Cloud baseline

GLM 5.3 FlashX — 0.4939 on Forge 1.0

z-ai/glm-5.3-flashx · scorer forge-1.0

GLM 5.3 FlashX, single model via OpenRouterz-ai/glm-5.3-flashx · single agentOct 7, 2026
Historic benchmarkForge 1.0 · not comparable with Forge 2.0

Forge 1.0 was frozen on 9 October 2026 and accepts no new runs; Forge 2.0 is its successor. The original results contained scorer faults. Corrections made from saved evidence on 15 October are dated on the affected run pages, with both original and corrected scores and the reason for the change. They reuse the original builds and observations; no benchmark or model was rerun, and the original inference charges remain. Uncorrected results retain known limitations: the top five original scores, 0.945 to 0.977, were too close to reliably rank the strongest models; two hourly-job tests missed Jira reads performed by a queue worker; and the old linter accepted an entity index with two range attributes that Forge refuses. Those remaining faults are not silently treated as fixed. Consult each run's correction note when comparing historical results. Forge 2.0 has separate results and a different task.

Overall
0.4939
Benchmark
Forge 1.0
Model build
25m 53s

Recorded inference cost

$2.19

From the saved OpenRouter generation billing summary (148 recorded requests). Includes auxiliary model calls made during the build. Scoring reuses the saved build; this amount describes the original inference spend.

  • Billing recovered 2026-10-18 from all 148 of this build's own generation ids with OpenRouter's generation records (148 billed, native token reconciliation passed). The earlier saved summary ($1.9790131, incomplete) was replaced; the benchmark score and run are unchanged.

Token usage

Tokens: 21.9M in · 118k out · 21.6M cached (exact, OpenRouter records)

22,018,476 tokens over 148 model calls: every call of the run, the model's own and the harness's helper calls, counted from OpenRouter records. Reasoning tokens as the provider reports them: 70,194 (not added to the total).

  • z-ai/glm-5.3-flashx-20260918

    In 21.9M · Out 118k · Cached 21.6M · Reasoning 70.2k · 147 calls

  • google/gemini-2.5-flash

    In 883 · Out 4 · Cached 0 · Reasoning 0 · 1 call

Tier breakdown

L · Lint and bundles0.9583
K · Platform currency1.00
T · Event pipeline1.00
R · Reconcile1.00
S · Storage1.00
B · Resolvers and permissions0.463
U · UI function0.812
V · Visual1.00
A · Rovo0.5417
E · Excellence0.5127

Scoring detail

How this score was built

Per-test scorer output, as posted by the app. Expand a section to inspect its tests.

Correctness combines the lowest score for each test across the scoring fixtures; economy measurements use their mean. The combined result can therefore be lower than every individual fixture’s final score. Critical defects reduce the weighted result before any score cap applies.

Earned credit before admission

(0.88 × 0.8655 core + 0.12 × 0.93 gate × 0.55 excellence) × 0.6000 critical = 0.4939

Critical defects compound a multiplier on the whole score (pre-severity 0.8232):

b_no_permission_leak 0.00 → ×0.60

The core (88% of the total) is the weighted mean of the nine measured tiers. The last 12% is the excellence slice: it unlocks in proportion to the perfection conditions below (16 of 18 met here), then pays out at the excellence tier's own measured mean. Core 0.7616 + excellence 0.0615.

t_event_rows 1.00r_heal_dropped 1.00u_widget_numbers 1.00u_ledger_table 0.14b_comment_adf_as_user 0.67k_consumer_shape 1.00k_current_apis 1.00k_dashboard_widget 1.00k_entity_declared 1.00k_llm_model_current 1.00k_manifest_semantics 1.00k_rovo_mcp 1.00k_rovo_skill 1.00k_runtime_risks 1.00k_widget_edit_bridge 1.00v_console_clean 1.00v_csp_clean 1.00l_lint_warnings 1.00

Earned score 0.494 · Cap 0.799 · Final score 0.494

The final score is the lower of the earned score and the cap. Meeting a cap's conditions adds no points.

  • current platform, complete surfaces: maximum 0.799u_ledger_tablea_action_result
  • production robustness: maximum 0.809b_no_permission_leakb_comment_exactly_onceu_llm_explainu_ledger_sortb_hidden_counta_action_permissionsl_scopes

Every invoked resolver exists and answers structured errors, nobody sees changes to issues they cannot browse (not on screen, not in an LLM prompt, not in a Realtime payload), the hidden-change count is right, and each click posts exactly one ADF comment as the viewer.

b invoke contract

1.00

34/34 invokes returned a defined, structured result; a Jira 500 on getSprint's read: handled

b no permission leak

0.00

hidden issue data shown to a person: ['sprint-action-749-light-800x600: SEC-340', 'sprint-action-749-light-800x600: SEC-401', 'sprint-action-749-light-800x600: SEC-449', 'sprint-action-749-light-800x600: data-change-id 8

b hidden count

0.1111

worst of 3 scoring sites (seed dd0b1b05f0a815e9): 1/9 hidden counts equal the oracle

b comment adf as user

0.6667

worst of 3 scoring sites (seed 133ce75a81764f1a): 4/6 comments are ADF as the viewer naming key, sprint and creep; ['SEC-449 status 404 via user', 'SEC-449 status 404 via user']

b comment exactly once

0.00

worst of 3 scoring sites (seed 133ce75a81764f1a): not exactly once: ['post@749: 0 comment(s), 0 success flag(s)', 'doubleClick@749: 0 comment(s), 0 success flag(s)']

b realtime payload clean

1.00

20/20 realtime payloads hold sprint ids only

Run notes and corrections

Scorer correction (2026-10-15): 0.2963 → 0.4939 from the same saved build. The original scorer applied a duplicate-comment penalty when an action failed its comment-count or success-flag check, even though no duplicate was recorded. All retained fixtures recorded at most one comment per action. The failed comment-action checks and original admission ceiling remain in the score. No model, benchmark, candidate application or browser probe was rerun. Original inference charges are unchanged. Legacy observation files lacked an embedded original observation hash; this correction uses their retained native build, fixture identities and original functional-grade matches.

Graded browser recording

Loading duration…

Watch the full graded browser recording at its original speed: the dashboard widget, the sprint action and the not-started sprint, as the scorer drove them on the run's own seed. The test results and screenshots on this page provide the wider evidence. Pause or scrub to inspect a view.

Ready to play0:00 / —:—
Open recording file

Drag the timeline to seek. Keyboard: arrow keys move 5 seconds; Home and End jump to the start and end.

Full graded browser recording: dashboard widget (no config, edit + Save, light and dark, 380 and 1180 px, a second instance), sprint action modal (sort, router, select, comment post through the 429 retry, double click, forbidden post, close) and the not-started sprint

Screenshots

Browser screenshots of the built app, taken during grading. Each caption names the view.

1Every captured surface
2Widget edit view
3Widget before configuration
4Sprint not started
5Sprint action · dark

Run details

Model
z-ai/glm-5.3-flashx
Engine events
0
Repair rounds
0
Started
Oct 7, 2026, 06:12 AM
Finished
Oct 7, 2026, 06:46 AM
Re-scored
Oct 15, 2026, 07:14 AM (same build, corrected scorer)

How the benchmark is scored