Latest writing / Tutorials
Durable idempotency: commit identity and result together
Use a framed journal, process crashes and 99 truncated prefixes to explain durable idempotency.
Start readingWriting
Understand the question. Take away a method.
Runnable tutorials, sourced mechanisms and engineering decisions.
59 entries · Runnable tutorials, sourced mechanisms and engineering decisions.
01 / BUILD · Tutorials
Outbox delivery: local durability is not remote exactly-once
Compare keyed and unkeyed receivers to separate lost replies, missing receipts and duplicate remote effects.
02 / BUILD · Tutorials
Conversation recovery: reconcile intent, receipt and whole turn
Halt the JVM at three commit boundaries and rebuild whole-turn history from durable facts.
03 / EXPLORE · Engineering cases
MuJoCo controller tuning: best candidate is not a passing controller
Select on 54 development episodes, then check position error and stable completion on 36 held-out episodes.
04 / EXPLORE · Engineering cases
MuJoCo mass transfer: can the same gains still work?
Compare fixed and mass-scaled gains in 36 episodes, including lower mean error without a velocity-contract pass.
05 / EXPLORE · Engineering cases
MuJoCo observation loss: zero force is not a stop
Compare stale feedback, zero force and fading in 27 episodes to distinguish timeout from physical braking.
06 / UNDERSTAND · Mechanisms
Block KV decoding: less workspace, more time
Separate cache payload, work arrays and observed runtime, retaining 810 timings including slower block decoding.
07 / UNDERSTAND · Mechanisms
Packed KV groups: what does 4-bit actually save?
Store real nibbles and audit group-scale overhead, padding bytes and attention output error.
08 / UNDERSTAND · Mechanisms
KV residual windows: do recent tokens deserve higher precision?
Compare early-focused and recent-focused attention to explain unequal gains from the same float32 window.
09 / BUILD · Tutorials
Java idempotent retries: no reply does not mean no effect
Seven loopback HTTP fault cases explain idempotency keys, concurrent admission, payload conflicts and process-crash gaps.
10 / EXPLORE · Engineering cases
MuJoCo velocity estimation: smoother signals can mean worse control
54 actual simulation rollouts compare finite differences, EMA and an oracle reference across position noise and observation delay.
11 / UNDERSTAND · Mechanisms
KV quantization axes: why Key and Value need separate reasoning
162 archived INT8 comparisons examine tensor, token and K-channel/V-token scales, including cases where the hybrid loses.
12 / BUILD · Engineering cases
Java history and request windows: can a trimmed fact return?
Follow narrowing, commit, restart and expansion to separate durable conversation history from one request context.
13 / BUILD · Engineering cases
Durable Java turns: what survives a process crash?
Checksummed append records, 142 truncated prefixes and three halted JVMs distinguish local recovery from remote exactly-once execution.
14 / BUILD · Engineering cases
Java retries: three timeouts are not one deadline
Seven local HTTP conditions separate retryability, Retry-After, attempt limits and elapsed-time budgets.
15 / EXPLORE · Engineering cases
Delayed MuJoCo feedback: can an old observation predict now?
Twenty-four paired traces compare delay, position noise and constant-velocity prediction, retaining incomplete recoveries.
16 / EXPLORE · Engineering cases
MuJoCo obstacle clearance: a point path is not a body path
Six contact trajectories compare direct, edge and clearance paths while separating geometry from controller tracking.
17 / EXPLORE · Engineering cases
A second MuJoCo scene: why does the actuator overshoot?
Compare mass and PD damping in a planar force-controlled actuator, separating target crossing from stable completion.
18 / UNDERSTAND · Mechanisms
Profiling KV reuse: saved work versus added overhead
Profile 909 queries across signatures, lookup, copying and continuation, including the remaining slowdown.
19 / UNDERSTAND · Mechanisms
KV quantization: how many bytes, and at what error?
Compare tensor and row-wise int8 quantization, including scale bytes and attention error under outliers.
20 / UNDERSTAND · Mechanisms
Tiled attention: preserving results without a full score matrix
Derive online softmax from a failing block-average baseline and separate smaller temporaries from measured speed.
21 / BUILD · Engineering cases
Java LLM practice (10): why is the reader still running after cancellation?
30 local HTTP connections separate conversation invalidation, reader exit and server work, comparing Future interruption with socket closure.
22 / EXPLORE · Engineering cases
Embodied AI practice (14): does landing in the bin mean the task is complete?
Eight MuJoCo trajectories, three passive readouts and recorded-state replay separate historical placement, current evidence and qualified completion.
23 / UNDERSTAND · Mechanisms
LLM mechanics (7): why is caching slower even when it saves computation?
4,860 measured queries compare three admission policies, connecting byte budgets, working sets, computation and local elapsed time.
24 / BUILD · Engineering cases
Java LLM practice (9): why can a cancelled answer overwrite a new conversation?
18 controlled thread schedules test request ownership for streaming previews, completion and errors, with cancellation and bounded deduplication.
25 / EXPLORE · Engineering cases
Embodied AI practice (13): where should recovery resume after interrupted lowering?
18 MuJoCo rollouts and synchronized replay connect phase contracts to recovery budgets, separating false holds, physical placement and controller abort.
26 / UNDERSTAND · Mechanisms
LLM mechanics (6): why can a larger prefix cache still get zero hits?
324 requests across cold-start traces and 648 arrays test byte-bounded prefix checkpoints, working sets, isolation and LRU eviction.
27 / BUILD · Engineering cases
Java LLM IV: which concurrent response may commit?
Reproduce reversed response order with real Java threads, then enforce versioned commits, complete turns and bounded receipts.
28 / Retrieve · Engineering cases
Java LLM practice (7): a real citation can still support a wrong answer
22 frozen configuration cases separate citation existence, revision scope, field support and completeness in a Java answer gate.
29 / Retrieve · Engineering cases
Java LLM VI: validate retrieval evidence before judging RAG answers
Implement BM25 and citation boundaries over ten owned synthetic documents and fourteen frozen queries, retaining misses and false hits.
30 / BUILD · Engineering cases
Java LLM practice (8): should an interrupted stream enter conversation history?
22 loopback HTTP cases separate incremental preview from committing history, testing UTF-8, SSE framing, cancellation and abnormal endings.
31 / BUILD · Engineering cases
Java LLM V: why can a timed-out tool keep running?
Build an allowlisted read-only executor and reproduce the distinction between timeout, cancellation and termination.
32 / EXPLORE · Engineering cases
Embodied AI practice (12): what should the gripper do after recovery is exhausted?
12 MuJoCo rollouts and synchronized three-view replay compare hold, open and open-retreat, separating terminal decisions from physical consequences.
33 / EXPLORE · Engineering cases
Embodied AI practice (11): placement completed, but is it still valid?
Seven MuJoCo rollouts and a paired replay separate historical completion from current validity, including false alarms and expiry boundaries.
34 / EXPLORE · Engineering cases
Embodied AI VIII: what should the gripper do after a drop alarm?
Fifteen fixed-scene rollouts compare hold, open and open-then-retreat after identical alarms.
35 / EXPLORE · Engineering cases
Embodied AI X: why can a recovery budget stop a recoverable task?
Twelve rollouts compare unlimited recovery, a two-resume cap and a 400 ms cooldown, separating bounded behavior from task completion.
36 / EXPLORE · Engineering cases
Embodied AI IX: a release command is not verified placement
Apply single-capture and sustained-window checks to five actual trajectories, including forged observations and post-completion disturbance.
37 / UNDERSTAND · Mechanisms
LLM mechanics I: should changing future tokens affect earlier outputs?
From a two-row hand calculation to a two-layer numeric experiment: causal masks, softmax and future-information leakage.
38 / UNDERSTAND · Mechanisms
LLM mechanics II: what does KV caching save, and how do we verify it?
Compare prefix recomputation, token decoding and chunk caching with output errors, projected-row counts and actual array payloads.
39 / UNDERSTAND · Mechanisms
LLM mechanics (5): after editing history, can we recompute only the changed suffix?
27 fixed conditions and 81 raw arrays test longest-prefix cropping, continuation positions, sliding-window fallback and scope isolation.
40 / UNDERSTAND · Mechanisms
LLM mechanics (4): when a prefix changes, is the old KV cache still valid?
Three seeds and six conditions compare unchecked and keyed prefix reuse; numerical equality does not establish permission to share across scopes.
41 / UNDERSTAND · Mechanisms
LLM mechanics III: why is a four-slot cache not four-token recomputation?
Separate absolute positions, cache slots and retained representations with a two-layer sliding-attention experiment.
42 / EXPLORE · Engineering cases
Embodied AI Practice VII: Why a Transfer Rule Stops Normal Lowering
21 MuJoCo rollouts separate stage-specific grasp checks, stop actions and final placement. A false alarm and lingering finger contact reveal two different failure modes.
43 / BUILD · Engineering cases
Java LLM practice 3: don't let a failed request change conversation history
Candidate snapshots, complete-pair commits and a UTF-8 request budget: 41 offline checks for history that survives failed requests.
44 / BUILD · Engineering cases
Does TypeScript validate model output? A 32-case comparison with Java
Assertions, unknown and raw JSON: 22 false accepts, four object-validation blind spots and a bounded Java/TypeScript comparison.
45 / EXPLORE · Engineering cases
Embodied AI VI: what do recovery gating and replanning each solve?
A 30-rollout ablation reduces first target jumps from 50–97 mm to about 0.25 mm, without improving placement outcomes in this fixed matrix. Separate recovery evidence, command continuity and task success.
46 / EXPLORE · Engineering cases
Embodied practice 5: communication is back—may transfer resume?
18 MuJoCo rollouts reveal unsupported resumptions hidden by eventual success; validate fresh evidence and replan the remaining path.
47 / UNDERSTAND · Engineering cases
After 24 correct answers: designing a more discriminating context experiment
Remove wording and identifier shortcuts, freeze 128 similar-distractor requests, and verify the materials with Java 8 and an independent audit. Offline preparation is not a new model result.
48 / EXPLORE · Engineering cases
Embodied practice II: if the grasp failed, skip the transfer
An 18-rollout MuJoCo comparison of a lift/contact gate, with recorded trajectories, videos, runnable code and explicit limits.
49 / EXPLORE · Engineering cases
Embodied practice IV: messages arrive, but observations expire
45 MuJoCo rollouts separate sampling intervals, bad-sample counts and capture age, showing how old messages can hide grip loss and why duration thresholds still have limits.
50 / EXPLORE · Engineering cases
Embodied practice III: a confirmed grasp still needs monitoring
A 27-rollout MuJoCo comparison of transient observation loss and forced opening, with false cancellations, a 40 ms confirmation delay, code and recorded trajectories.
51 / UNDERSTAND · Experiment reviews
Does moving an answer into the middle make a model miss it? An interrupted controlled trial
Six synthetic cases, four context conditions and 29 recorded requests. A Java experiment that separates retrieval results from HTTP 429 failures.
52 / UNDERSTAND · Engineering cases
LLM experiment II: make the comparison valid before interpreting position
A Java 8 paired-block protocol with 24 live Agnes 3.0 requests, independently audited scores and reported token usage, including the limits of a perfect observed score.
53 / BUILD · Tutorials
Java LLM Practice II: How to validate model JSON after HTTP 200
Build a Java 8 output boundary from a real fenced response and three timeouts, with strict validation, a bounded format adapter and reproducible evidence.
54 / EXPLORE · Tutorials
Embodied AI Practice I: Pick up a cube in MuJoCo, then deliberately miss
Run a contact-based pick-and-place task, compare0,25 and50mm target biases, and preserve videos, state trajectories and explicit success criteria.
55 / BUILD · Engineering cases
Inside AI Radar — from an RSS response to a published page
Trace the site's actual collection windows, daily quotas, deduplication and publishing checks, then diagnose a successful run that produced no visible update.
56 / EXPLORE · Mechanisms
Reading OpenVLA — from an image and instruction to robot actions
Follow preprocessing, visual projection, action tokens and unnormalization, with particular attention to units, boundary indices and the environment interface.
57 / UNDERSTAND · Mechanisms
KV Cache explained — why send the conversation history again?
Derive K/V reuse from causal attention, calculate cache memory, and separate inference caching from prefix reuse and chat history.
58 / BUILD · Tutorials
Calling an LLM with Java 8 — from one request to a conversation
Three runnable Java demos explain HTTP requests, JSON parsing, timeouts and conversation history.