SCI format

SCI is Weft's ASCII video archive format. Files use .sci and MIME type application/vnd.weft.sci. One visual representation is supported, with synchronized Opus audio and explicit presentation timing. The format contains ASCII symbols and appearance attributes, never a conventional video stream or source pixel planes.

The internal container version is 4. It is a compatibility field, not a filename extension or viewer setting. Older containers, two-color visual profiles and experimental scene-v2/v3 profiles are rejected. Re-encode their retained source media with the current encoder.

Header and metadata

Integers are unsigned and little endian. The fixed header occupies 32 bytes.

Metadata begins at byte 32 and occupies between 2 and 4096 UTF-8 bytes. The writer reserves 4096 bytes; readers follow offsets and ignore padding. Index offset is at most 65536. Archive size must equal the header declaration and cannot exceed 64 GiB.

Seek index

Each little-endian record occupies 48 bytes. Payload ranges are ordered, disjoint and within the archive. The final payload ends at the declared archive size. A reserved gap before payloads is allowed.

Presentation timing

Metadata declares chunks, frames, duration, nominal rational fpsNumerator and fpsDenominator, fps, timebase: 1000000, chunkSeconds: 1, video, audio, and qualities. FPS describes nominal source cadence, not a promise that variable-rate input is constant. Duration comes from presentation timestamps. Corresponding rendition/audio chunks must have identical start/end times and first frame numbers. Each rendition covers the entire timeline without gaps or overlap. Audio sample counts equal floor(end * 48000 / 1000000) - floor(start * 48000 / 1000000).

A video packet starts with its unsigned 32-bit presentation duration in microseconds, followed by a nine-byte packet header (tag, uncompressed length, compressed length) and one zlib stream. Durations are positive and at most one second. Long source holds split into identical presentation events without introducing new motion. Metadata records sourceFrames and averageFps separately from presentation event counts. Delivery chunks span at most two seconds, including a trailing hold event. Their sum must equal the indexed chunk duration. References reset at each chunk. Source timestamps normalize to the first video frame and are retained to microsecond precision. Sources above the configured frame-rate ceiling drop frames using source timestamps, rather than generating new motion.

The only visual profile is sci-ascii-scene-v1, mapper adaptive-layered-v2, font ascii-3x4-v2, transform reflect-xy-v1, packet format ascii-scene-zlib, and cellBytes: 28. Packet tags are 4 (independent commands), 5 (predicted commands), 6 (spatial attribute planes), and 7 (motion-predicted attribute planes). Tags 5 and 7 require a previous frame. All other tags and codec profiles are rejected.

The writer stages independently compressed ASCII frames before delivery packing. Packing reconstructs those compositions and selects temporal packets, then splits every rendition at a common boundary whenever any rendition would exceed 16 MiB. New boundary packets are independent. This bounds delivery payloads without changing fitted glyphs, colors, resolution, or source timing. Chunks may therefore be shorter than the one-second target. Temporary ASCII staging is limited to 64 GiB and is removed on completion or failure. Reserved index space may leave an unreferenced gap before payloads; viewers do not fetch that gap.

The writer can use a smaller preferred delivery budget, recorded as informational deliveryChunkBytes. It avoids preference-driven splitting until a group spans at least 20 ms; a hard-limit split takes precedence. One complete frame or that minimum group can exceed the preferred budget. References still reset at every actual boundary and audio retains the same indexed time range.

Adaptive glyph representation

The normative glyph table defines all 95 printable ASCII symbols, codes 32 through 126. Each symbol has a fixed 12-bit mask. Bit row * 3 + column selects foreground when set. For each reflected glyph sample, use column 2 - x when reflection bit 0 is set and row 3 - y when bit 1 is set. Scaling repeats this fixed mask; no frame can transmit a replacement mask. Changes to glyph masks require a distinct font identifier.

A composition partitions the raster into disjoint square groups of base cells. A base cell covers 3 by 4 pixels. A leaf covers 1, 2, 4, or 8 base cells on each axis, so its scale is 1, 2, 4, or 8. It paints an explicit ASCII space with a background RGB color, followed by zero to four printable ASCII foreground glyphs in order. Foreground layers paint only set bits of their fixed glyph masks; unset bits preserve the previous layer. Reflection bit 0 flips glyph columns and bit 1 flips glyph rows; other bits must be zero. Scaling uses nearest-neighbor replication of the fixed 3 by 4 mask. Colors are unsigned eight-bit RGB. No arbitrary source masks or pixel residuals are permitted.

The bounded expanded decoder state uses 28 bytes per base cell:

Unused slots are zero. Every base cell in a leaf repeats the same scale, background, and layers, with its own offsets. Leaves must fit completely inside the frame. Readers reject overlapping, incomplete, inconsistent, or out-of-bounds leaves.

Optional encoder color quantization selects RGB attributes from a uniform per-channel palette with 2^bits levels for 4 through 8 bits. For an input channel value v, compute i = floor((v * (2^bits - 1) + 127) / 255), then floor((i * 255 + floor((2^bits - 1) / 2)) / (2^bits - 1)). Quantization applies to averaged fitting samples and final fitted colors. Layers that no longer change reconstruction may be removed. The optional informational video.colorBits records preparation precision. Stored attributes remain eight-bit RGB and decoding rules are unchanged. Eight-bit preparation is the default and preserves prior full-precision output.

Palette and composition commands

Inflated command packets (tags 4 and 5) begin with an unsigned byte palette count, 0 through 255, followed by that many RGB triples. A color reference is one byte: 0 through 254 indexes the palette; 255 is followed by a literal RGB triple. Invalid palette indexes fail.

Commands fill the next unfilled base cell in row-major order:

• Command 0: scale exponent, ASCII space, background color reference, layer count, then each layer's ASCII code, reflection flags, and color reference. The reconstructed leaf fills its complete square region.
• Command 1, permitted only in tag 5: unsigned 32-bit base-cell index of a leaf origin in the previous reconstructed frame. Copy that complete glyph composition into the current position. Moving a composition changes its placement, not its symbols or masks.
• Command 2, permitted only in tag 5: positive unsigned 16-bit run count and unsigned 32-bit source base-cell index. Copy that many consecutive scale-one leaves from the previous frame. The destination must also be consecutive unfilled cells.

Commands must fill the entire frame exactly, with no trailing bytes. Previous state is immutable while constructing the next frame. Every packet is independently inflated with an exact length check and rejection of trailing compressed data. Inflation is limited to baseCellCount * 38 + 1024 bytes. Compressed payloads remain limited to 16 MiB. No reference crosses a delivery chunk.

Attribute planes

Tag 6 contains exactly baseCellCount * 28 bytes, arranged in 28 attribute planes. Each byte predicts from the left, above, and upper-left value using the Paeth predictor. Missing neighbors are zero. Subtract the predictor modulo 256 when encoding; add it modulo 256 when decoding. Paeth selects the nearest neighbor to left + above - corner, with ties preferring left, then above, then corner.

Tag 7 starts with signed little-endian 16-bit horizontal and vertical motion offsets in base-cell units, each between -64 and 64. The remaining bytes are 28 attribute planes. Each reconstructed destination attribute predicts from the previous frame at (x - dx, y - dy), or zero outside the frame. Attribute subtraction/addition is modulo 256. The complete reconstructed composition geometry and ASCII symbols must validate before rendering. These residuals describe glyph attributes, never arbitrary pixel samples.

The encoder estimates a global cell motion offset from repeated complete glyph compositions. It compares zero-motion prediction, the estimated motion prediction, spatial planes, and composition commands using their actual compressed sizes. Every choice reconstructs exactly the same ASCII composition.

Audio

Metadata declares audio: null when no audio is present. Otherwise it declares codec: "opus", sampleRate: 48000 and channels: 2. Each audio index record stores one complete independent Ogg Opus stream for its presentation interval. Audio payloads have no visual packet headers or duration prefixes.

Ogg pages use version zero, one stream serial, consecutive page sequence numbers beginning at zero, correct page checksums and consistent continuation flags. The first page marks beginning of stream and the final page marks end of stream. The first packet is a 19-byte OpusHead, version 1, stereo, channel mapping family 0; the next is OpusTags. Packet assembly is bounded to 65536 bytes. Each Opus data packet represents at most 5760 samples per channel (120 ms).

Decode at 48 kHz, apply the Opus header's pre-skip, and trim trailing samples using the final Ogg granule position. The final granule must equal the indexed sample count plus pre-skip. The decoded interval must supply exactly the indexed number of samples per channel. Audio chunks schedule at their indexed presentation start, on the same clock as visual events. Seeking begins with the independent audio chunk and skips the portion before the requested time. These audio codecs serve synchronized sound only; they never carry the visual representation.

Encoder and evidence

The encoder first fits reflected two-color glyphs, then greedily adds foreground glyph layers only when they reduce squared RGB error. It jointly refits each surviving layer color to the optimal rounded mean of its visible source pixels and removes completely occluded foreground layers. This cannot increase patch error. It considers an adaptive tree through scale eight. Exact coarse reconstructions short-circuit finer searches; otherwise child compositions compete with their parent. A byte penalty controls the local distortion versus serialized composition-size objective; zero penalty does not permit increased error from merging. The local size model precedes entropy coding and is not a globally optimal rate-distortion solver. Complete and predicted packets compete using their actual compressed byte lengths. A bounded patch cache reuses repeated source patches.

The default fidelity profile permits four layers and uses zero spatial byte penalty. The requested delivery bitrate is a research target recorded separately from the achieved rate; exceeding it never silently reduces resolution, source cadence, or fidelity settings. Research sweeps report actual compressed sizes and quality before changing production defaults.

Archives are limited to six hours, 100000 index records, 64 GiB total, 4096 pixels wide and 1080 pixels high. Dimensions must be divisible by three horizontally and four vertically. Each visual or audio delivery payload is at most 16 MiB. Supported rendition targets are 360p, 480p, 720p and 1080p, limited by source height.

Readers validate declarations before allocation, enforce exact inflation lengths and reject malformed glyphs, geometry, timing, checksums or unsupported identifiers. A failed decode cannot publish partial reconstructed output. Browser workers run the matching Rust decoder through WebAssembly; missing support fails explicitly.

These bounds and the custom extension do not establish better compression, indistinguishable quality or lower resource use. See research status for evidence and remaining targets.