GOLEM

How Kiro built it

Kiro wrote and ran GOLEM's own specs. This page is generated from site/data/kiro.json, which node scripts/kiro-stats.mjs counts from the files in this repository (generated 2026-11-19 at commit 3b48f4a).

16specs, 3 of them bugfix specs written from audits
216requirements with 2,147 EARS acceptance criteria
273correctness properties in the designs
699tasks completed. 519 property assertions (fc.assert) in 404 test files

The best session: a Kiro agent fixing the robot under the gate

The task: make requirement 1.3 hold (a 100 N push from any direction), changing only controller/reflex.ts, without breaking any line that holds today. The agent was golem-tuner from the kit pack, run headless in Kiro CLI on Opus 5.5. It took 13.6 minutes and 11.3 credits.

  1. It used GOLEM to look first. Through the MCP server it called list_requirements, falsify, get_counterexample, explain_margin and replay before touching code. The counterexample: 55.6 N, early in the stride (this session ran before the push-angle convention was fixed, so its angles use the old convention).
  2. Its first reflex made 1.3 hold and broke two other lines. Correcting forward speed too, it fell on a 6.89° slope (3.1) and at 30.2 ms of sensor latency (6.1). GOLEM caught both on the next golem test.
  3. It narrowed the reflex. Sideways only, and only sudden changes: smooth over 0.06 s, subtract a 1.5 s average, so a push counts and a slope’s steady drift does not. 3.1 and 6.1 held again.
  4. A different search seed broke it again. With a gain of 1.6, 1.3 fell at 79.9 N from 108°: the command outran the robot and it kept accelerating sideways. A gain of 0.8 held under the default search and three more search seeds.
  5. It stopped honestly. The gate was still red, only on the three lines marked Expected_Break, and it said so: the gate did not read the Expectation table yet. It also reported that the MCP server kept running the old reflex after the edit, and that our push-angle convention looked backwards. All three were real bugs, and all three are fixed.

The reflex ships as a switch, GOLEM_REFLEX=kiro. Re-measured afterwards over the full 4,860-trial stability map, the weakest push it survives rose from 62.5 N to 101.6 N, and all 36 directions hold 100 N (evidence). The prompt and the full session log are in evidence/results/tuner/; the logs of the other Kiro runs are in evidence/kiro-sessions/.

Every spec

16 specs counted, 2,147 EARS criteria. 1 more is listed as not started and left out of every total. Bugfix specs have no requirements file of their own: they list tasks that fix the defects of an audit. The G1 and candy rows are the robots' own requirement files, which GOLEM tests.

Specs across .kiro, apps/stagger/.kiro and apps/candy/.kiro
SpecTypeRequirementsEARS criteriaPropertiesTasks done
candy-appfeature151943780/85
golem-cloudfeature2017728not started, 65 open
golem-earsfeature131593677/77
golem-ears-hardeningbugfix00020/20
golem-evidencefeature2024606/6
golem-falsifyfeature212283891/91
golem-falsify-hardeningbugfix00016/16
golem-kitfeature2829949125/125
golem-signaturefeature933828/28
golem-simfeature151742060/60
golem-sim-hardeningbugfix00015/15
golem-sitefeature2722307/7
golem-studio-uifeature3035248129/129
stagger-appfeature212143745/48
candy-sorterfeature8800
hover-holdfeature3600
g1-walkingfeature61100

What is in .kiro

Steering

design-language.md, ears-for-robots.md, experience.md, interfaces.md, mujoco-wasm.md, product.md, structure.md, tech.md, testing.md. Two of them bind the look and the experience of the site and the studios: design-language.md and experience.md.

Hooks

Hooks
WhereHookTrigger
this repositoryTypecheck gateStop
this repositoryUnit tests for the saved engine filePostFileSave
this repositoryCompile robot requirementsPostFileSave
apps/candyCheck requirements on savePostFileSave
apps/candyFalsify controller on savePostFileSave
apps/candyGate the agentStop
apps/candyTest after spec taskPostTaskExec
apps/candySummon the ghostsManual
apps/hoverCheck requirements on savePostFileSave
apps/hoverFalsify controller on savePostFileSave
apps/hoverGate the agentStop
apps/hoverTest after spec taskPostTaskExec
apps/hoverSummon the ghostsManual
apps/staggerCheck requirements on savePostFileSave
apps/staggerFalsify controller on savePostFileSave
apps/staggerGate the agentStop
apps/staggerTest after spec taskPostTaskExec
apps/staggerSummon the ghostsManual
golem init packCheck requirements on savePostFileSave
golem init packFalsify controller on savePostFileSave
golem init packGate the agentStop
golem init packTest after spec taskPostTaskExec
golem init packSummon the ghostsManual

Agents, MCP servers, skills, workflow, Power and Crew app

Which Kiro surface did what

291 commits counted. By spec prefix: candy-app 26, golem-kit 25, golem-ears-hardening 21, golem-studio-ui 18, golem-falsify-hardening 17, golem-sim-hardening 16, golem-sim 16, golem-falsify 15, evidence 12, site 9.

Three independent audits

After the engine specs were done, each module got an independent review with probes against real inputs, not the module's own tests. Each review became a Kiro bugfix spec. The golem-auditor agent, which now does this job inside Kiro, was written afterwards; its first audit, of the Expectation-aware gate, is in docs/reviews/gate-audit.md with its session log.

golem-ears

An independent review of the EARS compiler and monitors. Verdict: FIX, not rebuild.

Bugfix spec: golem-ears-hardening, 20 of 20 tasks done.

docs/reviews/ears-audit.md

golem-sim

An independent review of the MuJoCo episode runner. Verdict: FIX.

Bugfix spec: golem-sim-hardening, 15 of 15 tasks done.

docs/reviews/sim-audit.md

golem-falsify and pool

An independent review of the falsifier and trial pool. Verdict: KEEP; fix defects 1 and 2 before the demo.

Bugfix spec: golem-falsify-hardening, 16 of 16 tasks done.

docs/reviews/falsify-audit.md

What went wrong

The push direction never reached the simulator

All tests passed, and the G1 pushes ignored their direction. The ears audit found it (defect 1): the compiler matched a dimension to the perturbation only by name, and the direction slot's dimension is called push.angle. A test hid it with a loose toMatchObject. The first run of the stability map was made on that build, so every direction gave the same boundary. That run is kept as invalid in evidence/results/invalid/, and the valid map was re-run after the fix. See docs/reviews/ears-audit.md.

The fallen margin was too flat for guided search

A fall scored a fixed −0.01 m. The cross-entropy search needs a slope, and flat robustness made it lose to random sampling on a synthetic needle (9/30 against 21/30 at budget 1000 before the fix, the falsify audit's defect 2; 15/30 against 22/30 after it). The G1 app then got a graded margin (height and tilt headroom, same verdict), and the benchmark was run again. On the real G1 lines the arms still tied at budget 96. See the benchmark.

The scheduler doubled frames

In the studio, a render tick that called requestOnce while a reason to render was held queued a second animation-frame callback. The callbacks doubled every frame, and the G1 page took seconds per frame. It was found in the browser while building the signature moments, fixed in RenderScheduler, and given a regression test (commit 3d9f56c).

Smaller ones, found while building this site

Counts above: 699 tasks done of 707, counted in the markdown task lists. The golem-cloud spec is written and not started; it is not part of this build.