Cloud baseline
Gemini 4 Carbon, single model via OpenRouter
google/gemini-4-carbon · scorer forge-2.0
- Overall
- 0.5425
- Benchmark
- Forge 2.0
- Model build
- 115m 53s
Recorded inference cost
At least $18.28
From the saved OpenRouter generation billing summary (118 recorded requests). This is a lower bound; the billing record is incomplete. Includes auxiliary model calls made during the build. Scoring reuses the saved build; this amount describes the original inference spend.
- 6 telemetry call(s) carried no response id to bill
- recovered native tokens exceed session counters without complete per-call proof: prompt -1401, completion -6
Token usage
Tokens: at least 28.7M in · 193k out · 21.5M cached (lower bound, OpenRouter records)
28,925,643 tokens over 124 model calls: every call of the run, the model's own and the harness's helper calls, counted from OpenRouter records. 6 calls have no token count, so these are a lower bound. Reasoning tokens as the provider reports them: 98,897 (not added to the total).
6 calls of the run have no readable OpenRouter record; their tokens are not in the totals.
google/gemini-4-carbon
In 28.7M · Out 193k · Cached 21.5M · Reasoning 98.9k · 123 calls
google/gemini-2.5-flash
In 1.4k · Out 6 · Cached 0 · Reasoning 0 · 1 call
Tier breakdown
Each bar is the tier's mean, from 0 to 1 (for E, its mean × the excellence gate). The percentage is the tier's share of the score. Mean × share, added over all tiers, is the tests earned figure below.
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, followed by the reliability multiplier, before any score cap applies.
The rules behind these numbers: Forge 2.0 requirements, weights and score caps.
The score, step by step
0.9202 tests earned × 1.0000 critical defects × 0.5895 reliability = 0.5425 final score
tests earned = 0.15 × 0.8997 v1 + 0.7852 v2 (out of 0.85)
v1 = 0.88 × 0.9203 tiers L–A + 0.12 × 0.9294 gate × 0.8057 excellence
Reliability: every tier in the breakdown above, except Excellence, is a group of tests. Each group of tests multiplies the score by 1 − 0.10 × the shortfall of its worst test: by 0.90 when that test fails completely, by 0.95 when it scores 0.5.
- K Platform currency: worst test
k_widget_edit_bridge, score 0.75 × 0.9750 - T Event pipeline: worst test
t_reestimate_followed, score 0.5 × 0.9500Already counted in this group: t_event_rows, t_out_of_order
- R Reconcile: worst test
r_pagination, score 0.9318 × 0.9932 - B Resolvers and permissions: worst test
b_comment_adf_as_user, score 0.5 × 0.9500Already counted in this group: b_hidden_count
- U UI function: worst test
u_widget_live, score 0 × 0.9000Already counted in this group: u_widget_numbers, u_widget_chart, u_widget_edit_config, u_ledger_table, u_ledger_sort, u_llm_explain
- A Rovo: worst test
a_action_result, score 0.6 × 0.9600Already counted in this group: a_action_permissions
- R2 Dosing: worst test
r2_person_never_quota_refused, score 0 × 0.9000 - R4 Changing world: worst test
r4_sprint_close_final, score 0 × 0.9000 - R7 Custom field: worst test
r7_values_fresh, score 0.8863 × 0.9886Already counted in this group: r7_writes_accepted
- R10 Native JQL: worst test
r10_empty_and_closed, score 0.75 × 0.9750Already counted in this group: r10_precomputation_refresh, r10_shared_permissions
v1 is this run's score on the tests carried over from Forge 1.0: 15% of the total. v2 is the requirements R1–R10, each at its share of the score: the other 85%. Critical defects multiply the whole score, and so does each group of tests, by 1 − 0.10 × the shortfall of its worst test.
Excellence gate 0.9294: the average of the 18 conditions below, each counted at the value shown. Green is met in full; red is not, and lowers the gate. The 3 excellence tests are paid in proportion to the gate; together they are 1.8% of the total.
Earned score 0.5425 · No cap applies · Final score 0.5425
Under a cap the final score is min(earned, cap − 0.05 × (1 − earned)): always below the cap, and closer to it the more the run earned. With no cap it is the earned score. Avoiding a cap adds no points.
Forge's own linter reports no errors (warnings cost points), every manifest function bundles and loads, packages resolve at the kit's pins, the manifest rules the linter misses, and only the scopes the app's calls need.
l deployable
1.00lint: 0 errors, every stage reached
l bundles load
1.009/9 functions bundle, load and export their handler
l lint warnings
1.000 distinct lint warning(s): []
l real packages
1.009/9 bundles resolve @forge/* at the kit pins only
l manifest rules
1.006/6 manifest rules met:
l scopes
1.00scopes: missing [], extra []
Current modules and APIs: dashboards:widget with its edit API, the rovo:skill and rovo:mcp wiring, a Forge LLM model the platform lists as active, no /rest/api/3/search, no @forge/api storage, a nodejs22.x or nodejs24.x runtime, the consumer shape, the declared KVS entity, and deploy readiness: a manifest and permissions real Forge would accept and no code predicted to fail there.
k dashboard widget
1.00dashboards:widget 2/2: {'custom_ui_resource': True, 'edit_resource': True}
k widget edit bridge
0.75worst of 3 scoring sites (seed 3c04f61f28a4a511): 3/4 board picks reached the dashboard through the widget edit API (updateConfig/onProductSave)
k rovo skill
1.00rovo:skill 7/7:
k current apis
1.00current APIs 5/5:
k consumer shape
1.00consumer shape 3/3
k entity declared
1.00entities with a sprint-partitioned, ranged index: ['scope-change', 'scope-ledger', 'sprint-issue']
k v2 surfacesdiagnostic, no weight
1.00v2 surfaces declared: ['jira:adminPage', 'webtrigger', 'scope-ledger']
k rovo mcp
1.00rovo:mcp 4/4:
k llm model current
1.0013 LLM call(s); model ids list() does not return: []; llm module with claude: True
k manifest semantics
1.00deploy readiness, manifest + permissions: 17 rules checked, none failing here
k runtime risks
1.00deploy readiness, predicted runtime failures: 12 rules checked, none failing here
Issue updates flow trigger to queue to consumer: one ledger row per change under duplicate, reordered, dropped and concurrent deliveries, multi-sprint changelog values, re-estimates followed, Retry-After honoured, no user context in background work.
t trigger handoff
1.001023/1023 deliveries handed off correctly
t event rows
0.9942worst of 3 scoring sites (seed 3aaa6d6b73d37669): event rows: 170/172 exact (precision 1.00, recall 0.99)
t no double count
1.00every change holds at most one ledger row in every phase (duplicated deliveries included); 20 concurrent group(s) raced (independent 20)
t out of order
0.8571worst of 3 scoring sites (seed d031c3ee22136508): permuted deliveries: 11/11 rows exact, 1/3 sprints with oracle numbers
t multi sprint parse
1.00123/123 multi-id Sprint changes recorded exactly
t reestimate followed
0.50worst of 3 scoring sites (seed 3aaa6d6b73d37669): 3/6 re-estimated sprints show the oracle numbers
t retry after honoured
1.00InvocationError retryAfter 30s >= 30s
t no user in async
1.00background work is asApp only
The first scheduled run starts the backfill of every change since each active sprint started, removals included; later runs heal what the event stream missed and a run with nothing new writes nothing; they page through results in each endpoint's own style, never retry before Retry-After and finish inside the module timeout, queue continuations included.
r backfill complete
1.00130/130 historical changes v1 never recorded, recorded one virtual hour after the upgrade (the backfill is dosed; v1 rows are R1's, at 2 h)
r removals found
1.0024/24 removed-to-backlog changes found
r heal dropped
1.004/4 dropped changes healed exactly once; 0 other row(s) added
r pagination
0.9318worst of 3 scoring sites (seed 3aaa6d6b73d37669): 82/88 reads the site served in two or more pages walked to their end through the cursors the site offered (/rest/agile/1.0/board 16/16, /rest/agile/1.0/board/{id}/sprint
r rate limit
1.00run 1 (backfill) request 2: waited 2.7 virtual s in-invocation; run 1 (backfill) continuation page: waited 2.7 virtual s in-invocation; run 2 (hour-1) request 2: waited 2.7 virtual s in-invocation
r as app
1.00235/235 Jira calls of the scheduled runs and their queue continuations asApp (scheduled 235)
r completes in timeout
1.008/8 invocations of the scheduled runs and their queue continuations inside their module timeout (scheduled 8)
r idempotent rerun
1.000 ledger write(s) (scope-ledger) on a no-change run over 363 rows
Forge KVS with the storage scope, ledger reads through the declared entity index in change-time order, no KVS limit errors.
s storage scope
1.00KVS used, no 403 from the KVS proxy
s entity index used
1.00755/755 ledger reads are entity queries on a declared index
s index order
1.006/6 tables list changes in change-time order
s limits
1.00no KVS limit errors
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
1.0052/52 invokes returned a defined, structured result; a Jira 500 on sprintLedger's read: handled
b no permission leak
1.00no hidden key, summary or change id reached a person
b hidden count
0.8333worst of 3 scoring sites (seed 3c04f61f28a4a511): 15/18 hidden counts equal the oracle
b comment adf as user
0.50worst of 3 scoring sites (seed 3aaa6d6b73d37669): 6/12 comments are ADF as the viewer naming key, sprint and creep; ['25740 status 201 via user', '24977 status 201 via user', '24606 status 201 via user']
b comment exactly once
1.0012 click/double-click gesture(s), none posted more than one comment
b realtime payload clean
1.00188/188 realtime payloads hold sprint ids only
The dashboard widget's numbers and chart, its board choice saved through the host, live updates through Forge Realtime, the sprint action's ledger table, sorting, issue links, comment flow, the Forge LLM explanation, close and not-started states.
u widget loadsdiagnostic, no weight
1.004/8 configured widget views render a sprint with its numbers
u widget numbers
0.50worst of 3 scoring sites (seed 3c04f61f28a4a511): widget numbers: 4.00/8 views exact (sprints by startDate, four §1 numbers each)
u widget chart
0.50worst of 3 scoring sites (seed 3c04f61f28a4a511): 4/8 charts on one linear scale within 1px; ['no single svg[data-testid=chart]', 'no single svg[data-testid=chart]']
u widget edit config
0.70worst of 3 scoring sites (seed 3c04f61f28a4a511): edit/config 7/10: pick0_view_shows_board, pick0_reopen_pressed, second_instance_own_board
u ledger table
0.9855worst of 3 scoring sites (seed 3aaa6d6b73d37669): ledger tables: 5.91/6 exact (rows, cells, default order)
u ledger sort
0.8333worst of 3 scoring sites (seed 3aaa6d6b73d37669): 15/18 at-toggle states ordered with aria-sort (6 after another sort was active)
u issue router
1.006/6 issue keys open /browse/<KEY> through the Forge router
u comment flow
1.0032/32 comment-flow steps right (select, one comment and a success flag per click and double click, forbidden error flag, modal works)
u modal close
1.006/6 close clicks reached bridge close
u not started
1.00future sprint shows not-started only
u widget live
0.00worst of 3 scoring sites (seed 3c04f61f28a4a511): moved sprints showing the new numbers live 0/1; gate failed: subscribed; 0 idle invoke(s)
u llm explain
0.9565worst of 3 scoring sites (seed 3aaa6d6b73d37669): explain 22/23 over 5 scripted answers: 1:digits:every_number_is_a_ledger_number
Atlassian design tokens with 4.5:1 contrast in light and dark, a painted surface, no Content Security Policy violations or console errors, and the widget readable at 380 px wide.
v theme tokens
1.0029/29 surfaces themed with --ds-text* at >= 4.5:1
v dark mode
1.0029/29 surfaces paint a --ds-surface* background in their mode
v csp clean
1.000 CSP violation(s), 0 failed asset request(s)
v console clean
1.000 console/page error(s) on nominal scenarios: []
v widget sizes
1.004/4 widget sizes without horizontal overflow, every sprint visible, no number clipped
The get-sprint-scope action returns exact numbers and the visible changes for the invoking person, returns errors instead of throwing, and the skill's SKILL.md tells the agent how to use it.
a action result
0.603/5 active sprints answered exactly
a action errors
1.002/2 bad sprintId inputs return {error} without throwing
a action permissions
0.75worst of 3 scoring sites (seed 3c04f61f28a4a511): 9/12 per-person answers list exactly what that person may browse
a skill instructions
1.00SKILL.md body 5/5: result fields named ['sprintName', 'committed', 'added', 'removed', 'creepPercent', 'hiddenChanges', 'changes']
Few Jira requests in the first scheduled run and per relevant change in triggers and consumers, and surfaces that paint after one round trip — paid only in proportion to the excellence gate.
e reconcile economy
1.00mean over 3 scoring sites: backfill 32 background Jira calls (by lineage, every phase) / optimum 20 (incl. 2 forced 429 repeat(s)) = 1.60x; 1.0 at <= 2.45x; 2 2.0 call(s) not counted (field-value writes, person-facing, w
e event economy
0.417mean over 3 scoring sites: 974 Jira calls (by lineage, every phase) / baseline 214 (213 delivered relevant change(s) x 1 read + 1 forced 429 repeat; a normalizer, not a reachable optimum) = 4.55x; 1.0 at <= 2.86x; 151 fi
e ui round trips
1.00mean over 3 scoring sites: round trips before first paint: {'widget-view': 0, 'widget-edit': 1, 'sprint-action': 1}
Every v1 row moves to the new scope-ledger entity exactly once, with its change id and time, while events keep flowing and across invocation limits, within 2 virtual hours of the upgrade, with its progress in the admin panel; v1's entity stays intact.
r1 v1 rows migrated
1.00h2: 59 of 59 v1 rows in scope-ledger with the source attributes recorded by the probe
r1 v1 rows intact
1.00all 59 v1 source rows intact; 59 final copies preserve recorded v1 attributes at every observed checkpoint
r1 events during migration
1.0048 of 48 changes made while the migration ran are in scope-ledger at the end
r1 progress visible
1.006 of 6 panel reads show the migration right (total 59; complete after the h2 read)
On a site of about 1,000 issues, background work stays inside its share of the 2,400-point hour, no person-facing request meets a quota 429, and Retry-After and RateLimit-Reason (quota, burst, per-issue write) decide every retry; the backfill still completes.
r2 background share
1.0010 of 10 hours kept background within 70 % (1680 points; the contract default; the probe recorded no change); busiest hour 1680; mandatory field-write lower bound 7 (reads additional)
r2 person never quota refused
0.00worst of 3 scoring sites (seed 3c04f61f28a4a511): 5 of 971 person-facing requests refused for quota (e.g. /rest/agile/1.0/sprint/{sprintId}, /rest/agile/1.0/sprint/{sprintId}, /rest/agile/1.0/sprint/{sprintId} (+2))
r2 rate limit reaction
1.005 of 5 429s answered by their reason (quota: all background paused; burst: the endpoint; per-issue: the issue)
r2 backfill complete
1.00130 of 130 pre-upgrade changes v1 never recorded are in scope-ledger, right
Work longer than one invocation continues in the next, no function waits past its limit (a long Retry-After is handed back, not waited out), and a killed invocation leaves no duplicate.
r3 never killed
1.009 of 9 functions never exceeded their limit
r3 long retry after deferred
1.006 of 6 invocations handed a Retry-After longer than their limit stopped and resumed later (results: ok, retry)
r3 no duplicate rows
1.00every (changeId, sprintId) appears once at each of 7 checkpoints; 20 concurrent group(s) raced (independent 20)
Sprints close, issues move to another board's sprint, a board's estimation field changes, issues are deleted and a person loses browse permission while the app runs; the ledger and every view come out right.
r4 sprint close final
0.00worst of 3 scoring sites (seed 3aaa6d6b73d37669): 0 of 1 closed sprints kept exactly their final ledger; sprint 1188: 2 rows gone (e.g. ('1491846', '1188'), ('1491885', '1188'))
r4 move new board
1.0017 of 17 cross-board moves recorded with the new board's estimate
r4 field switch
1.0042 of 42 rows after the switch use the new field; 20 of 20 earlier rows keep their estimate
r4 deleted history
1.002 of 2 rows of deleted issues kept as history with deleted: true
r4 permission revoked
1.003 answer(s) after a browse revoke showed ledger data and none of the newly hidden issues
A jira:adminPage built in UI Kit with the contract's controls, and every admin resolver checking Jira's ADMINISTER permission on the server, never trusting the payload.
r5 panel labels
1.008 of 8 §2.6 controls found by label; new secret shown once: yes; then masked ••••<last4>: yes
r5 admin actions apply
1.002 of 2 admin-page invokes of the admin's save and rotate applied
r5 nonadmin refused
1.004 non-admin/forged admin actions landed no attempted effect in storage
Signed deployment events: a missing or wrong signature, a tampered body or a stale timestamp is refused with no effect, a replay applies once, a valid event records the deployment, and the secret is never disclosed.
r6 unsigned no effect
1.005 invalid events (bad-signature, stale, stale-future, tampered, unsigned) had no effect
r6 replay once
1.001 sequential replay(s) and one concurrent pair had one acceptance and no duplicate Jira/deployment effect at any checkpoint; bookkeeping writes 0
r6 status codes
1.0010 of 10 cases answered with their status
r6 valid applies
1.003 of 3 validly signed events accepted (202) and recorded as `Deployed to <env>` on every ledger row of the issues they name
r6 secret never disclosed
1.00the CI secret (panel) never appeared outside its one-time display
The read-only scope-status field holds committed or added +points for an issue in an active sprint, removed for one that has left an active sprint since that sprint started, and nothing for any other issue, written in bulk through Jira's app field-value API and fresh within the virtual hour.
r7 values fresh
0.8863worst of 3 scoring sites (seed 3c04f61f28a4a511): mean over 7 checkpoints; worst h4: 284 of 358 issues right; BILL-162 None (want added +13); OPS-104 'removed' (want )
r7 writes accepted
0.9748worst of 3 scoring sites (seed 3c04f61f28a4a511): 155 of 159 bulk field writes accepted; refused: 404
Tool calls validated against the viewer's sprint scope, back-off on a 429 without Retry-After, answers with no finish reason refused, a 10-minute cache, and the admin's kill switch and daily token budget enforced on the server.
r8 tool call scope
1.002 of 2 manipulated calls respected function and visible sprint scope
r8 failure handling
1.002 of 2 LLM failures handled
r8 cost controls
1.003 of 3 cache/kill-switch/budget cases met their full outcome
The widget and the sprint action each make at most one invoke and load at most 150 KB before their first data paint, with no external origin; the admin page renders after one invoke.
r9 widget boot
1.00widget-view: 1 invoke(s) before first paint (<= 1), 102781 bytes of JS+CSS (<= 153600), 0 external request(s)
r9 sprint boot
1.00sprint-modal: 1 invoke(s) before first paint (<= 1), 107238 bytes of JS+CSS (<= 153600), 0 external request(s)
r9 admin boot
1.00admin-page: 1 invoke(s) before first paint (<= 1)
scopeChanged(sprintId) is a jira:jqlFunction returning native issue predicates from the ledger. Jira shares its precomputed expansion across users, applies each viewer's browse permissions, and keeps the same cached mapping fresh through the documented precomputation update API after changes and reconciliation. Empty, future and closed sprints, IN and NOT IN, and invalid arguments are checked by running actual native JQL.
r10 manifest
1.00scopeChanged manifest declares issue IN/NOT IN, required sprintId, and bound function
r10 initial membership
1.0024/24 initial_membership cases correct
r10 precomputation refresh
0.8333worst of 3 scoring sites (seed 3c04f61f28a4a511): 20/24 precomputation_refresh cases correct; refresh: missing=['34870', '36071', '37988'], extra=[], invoked=0; refresh: missing=[], extra=['34870', '36071', '37988'], inv
r10 shared permissions
0.8889worst of 3 scoring sites (seed 3aaa6d6b73d37669): JQL disclosure not observed: 64/72 shared_permissions cases correct; shared-user: missing=['25036', '27166'], extra=[], invoked=0; shared-user: missing=['25036', '27166']
r10 empty and closed
0.75worst of 3 scoring sites (seed 3aaa6d6b73d37669): 3/4 empty_and_closed cases correct; closed: missing=['25036', '27166'], extra=[], invoked=0
r10 invalid arguments
1.004/4 invalid_arguments cases correct
Run notes and corrections
Scoring load: this build was scored once in the app while other work on the Mac pushed the load above the limit of 8 we normally score under on this 16-core Mac (sampled 1-minute load 8.1 to 8.96). It was not scored a second time. Build 2 of this pair was scored twice, at loads up to 12.57, and gave 0.7862 both times with all 105 check scores identical.
Graded browser recording
Loading duration…Watch the full graded browser recording at its original speed: the dashboard widget, the sprint action, the not-started sprint and the Forge LLM cases, as the scorer drove them on the run's own seed, the first of the three scoring sites. The test results and screenshots on this page provide the wider evidence. Pause or scrub to inspect a view.
The UI Kit admin panel is not in the recording: it is graded from its component tree.
Drag the timeline to seek. Keyboard: arrow keys move 5 seconds; Home and End jump to the start and end.
SHA-256: 1d6759ea6338c2ec06a86363ce48e2dd84dd50a49f2039e727c53adb344013f6
Screenshots
Pictures from one of the three scoring sites, the one on the run's own seed. The dashboard widget and the sprint action are captured from the built app in a browser during grading. The admin panel picture is drawn by the benchmark from the app's component tree, not by Jira, and nothing is graded from that picture. A test keeps the score of its worst site (the excellence measurements keep their mean), so a test can fail on a site these pictures do not show. A detail that opens "worst of …" names the seed of the site the test scored worst on, which can be the pictured one.
Run details
- Model
- google/gemini-4-carbon
- Engine events
- 0
- Repair rounds
- 0
- Started
- Nov 12, 2026, 06:57 PM
- Finished
- Nov 12, 2026, 09:53 PM
- Call budget
- With wrap-up nudge, version 2: as the call budget ran low, the agent harness told the model to start no new work and finish