correctness depends on.

Gap: none for lookup. The whole truth file is still held in memory per cached item, which is a memory question rather than a query-cost one and is untouched here.

JR-007 — Identity is public identifiers

Each actor carries imdb_id, tmdb_id and jellyfin_id, any of which may be "". Resolution prefers jellyfin_id (a Jellyfin Person GUID) when non-empty, and otherwise matches imdb_id/tmdb_id against the item's People ProviderIds. Never a name alone — names are ambiguous and unstable (SR-001).

jellyfin_id is the exception that proves the rule: it is meaningful only on the instance that produced it, which is exactly why the exchange strips it (JR-034).

Current: implemented. Gap: none.

---

2. Truth-data sources and precedence — JR-008 … JR-011

JR-008 — Sidecar discovery

For Movie.mkv, the plugin looks for Movie.jray.json beside it — suffix configurable, default .jray.json, resolved from the item's media source path.

Current: implemented. Gap: none.

JR-009 — Managed truth push

PUT/DELETE .../Truth let a worker that cannot write beside the media file deliver results over HTTP. Stored under the plugin's configuration directory, keyed by item id, independent of the library filesystem.

Current: implemented. Gap: none.

JR-010 — Precedence and provenance

There are now three sources: a sidecar file, a push from a local worker, and a manifest fetched from a server. Managed truth — pushed or fetched — takes precedence over a sidecar.

Fetched manifests are stored through the managed store, so precedence stays a two-way rule rather than a three-way one, and the read path does not learn about the exchange at all.

But the three are no longer interchangeable, so provenance is recorded with the stored truth: which source it came from, and for a fetched one, which server and at what match tier. A loose-tier fetch from a third-party server and a locally-computed sidecar are not the same claim, and JR-036 requires the difference be surfaceable.

Current: satisfied. Two-way precedence in TruthDataService, and TruthProvenance records source, server, tier, applied offset and caveat. GET .../Provenance serves it.

Two decisions worth keeping. Provenance is stored beside the truth file, never inside it — injecting fields would mean the bytes served back are not the bytes the producer wrote, which is the property JR-004 turns on (UT-026 pins this). And the applied offset is recorded because it is otherwise unrecoverable: once JR-030 shifts the windows they look native, and nothing would say they had been shifted.

A sidecar's provenance is derived rather than stored — it is local, and its timestamp is the file's own. Gap: none.

JR-011 — Caching

Loaded truth is cached in memory for a configurable duration. Any write — managed PUT, DELETE, or a stored fetch — invalidates that item's entry immediately, so a push takes effect without waiting for expiry.

Current: implemented. Gap: none.

---

3. Read API — JR-012 … JR-014

JR-012 — GET /Plugins/JRay/Items/{itemId}/Timeline

Returns the full truth file (§1), or 404 if no truth data exists from any source.

JR-013 — GET /Plugins/JRay/Items/{itemId}/jray?t={seconds}

Returns an extensible "context at time t" envelope, or 404:

{
  "actors": [
    { "name": "Steve Buscemi", "imdb_id": "nm0000114", "tmdb_id": "884", "jellyfin_id": "abc123-guid" }
  ]
}

Future fields (locations, trivia, and per JR-005 a presence caveat) are added to this object without changing the route, so clients must ignore unknown keys.

JR-014 — Authorisation

Route	Requires
Timeline, jray?t=	Authenticated Jellyfin user token
Truth, Tasks/*, Policy/*, Coverage/*, and §6's fetch routes	Administrator role
ClientScript	Anonymous — it is injected into a page served before login

Current: all three implemented as stated. Gap: none.

---

4. Work discovery, policy and coverage — JR-015 … JR-019

JR-015 — GET /Plugins/JRay/Tasks/Pending?limit=10

A random sample (default 10, max 100) of movies and episodes with no truth data:

[ { "item_id": "abc123-guid", "path": "/data/movies/Movie.mkv", "name": "Movie" } ]

Randomness is the design: repeated polling spreads work across the backlog without the server tracking who is working on what, and two workers polling concurrently mostly do not collide. An empty array means nothing is left, or that everything remaining is a virtual/missing-path item.

JR-016, JR-017 — Prioritise / ignore rules

Rules steer the queue: each targets a genre, a series, or an item, and either prioritises (front of the queue) or ignores (hidden entirely). This is how an admin says "never extract anime", "this series first", or "skip this one".

Resolution picks the most specific match: Item > Series > Genre. A rule is keyed by scope + value, and setting one replaces any existing rule for the same key — so a single target can never be simultaneously prioritised and ignored. Cross-scope conflicts (a prioritised series inside an ignored genre) are resolved by specificity: the series wins.

{ "scope": "Genre", "value": "Anime", "action": "Ignore", "label": "Anime" }

value is a genre name, a series id GUID, or an item id GUID; genre matching is case-insensitive. Persisted to policy.json in the plugin's configuration directory.

JR-017 is the constraint worth stating separately: rules affect work discovery only. They never change the overlay or the read endpoints. An item you ignore for extraction still shows its overlay if truth data happens to exist — because the rule expresses "don't spend compute here", not "pretend this doesn't exist".

Endpoints: GET/PUT /Plugins/JRay/Policy/Rules, and DELETE /Plugins/JRay/Policy/Rules?scope=&value= (idempotent).

JR-018 — GET /Plugins/JRay/Coverage

How much of the library has truth data, overall and by media type and genre:

{
  "total":  { "total": 1200, "covered": 300, "pending": 850, "prioritised": 40, "ignored": 50 },
  "by_media_type": [ { "label": "Film", "counts": { "…": 0 } } ],
  "by_genre":      [ { "label": "Anime", "counts": { "…": 0 } } ]
}

prioritised is a subset of pending. Percent done is covered / (total - ignored) — ignoring a genre does not drag the percentage down, because ignored items are intentionally out of scope, not outstanding work. An item counts toward every genre it carries, so genre rows overlap and need not sum to the library total.

JR-019 — Pickers

Coverage/Genres, Coverage/Series, and Coverage/Items?search=&limit= populate the rule editor's dropdowns, each returning [{ "value": …, "label": … }]. An absent search returns [].

Current (JR-015 … JR-019): all implemented. Gap: none functionally — these traced to no requirement until this register existed, which by the gate's own definition read as scope creep. They serve PR-003 (fully automatic, no per-title manual work): steering a queue is how automation is directed without becoming per-title labour.

---

5. Player overlay — JR-020 … JR-024

JR-020 — The overlay

Jellyfin has no plugin hook for player UI, so jRay adds <script defer src="/Plugins/JRay/ClientScript"></script> to the web client's index.html. The script listens for the player's pause event, calls jray?t= for the current item and timestamp, and renders the scene's cast.

Per JR-005 it presents scene membership, not instantaneous visibility.

JR-021 — jRay never injects into index.html on disk

File Transformation is a hard requirement, not a preference. There is no on-disk patching fallback.

The prohibition is on injection, not on writing: JR-022's migration must write to the file in order to remove a legacy patch. Stating it as "never writes" would put the two requirements in contradiction, and the static check would have to be disabled to let the migration through — so the check is that no code path adds the script tag.

At startup jRay looks for the File Transformation assembly via AssemblyLoadContext and, if present, calls Jellyfin.Plugin.FileTransformation.PluginInterface.RegisterTransformation by reflection — no compile-time dependency, so jRay loads normally when it is absent. The payload:

{
  "id": "2c9b5a41-6ad0-4c1e-9f7d-1d1e6b0d5a90",
  "fileNamePattern": "index.html",
  "callbackAssembly": "<jRay assembly full name>",
  "callbackClass": "Jellyfin.Plugin.JRay.Services.FileTransformationRegistration",
  "callbackMethod": "TransformIndexHtml"
}

File Transformation matches callbackAssembly against Assembly.FullName exactly, so the full display name is sent. The callback is a public static method taking a payload with a contents string and returning the transformed string; the payload binds with Newtonsoft, which matches property names case-insensitively. Registration is unconditional at startup — the callback itself checks the "enable overlay" setting per request, so toggling takes effect without re-registering.

Writing to index.html is rejected because it is destructive in ways a plugin cannot clean up after:

• It outlives the plugin. Uninstalling jRay leaves the patch in a file jRay no longer owns.
• It breaks on upgrade. A web-client update replaces the file, discarding the patch — or preserves one pointing at an endpoint that has since changed.
• It collides. Another plugin patching the same file races with jRay, and the loser's edit is lost with no diagnostic.
• It is a second code path, and it is the one nobody runs, so it is the one that rots.

Current: satisfied. WebClientPatchService is removal-only — the injection capability is deleted, not switched off, since dead code with a live signature is what a later refactor re-enables by accident. Enforced by scripts/checks/no-index-injection.sh, which was verified to fail on a reintroduced injection rather than merely to pass today. README.md no longer advertises a fallback. Gap: none.

JR-022 — Migrate away from earlier on-disk patches

Users upgrading from a version that patched the file must not be left with a stale injection. On startup jRay removes any on-disk patch bearing its own <!-- jray-overlay --> marker — unambiguous, and touching nothing another plugin added.

Current: implemented — WebClientPatchService.RemoveLegacyPatch runs at every startup and is a no-op once the marker is gone. The strip itself is factored out as RemoveInjection so it is unit-testable without a filesystem. Gap: no test executes it yet, so this stays In Progress rather than Done — there is no test project in this repo.

JR-023 — A hard dependency Jellyfin cannot resolve

Jellyfin has no plugin dependency mechanism. A manifest cannot declare that another plugin is required, and nothing will install one. File Transformation documents only an end-user repository URL and a reflection integration for plugin authors; there is no NuGet-style dependency to take.

jRay must not bundle the assembly. A bundled copy would sit in a different AssemblyLoadContext from the real one — precisely the failure the reflection integration exists to avoid — on top of licensing and version skew. The dependency is satisfied by the user installing the real plugin.

So:

1. Detect at startup and say so — log a warning naming the plugin and its install URL, and disable only the overlay. Every other feature works.
2. Surface it where it can be acted on — the configuration page shows dependency status: satisfied, or missing with the manifest URL https://www.iamparadox.dev/jellyfin/plugins/manifest.json and a one-line instruction. A warning only in the server log is one nobody reads.
3. State it as a prerequisite in install docs, before the jRay install step.

Optionally, publish jRay through a repository manifest that also lists File Transformation, so one repository URL surfaces both. This is not a dependency mechanism; it removes a step and the chance of installing the wrong thing.

Current: all three implemented, and the detection half is tested (UT-012…015). Startup detection and registration, a warning naming the plugin and its install URL, and a status banner on the configuration page fed by GET /Plugins/JRay/Status/Dependencies. The README states the dependency before the install step rather than after it.

Both failure-path log messages previously said the overlay was "falling back to patching index.html on disk" — a claim JR-021 made false, and the worst place to leave one: an admin reading it while debugging a missing overlay would go hunting for a patch that no longer exists. UT-013 asserts the message names the install URL and does not claim a fallback.

Gap: the config-page banner is T4, verifiable only against a live server.

JR-024 — Names render as text, never markup

Every string that reaches the overlay — actor names above all — is rendered as text. With §6 the source of those strings may be a third-party server, and the server spec §5a names this the single most important client-side control, because it holds even when every other check is bypassed.

It is stated here as a plugin requirement because the server cannot enforce it and the DOM is jRay's.

Current: satisfied. Web/jray-overlay.js uses textContent throughout — no innerHTML, no insertAdjacentHTML. Gap: nothing behavioural. It holds today by construction rather than by rule, which is what the static check exists to keep true once §6 makes remote strings reachable.

---

6. Manifest exchange client — JR-025 … JR-037

The wire format, tiers and server behaviour are specified in ../JRay-public-server/SPEC.md. This section owns the client half, which previously lived in that document's §9 — an inversion, since those are obligations on this repo.

A Jmanifest is this truth file plus a portable identity block and a cut fingerprint, minus the installation-local fields.

JR-025, JR-026, JR-037 — Server list and resolution

The plugin queries a configured ordered list, not a single URL. Per server: Url, Name, Token (contribution only), Enabled, AllowContribute, TrustLevel (Full / FetchOnly). A community entry ships pre-configured but disabled.

First acceptable wins — servers are tried in order, and the first result clearing the configured tier is taken. Order is the user's trust ranking, made explicit. Best-match-across-all would multiply egress and leak the library to more parties for a gain the ordering already expresses.

For a series, first-match applies per episode (JR-026): fetch the bundle from server 1, then query server 2 only for what is still missing. Series are commonly split across sources, and this is where multiple servers earn their keep.

Failure isolation (JR-037): an unreachable or failing server is skipped after a short timeout (5 s connect, 30 s read) and marked failed with exponential backoff. One dead server must never stall a library sweep; failures surface per-server in the config page.

Current: ManifestServer and PluginConfiguration model all of this. Gap: nothing consumes them — there is no HTTP client.

JR-027 … JR-029 — Every server is untrusted

Everything in the server spec §5a is a property of a correctly operated server. Pointing the plugin at an arbitrary URL inherits none of it. jRay therefore re-applies client-side what a server applies on upload, including for the default server:

• JR-027 — validate on receipt. Downloaded manifests go through the same strict schema as uploads: unknown fields rejected, sizes capped, and windows bounds-checked against the item's real runtime. A manifest is never trusted because a server served it.
• JR-028 — size caps enforced while streaming, so an unbounded body is aborted rather than buffered. 2 MiB single manifest, 25 MiB bundle.
• JR-029 — HTTPS required for non-loopback servers, with certificate validation never disabled. A plaintext server would let any intermediary rewrite actor overlays.
• TrustLevel: FetchOnly — the default for user-added servers — accepts manifests but never contributes and never sends inventory beyond the single item queried.

The honest framing for the config page: adding a third-party server means trusting its operator not to serve you deliberately wrong actor data. The controls above bound the damage to bad overlay content; they cannot make wrong data right.

JR-030 — Offsets are applied before storage

When a match carries a non-zero offset (the audio tier — §8), the plugin must add it to every scene window before storing.

The stored truth file is always in the local file's own timebase. This is what keeps the offset out of the read path entirely: Timeline, jray?t= and the overlay never learn that an offset existed. An offset applied at read time would have to be applied identically in three places and would be wrong in the fourth.

JR-031, JR-032 — Endpoints

Mirroring the existing Truth and Tasks controllers:

Route	Purpose
POST /Plugins/JRay/Items/{itemId}/Fetch	Resolve across servers in order; on a match at or above the configured tier, apply JR-030 and store via the managed store
POST /Plugins/JRay/Series/{seriesId}/Fetch	Bundle fetch with per-episode gap-filling
GET /Plugins/JRay/Servers/Status	Per-server reachability and last error, for the config page
POST /Plugins/JRay/Items/{itemId}/Identify	Compute the audio signature and search by content, for items of unknown providence

JR-032: Identify never stores automatically. It returns candidate titles with scores and offsets; storing one is a separate confirmation step. Content identification is a guess about what a file is, and a wrong guess silently attaches another film's cast to it.

JR-033 — Scheduled sweep

A scheduled task walks items with no truth data and attempts a fetch, reusing the Tasks/Pending backlog logic — including its policy rules — and the batch exists endpoint, so a sweep is a handful of requests per server rather than one per item.

JR-034, JR-035 — Contribution

On a PUT .../Truth from a local worker, if contribution is enabled: strip movie and jellyfin_id, attach identity from the item's ProviderIds and its measured runtime, and POST to each contribute-enabled server. For a series, batch into one bundle upload rather than per-episode posts.

Stripping is a requirement, not hygiene. movie leaks the contributor's directory layout and jellyfin_id is a GUID from their database — meaningless elsewhere and mildly identifying. The server rejects both, but the plugin must not send them in the first place.

Contribution is never fanned out. A manifest goes only to servers with AllowContribute set, each an explicit choice.

JR-035: uploads set Expect: 100-continue, so a server rejecting on size or auth does so before the body is transmitted. This matters most for bundles, where a rejected upload would otherwise push tens of MiB pointlessly.

JR-036 — Match tier is the user's dial

The configured minimum tier (audio / runtime / loose) gates what may be stored. A loose match — runtimes within ±30 s — is plausibly a different trim of the same cut, so it is surfaced as a caveat in the UI, not applied silently. Per JR-010 the tier is recorded with the stored truth, which is what makes surfacing it possible after the fetch has finished.

There is no exact tier here, and the plugin sends no video_hash. The server spec §3 defines exact as an equal OpenSubtitles file hash, and it is the strongest technical signal available — it identifies a specific file, so it cannot produce a false positive. That is exactly why it is withdrawn.

A TMDB id discloses "some copy of this film", which is what a library catalogue discloses. A file hash discloses this exact release, which turns a catalogue lookup into a release-identification service and turns a server's database into a mapping from file fingerprints to the instances holding them. That is a far more specific disclosure than PR-005 permits, and a dataset no volunteer operator should be asked to hold.

The audio signature is the deliberate replacement: derived from content, it identifies the cut rather than the copy, so two different encodes of the same edit agree. It answers the question the exchange needs — "do these timings apply to this media?" — without answering the one it must not. audio is therefore the top tier.

A server may still hold hashes contributed by other clients; this plugin simply never participates, and MatchTier has no Exact member so no code path can come to depend on one.

Current: MinimumMatchTier exists in configuration, defaulting to runtime, and ManifestExchangeClient rejects a below-tier match. Gap: the caveat is returned by the fetch endpoint but not yet displayed in the overlay.

---

7. Egress and privacy — JR-038 … JR-041

Contribution reveals to a server operator that some instance holds a given title. Fetching reveals the same. That is inherent to the exchange — which is why the requirements here bound it rather than claim to remove it.

• JR-038 — opt-in, off by default. Manifest sharing, contribution and audio signatures are three separate switches, all default off, and the pre-configured community server ships disabled. No traffic leaves an installation until an admin acts.
• JR-039 — no library-wide inventory in one request. The batch exists endpoint is capped at 100 items and sweeps are paced. A single request enumerating a library is a fingerprint of it, which is the thing PR-005 exists to prevent.
• JR-040 — the config page says plainly that each configured server multiplies the exposure. First-match resolution limits it — later servers are queried only for what earlier ones lacked — and that is worth stating too.
• JR-041 — the plugin never touches gallery data. No reference faces, no embeddings, fetched or stored or transmitted. There is no such code path and there must not be one (SR-005). Verified by static check, mirroring the server's UR-012.

Current: JR-038 holds — every switch defaults off. JR-041 holds vacuously, there being no gallery code. Gap: JR-039 and JR-040 are unimplemented, alongside the exchange client itself.

---

8. Audio signature — JR-042 … JR-045

A content-derived fingerprint from the centre of a media file, used to identify a file of unknown providence and to recover the time offset between differently trimmed releases of the same cut. Construction is specified in ../JRay-public-server/SPEC.md §3 and must be implemented exactly:

1. Decode a 120 s window centred on the midpoint (runtime/2 ± 60 s) — avoiding logos and cold opens at the head, credits at the tail.
2. Downmix to mono, resample to 11025 Hz.
3. STFT: 4096-sample frame, 1024-sample hop (~93 ms, ~1290 frames), Hann window.
4. Log-magnitude spectrum over 300–3000 Hz.
5. 32 logarithmically spaced bins; record peak-bin index plus a 2-bit energy class.
6. One byte per frame → ~1290-byte array, base64-encoded.

JR-042 — no new dependency. FFmpeg performs decode, downmix and resample, using the binary Jellyfin already ships, reached via IMediaEncoder.EncoderPath from MediaBrowser.Controller.MediaEncoding. The plugin implements only a small fixed FFT and bin-peak extraction.

JR-043 — bit-exactness is verified, not assumed. The pipeline computes this signature too (extraction IR-004), deliberately: files never processed locally still get one from the plugin. Two independent implementations of one fingerprint are only useful if they agree exactly, so a golden-vector fixture is shared between the two repos — a short WAV and its expected signature, committed in both. It is CPU-only DSP, which is why this cross-repo check can be a binding CI test rather than an aspiration. Extraction's counterpart is IR-005.

JR-044 — media shorter than 120 s. The window underflows, so no signature is emitted and no sync offset is applied. Such items fall back to the runtime tier, which is adequate: a 90-second extra is not content whose cut alignment matters. Both producers must apply the identical rule, or they diverge on exactly the short items most likely to be misidentified. Extraction's counterpart is IR-007.

JR-045 — the signature carries its own v1: prefix, separate from schema_version. Emit and honour it, so a future change to the DSP chain is detectable rather than silently producing non-matching signatures. Extraction's counterpart is IR-008.

Matching — sliding ±600 frames (≈±56 s), scoring the fraction of overlapping frames whose peak bin matches — is a consumer concern and belongs to this plugin. Offsets are applied client-side per JR-030; manifests are never rewritten.

Current: JR-042 and JR-043 hold. AudioSignature implements the construction above — band table, periodic Hann, radix-2 FFT, band-mean peak, energy class, packing — and AudioSignatureService decodes the centre window by running the FFmpeg binary IMediaEncoder.EncoderPath names, so the plugin gained no dependency. fixtures/audio/ holds the extraction repo's three golden-vector files byte-identically, and the computed signature equals the recorded vector exactly (UT-038 … UT-044). The binding check regenerates the fixture PCM from make_fixture.py's arithmetic and verifies it against the recorded decode checksums, so it runs on a CI host with no codec at all; the two tests that exercise the real FFmpeg decode self-skip without a binary.

JR-044 and JR-045 hold as well, both halves. AudioSignatureMatcher is the reader they were waiting on: it implements the ±600-frame slide above, scoring the fraction of overlapping frames whose peak band agrees, and returns the tier and the offset. TryParseFrames refuses any prefix but v1:, so a future producer's v2: drops the item to the runtime tier rather than being scored as if it were understood; and a runtime under 120 s yields no match and therefore no offset, read off the runtime rather than inferred from a missing string, because the runtime is what both producers test (UT-045 … UT-052).

The offset has two terms, which is easy to miss: the recovered slide, and the difference between where the two windows are anchored. Both windows are centred on their own file's midpoint, so unequal runtimes start them at different absolute times. A release carrying 40 s of extra head material recovers 20 s from each term.

One parameter is not from the specification: an alignment must overlap by at least 64 frames (~6 s) before its score counts. Without a floor the extreme offsets compare a handful of frames, where a chance agreement scores 1.0 and beats the true alignment. It never binds on the real case — two full-length signatures still overlap by 688 frames at the widest offset.

Gap: the signature is computed and can now be read, but nothing yet calls either: ComputeAudioSignatures still gates nothing, no fetch path attaches a signature or consults the matcher, and no contribution carries one. That wiring belongs to the fetch and contribute paths (JR-031, JR-034), not here.

---

9. Human-in-the-loop association — JR-046

Proposed. See system spec §4, which owns the design.

The pipeline produces unidentified tracks — a face that is genuinely someone, sustained across many frames, that the gallery cannot name. A user watching the film usually knows exactly who it is. jRay's contribution is the review UI: show a cluster's context crops, let the user pick from the title's cast or search TMDB, and record the association for extraction to ingest into the local gallery.

The unit of review is a person, not a track. Unknown tracks are clustered upstream (AR-021), so the question is "who is this person, who appears in these twelve places?" rather than twelve disconnected questions. One answer resolves the cluster.

Deliberately left as a single TBD row rather than decomposed. It depends on extraction AR-021/AR-022 landing, and on system open question 2 — whether unidentified presence is published in the truth file at all, which determines whether this UI's work queue arrives with the truth data or needs a separate channel. Decomposing now would fix an interface against an undecided upstream.

---

10. Client: pushing results from a remote worker

Reference material for worker authors; the requirements are JR-009 and JR-015.

1. Authenticate. Create an Administrator API key (Dashboard → API Keys) and send it as X-Emby-Token: <key> or Authorization: MediaBrowser Token="<key>".

2. Find work. Poll GET /Plugins/JRay/Tasks/Pending?limit=10 rather than walking the library and checking each item.

3. Resolve the item id. The push endpoint is keyed by Jellyfin item GUID, not path:

GET /Items?Recursive=true&Fields=Path&IncludeItemTypes=Movie,Episode

Match Path against the file you processed — which requires the worker to see the file at the same path Jellyfin does; translate first if it mounts the library elsewhere. The mapping is stable until the file moves, so cache path -> itemId and re-resolve only on a miss.

4. Push. PUT /Plugins/JRay/Items/{itemId}/Truth with the truth file body. 204 stored (cache invalidated immediately), 400 unsupported schema_version, 401/403 key missing or not an administrator. The PUT is idempotent, so retrying on a network error is safe.

5. Optionally remove. DELETE /Plugins/JRay/Items/{itemId}/Truth always returns 204; the item falls back to its sidecar on the next read.

---

11. Open questions

1. sample_fps, gallery_size and pipeline_version move under extraction.* in JR-002. This aligns the truth file with the Jmanifest's block of the same name, and the bump is breaking regardless. It is a change this spec proposes rather than one inherited from SR-003's list — confirm, or keep them top-level.
2. Does route belong in the jray?t= envelope? JR-013 says the response is extensible and JR-005 says presentation must not over-claim. Exposing belief and route would let the overlay caveat a weak claim, but invites a UI that shows a number to a viewer who cannot act on it.
3. System open question 2 — unidentified presence. If published, the overlay could show "unidentified person" and JR-046 gets its queue from the truth file directly. jRay is the consumer that would have to display it, so this repo has a position to state.
4. Test-ID namespacing. UT-nnn/IT-nnn are per-component registers, so UT-001 will exist in both this repo and scene-actor-extraction. Fine while the gate runs per repo; ambiguous the moment a rollup spans them.