Changelog#

This file separates current source changes from tagged releases. Version 0.3.5 is the current development version. Historical entries retain the behavior and limits of their named version.

Unreleased#

No changes.

0.3.5#

Memory and persistence#

• Add typed GPU state with configurable object and payload capacity.
• Prepare text query vectors on the GPU and select exact scoped memory references.
• Retain source text and qualify two-stage semantic interpretation for exact model and processor identities.
• Preserve complete source spans, correction targets, actor labels and recorded context choices.
• Add source-backed appraisal with separate benefit, harm, regard and task trust evidence.
• Store supported task outcomes as exact evidence reviews for later matching tasks.
• Add explicit cold-payload offload and fetch with dependency and durability checks.
• Package models, runtime assets, scope state and recovery history in complete CCIR files.
• Preserve compatible creator-policy state through explicit versioned updates.

Service and controls#

• Add a separate HTTP gateway for standard completions and native request resources.
• Add scoped shared conversations, exact retry receipts, cancellation and saved-result state.
• Report model capabilities and qualified numerical controls independently of graphical clients.
• Bind fitted controls to exact models, turn formats, assets and acceptance evidence.
• Preserve affect state by private conversation or shared space through recovery.
• Add revision-bound policy controls through local commands and the native API.

Models and media#

• Add native JPEG and RGB8 input through the supported vision model pair.
• Add native WAV and PCM input through the supported audio model pair.
• Preserve media ownership, model routing, source bytes and copied-file recovery.
• Expose independent source, feature, worker and KV capacity settings.

Operation and integration#

• Complete refused prompt admission without discarding prior conversation state.
• Add instance tool defaults and per-conversation tool choices.
• Compile profiles with 32, 64, 128 and 256 agent slots.
• Embed required clock and Q8 decoder PTX modules in the executable.
• Report essential child failure during operation and final shutdown.
• Bound failure cleanup and preserve optional terminal behavior.
• Keep cold-read temporary state within the cancellation block for optimized builds.
• Reorganize the manuals around setup, operation, API contracts and file formats.

Current limits#

The default typed store permits 8192 object versions and 16 MiB of payload. Text query inputs permit 2048 UTF-8 bytes. Recall permits 16 references in 4096 rendered bytes.

Large-card profiles compile; full-capacity hardware execution remains unqualified. A runtime uses one GPU. The affect and quality file writers retain a 64-slot limit. Semantic memory, appraisal and numerical controls retain their documented exact-model qualification boundaries.

The numerical reference comparison remains unresolved. See support and qualification for the complete current boundaries.

0.3.0#

Release#

• Set version 0.3.0.
• Keep the two reranker class logits in registers without changing the score calculation.
• Report process-check failures even when an exception has no message.
• Require the expected illegal-address error in memory fault checks.
• Wait for the current boot's durable restore metadata before checking its journal.
• Keep ASCII window text ahead of Enter when an X11 input method is active.
• Check all local model files and permit each family's pre-tokenizer metadata.
• Correct model download commands and record verified file sources.
• State journal timing and quantized decay limits. Identify earlier benchmark versions.
• Correct the readout normalization formula and timer limits.
• Keep the saved instance name and device at startup.
• Preserve child startup output in the journal boot log and report its path on failure.
• Keep the New conversation control balanced when a persona import starts.
• Bound directory watches in the restore check without reducing its input count.
• Add fixed-input accuracy capture, reference validation, and complete comparison counts.
• State reranker vocabulary requirements and measured limits in model tool use.

Known limits#

• Some checked models make unrelated tool calls or omit the requested call.
• Per-instance and per-conversation tool selection controls are not included.
• A long persona conversation can repeat a prompt-capacity error without completing the input.
• The empirical numerical comparison remains incomplete as an accuracy qualification; see docs/17-accuracy.md.

Layer and state tables#

• Bind, select, and capture layers through type rows (d4746e4, 08e3df9, 499d7e6).
• Derive state allocation and cache pages from layer types (3e0bd21, 1cacc21, 3275948).
• Check mixed layer selection, compact cache addresses, and fresh descriptor placement (6db6f20, bb156da, 74c4876).
• Read tokenizer families from model metadata (03e742c).
• Capture attention without query and key head normalization (4676769).
• Read bounded model turn wraps from the store (a577091).

Control and operation#

• Show custom local model files and keep file presence separate from active roles (193ad6e).
• Inspect model headers asynchronously with bounded time and output. Clear results when the model file or executable changes (193ad6e).
• Keep the exact selected language model and available companion roles through the wizard (193ad6e).
• Add Restore controls and a larger default conversation window (193ad6e).
• Add a model file guide with measured hybrid, expert, and memory examples (193ad6e).
• Correct replay timeout and failure counts, and add host controls (193ad6e).

Tool history and restore#

• Keep native call arguments and render stored tool results as separate turns (01b8b2d).
• Use the selected vocabulary for memory embeddings and refuse invalid note provenance (01b8b2d).
• Refuse rejected token replay while permitting a reproduced sequence-open refusal (01b8b2d).

Tool call forms#

• Select bounded tool call forms from complete model template identities (38e0515).
• Use one selected row for parsing, tool instructions, call history, and result turns.
• Support tagged JSON, native bare JSON, and function-parameter forms without template execution.
• Refuse unsupported forms without disabling text conversation.
• Report invalid memory provenance to the model without saving a note.
• Retain every call argument and render stored results as separate tool turns.
• Check native forms, metadata lifetime, and prompt bounds at one slot and all slots.

Tool completion#

• Require a ready embedding pass before memory tools appear in a prompt (ba46a18).
• Refuse unavailable memory calls at once and state the required model role.
• Use a separate embedding vocabulary when it differs from the language vocabulary.
• Record each device tool result, including empty results and errors, without duplicate host replies.
• Bound automatic continuation and tool turns by one input budget and report exhaustion on the console.
• Check tool completion, vocabulary selection and turn bounds at one slot and all slots.
• Add a real attach-socket conversation driver that retains replies, tool results and elapsed times.

Model file checks#

• Read Q4_1, Q5_0, Q5_1, Q2_K and Q3_K weights in each matrix and flat reader (ab3ff3f).

• Read Q4_K, Q5_K and Q6_K super blocks (0c5bcbe).

• Check each packed reader against a double reference at batches of one and 64.

• Add required-file checks that compare each tensor element and refuse missing input.

• Select rotary pairs from a shared family table and refuse unknown families.

• Add aotx_models inspect for local files and bounded HTTP header reads.

• Share compiled tensor and tokenizer tables with the header report.

• Bound header allocations and refuse duplicate metadata keys and tensor names.

• Run the real wrap check in the architecture test.

• Name the cache rebuild check correctly and exclude untested results from its tally.

• Add batched expert selection, selected matrix products, and weighted output sums.

• Add full-width query and key norms for the expert layer.

• Supply a file-specific expert model manifest with explicit turn wraps.

• Compare exact tensor names when layer types share descriptor slots.

• Check distinct batches through the decode graph and accept an explicit process-check driver.

• Refuse incompatible expert routing rules, tensor dimensions and tensor types before binding.

• Keep all-slot prefill checks within the token capacity on each profile.

• Add a separate attention bias layer capture and bias-before-rotation kernel.

• Bind complete tensor suffixes and require correctly shaped F32 query, key and value biases.

• Preserve explicit head widths and derive an absent width for the bias layer.

• Supply a pinned Q8_0 model entry with explicit conversation spans.

• Check bias placement, grouped heads, file refusals and token bounds across sequence slots.

• Give each layer row its tensor slot meanings and derive descriptor capacity from row spans.

• Bind layer offsets without pointer arithmetic across descriptor members.

• Check mixed tensor types, model isolation, absent slots and descriptor bounds across layer batches.

• Refuse tensor rows that exceed the inspector name-mask capacity.

• Add fixed F32 recurrent matrices and convolution history outside the page pool.

• Add batched delta layers and gated paged attention with partial rotary turns.

• Rebuild fixed state through prompt replay and check every state byte across chunk partitions.

• Add a Unicode-mark tokenizer pattern and a verified split-pair family row.

• Check mixed layer selection, fixed allocation sizes, gates, carried state and slot isolation.

• Refuse recurrent sequences with slots outside the table and check the fault counts.

0.2.8#

Release#

• Set version 0.2.8 (the release commit).

The pair mode and the fixture tool#

• Add the pair mode of the quality score tool (90c62df).
• Add the fixture tool and the pair script (51fae80). The console line outcome and the fixture conversations.
• Extend the checks of the score and steer tools (b3743be).
• Make the armed result stand in for the call of a turn (2f054ad).
• Match the device template and tighten the checks (1ed271a).

The completed request, the cosine readout and the log-odds score#

• Make an armed tool result a completed request (78bf1b3).
• Read the probes as the cosine of the row (ed4e381).
• Score the pair judge by the log-odds in nats (ce1a1bb).
• Wait a fixed count of ticks for the quality rows (8037e69).
• Arm the ok result on a line with no tool field (d8a3b2c) and add the call arm of the outcome line (b5215f6).

The dials and the spoken voice coupling#

• Add affect dials and figures (bb0e2d2).
• State unavailable dial figures (a07bf76).
• Couple agent state to spoken voice (600655a).

The actuator figures#

• Carry three actuator figures in the affect trace (a18c60e). The budget spent, the entropy shift and the class frequency shift of a turn. The record grows to 92 bytes in a body of 192 and stays derived, so a build without the option is unchanged.
• Show the actuator figures beside their dials (cec8ba2).

The argument line and the profile checks#

• Complete a call over the room of the argument line (e731f73). A call the parser refuses on its bound ends with a tool error result, and the turn after it writes the reply.
• Make the profile checks read the profile (da10534). The agent table checks and the pass count of a text set read the slots of the build. The probe checks read its language role.

Documentation#

• Document the state, the actuators and the score tool (b670fbd). The text of version 0.2.7, which its tag does not hold.
• Document the quality instrument and the dials (6908222). The quality page, the affect page, the control page and the measured figures.
• Add the changelog section of version 0.2.8 (428a809).
• State what the affect option leaves out (94e385b).
• Set the zero point of the readout on plain replies (339353d). The measured axes and the findings of the measurement.
• Replace the conduct samples with current output (0b99d48).
• Requote the stream samples from one measured turn (f449d3f).

0.2.7#

Release#

• Set version 0.2.7 (the release commit).

The affect state and its restore#

• Add the affect state, its update law and its record (262ad1b).
• Derive the affect state line and check the restore (50e6d62).
• Keep the affect law with its open sequence (4e35d20).

The calibration#

• Add the dialogue neutral set and three conversations (2c5b704).
• Add the probe layer and the standardization set (eef4c3e).
• Measure the composite vectors in the calibrate mode (c903de0).
• Give the small talk conversation no tool outcome (2eb6ff5).
• Add the task set and the neutral reply set (e684e77).

The actuators#

• Couple affect to sampling and voice bias (53881a3).
• Build the bounded affect steer row (b28feda).

The capability instrument#

• Add the capability score tool and its check (f0efd16).
• Extend the axis check to the composite and the readouts (f39260d).

Documentation#

• Document the affect controller, the calibration and the score tool (b670fbd).

0.2.6#

Release#

• Set version 0.2.6 (the release commit).

The affect substrate#

• Add the optional affect data contract (b712dc4). The build option, the thirteen settings and the three record layouts.
• Add affect and quality drain streams (752540f). The two streams, the derive names, the whole-journal output and the identity check.
• Cut the agents table in the allowance test and keep failed identity runs (738e718).
• Add the affect sums, the probe loader and the read branch (afedda6).
• Add the affect turn node and its check (b282ab2).
• Mark verified tasks, refuse bad probe rows and mask the clock record (12ed10c).

The derivation tool and the calibration#

• Add the affect fixtures (cce6375).
• Add the axis mode of the steer tool (2d1cfe3).
• Add the calibrate mode and its check (e8ff2da).
• Refuse a flat probe scale, take the longer row and rewrite ten pairs (094c49f).

The quality stream#

• Measure quality in the embedding batch (c6633cc).
• Connect quality records to completed turns (ae1a769).
• Load the refusal phrases from the store and gate the page release (fbd5f8a).

The control client#

• Read affect and quality replica lines (a0d287c).
• Show affect and quality trace figures (3213880).
• Keep two instances apart in the control program (e7f6e7d) and build the instance name case without the option (b3dea6d).

Documentation#

• Document the affect substrate and the quality stream (a725a09). The affect page, the conduct page and the control page.
• Document the affect settings and the affect option (68ed95d).

0.2.5#

Release#

• Set version 0.2.5 (the release commit).

Decode controls and telemetry#

• Add agent decode controls (f8f32ef). Class A records set the sampler row of each agent. The token statistics stream, the think budget and the stop command for one agent come with them.
• Add conduct controls and measurement tools (b4fd0c4). The steer vector derivation gives a potency figure. The steer add follows each layer residual. The bias profiles and the page map stream come with them.
• Publish the device ring use (b0fc9f8) and show the ring occupancy and record the rates (a59f934).
• Split the matrix rate fixture (b570fc7).

Control client#

• Make first-run startup and voice reliable (a6fa503).
• Improve conversation and sync controls (1feb216).
• Add model and persona controls (28e4737).
• Repair the control program and document it (8904192).
• Wrap the reply text at the window edge (9b8db35).

Checks and repairs#

• Fix the findings of the review of the quality pass (5fa9c7c).
• Harden the steer tools, the conduct loader and the checks (7097b56).
• Keep one manifest line for each role (685c7bb).
• Name an empty settings file in the replay gate (96ca1dc).

Repository page#

• Add the divider graphic to the README (673d4e9) and update it (056546a).

0.2.0#

Release#

• Set version 0.2.0 and add the release documents and measured figures (6086667).
• Add the repository page (8306cd8).
• Add the terms table to the repository page (608de31).
• Replace the mark on the repository page (30b3bb4).
• Show the banner at full width on the repository page (5495311).
• Fix the release findings and update the verified figures (4b56f51).

Profiles and settings#

• Add the settings records, the card record and the common key list (1c97e00).
• Add the four build profiles and move device table limits into them (e382048).
• Add settings files and their disk-side readers (82b9117).
• Write a changed device setting at the tick commit (15515be).

Modules and tools#

• Add the import and remove record layouts (d63e39b).
• Add the device catalog and the import of modules (9c3a356, 47fbecb, 4c7bc3e).
• Add device and host tool modules, checks, examples and the tool SDK (5afe360, abd1722, fa07d81).
• Join tool nodes to the tick graph and make the tool contracts exact (c87564f, 7ec445a).

Terminal and conversation#

• Add the display mirror and the terminal program (bf61047, ddd5ee7, f2016e9).
• Add long input lines, transcripts and conversation memory (cd47376, 4e4b645, 316a637).

Model store#

• Add the model record, the store program, run-time model load and store checks (30ec4ca, c905dbd).
• Make replacement, fetch and catalog rules exact (d0a5b88).

0.1.0#

• 088c239 Add the repository layout, the gates and the clock module
• 4bba818 Add the seam
• 973c096 Add the surface
• 80a57ba Add the model file and the tokenizer
• 582948d Add the model kernels
• 5d730d1 Add decode through the tick
• 2a26563 Add agents and tools
• 7e70542 Add the documentation set
• b7d1257 Fold the block scale once in the PTX product
• a52bf6f Record the late verdict and pace the replay
• 9ecab9a State the late verdict and the pace of a replay
• 6297929 Hold the replay clock until a later tick is seen
• 50df09b Refuse a second granted line after a restore
• ff6ced1 Add the measured figures of one machine