Your LLM logs say what was asked. We add what happened next.

GitClear runs as a container inside your own Databricks pipeline and joins LLM request logs to git, pull requests, and org data—so every dollar of inference can be traced to code that did or didn't last.

The missing join

Millions in LLM spend, and no way to say what it bought.

Vendor dashboards report acceptance rates and lines suggested. Those measure enthusiasm, not durable output, so renewals get approved on anecdote.

Is your biggest token spender an anti-pattern to rein in, or the prototype every team should copy?

Finance sees

The invoice, by vendor and month.

Your git provider sees

Commits and pull requests, by author.

Nobody sees the link

Which tokens produced which lines, whether they survived review, and whether they were still present months later.

The opportunity

You're already logging the evidence.

Legal, Security, or a Works Council usually requires that LLM traffic never leaves the network unobserved. Every request and response is therefore already landing somewhere.

AI gateways and proxies

Portkey LiteLLM Kong AI Gateway Cloudflare AI Gateway

LLM observability

Helicone Langfuse Braintrust Datadog LLM Observability

Cloud invocation logs

Amazon Bedrock Google Vertex AI Azure OpenAI

Vendor telemetry

Anthropic Analytics API Claude Code OpenTelemetry GitClear hooks
bronze.llm_gateway_requests
prompt, completion, model_id,
tokens_in, tokens_out, timestamp

Whatever the source, it ends up in a Delta Lake table. It was built for compliance, and it is accidentally the most complete record of how your engineers work. Joining it to what happened next is the opportunity.

Deployment

A watermark processor that runs inside your pipeline.

You export nothing to GitClear. Our container reads from your tables and writes enriched silver and gold tables back to your account. Hourly or nightly runs resume from the last watermark.

bronze Sources never joined before

Git commits Substantive change, churn, moves, and copied code.
Pull requests Review turnaround, approvals, and release state.
HR and org data Teams and cost centers leaders manage.
LLM logs Model, tokens, cost, generated code, and task category.

GitClear container

$ gitclear-enrich run
resume from watermark
classify prompts in-cluster
write silver.*, gold.*
  • Astronomer / Apache Airflow
  • Databricks Workflows
  • Kubernetes, Argo, Dagster, or Prefect

silver gold Question-shaped tables

silver.ai_work_episodes One row per AI-assisted task.
silver.change_lineage Lines followed through moves and rewrites.
gold.model_task_roi Durable output per dollar.
gold.token_release_funnel Generated code that reaches production.

Row by row, the output connects this model to this task and the code that did—or did not—survive.

Decision-ready output

Every graph answers a decision someone has to make.

Is the frontier model worth its price?

Task × model ROI

Compare productivity, all-in cost, speed, revision risk, and steering by task. A less expensive model may match the frontier model on tests while the frontier earns its premium on cross-file refactors.

gold.model_task_roi  — enriched from your bronze tables
databricks unity catalog · airflow on astronomer · 5,200 ai work episodes
bronze → silver → gold
rebuilt nightly · 12 min p95
Bronze
bronze.ai_assistant_events
bronze.git_commits
bronze.pull_request_reviews
as your vendors ship it
conform
Silver
silver.ai_work_episodes
silver.change_lineage
one row per AI work episode
relate
Gold
gold.model_task_roi
gold.durability_curve
question-shaped marts
01 Task–Model ROI Matrix
“For this kind of task, which model has historically delivered the best value?”
best value
lowest all-in cost
fastest result
lowest revision risk
least steering
prompt category
Apex-R 2.4
premium reasoning
Apex-F 2.4
premium fast
Core 3.1
balanced workhorse
Lite 1.9
cheap, verbose
Local-32B
self-hosted
implement_feature
outcome: durable Diff Delta @ 90d
1.85×
$79.2 / success
80% @30d · n=262
1.70×
$79.8 / success
78% @30d · n=87
0.97×
$98.9 / success
74% @30d · n=612
0.37×
$212 / success
60% @30d · n=349
0.11×
$436 / success
49% @30d · n=146
root_cause_bug
outcome: verified diagnosis, no recurrence
2.34×
$81.9 / success
82% @30d · n=131
1.62×
$109 / success
71% @30d · n=44
0.81×
$154 / success
64% @30d · n=306
0.19×
$558 / success
44% @30d · n=175
0.04×
$1381 / success
34% @30d · n=73
dry_cleanup
outcome: duplication removed, retained @ 90d
1.63×
$52.1 / success
90% @30d · n=66
n=22 insufficient
1.00×
$55.5 / success
90% @30d · n=153
0.60×
$75.8 / success
86% @30d · n=87
0.17×
$156 / success
70% @30d · n=36
explain_code
outcome: explanation accepted, next task lands
1.60×
$21.0 / success
97% @30d · n=112
1.74×
$18.0 / success
97% @30d · n=37
1.00×
$21.5 / success
94% @30d · n=262
0.51×
$34.8 / success
83% @30d · n=150
0.15×
$70.0 / success
69% @30d · n=62
address_pr_feedback
outcome: thread resolved without reopen
1.71×
$50.3 / success
89% @30d · n=103
1.78×
$44.7 / success
92% @30d · n=34
0.98×
$57.1 / success
87% @30d · n=240
0.42×
$110 / success
73% @30d · n=137
0.11×
$238 / success
59% @30d · n=57
build_failing_test
outcome: fails before fix, passes after
1.66×
$32.1 / success
94% @30d · n=47
n=16 insufficient
1.04×
$33.6 / success
91% @30d · n=109
0.48×
$59.9 / success
78% @30d · n=62
0.14×
$117 / success
65% @30d · n=26
best in row tied with best (CI overlap) n < 25, no ranking shown
02 Token-to-Durable-Production Yield Curve
“Where does each model's apparent productivity disappear?”
Share of each model's initially attributed change still present in production, by funnel stage 0% 27% 53% 80% 106% right-censored: 12d exposure, survival estimate ↑ % of initially attributed change still present Apex-R 2.4 74% Apex-F 2.4 71% Core 3.1 56% Lite 1.9 29% Local-32B 24% generated accepted committed merged day 7 day 30 day 90 — pre-merge leakage — — post-merge decay —
Diff Delta is lineage-aware: moves, renames and reformatting keep their lineage, so code the developer rewrites stops counting toward the model.
A clipped section of two gold tables from our design study, drawn in HTML/CSS so it scales. Synthetic data : model names are placeholders and no real measurement is implied.

How many of our tokens never ship?

Token release funnel

Follow generated lines through commit, review, merge, release, and survival at 90 days. Drop-off between stages shows exactly where spend stops producing value.

Security

Aggregated by design, so projects don't stall in review.

Enterprise results roll up by team, model, or cost center. This keeps the enriched tables clear of unnecessary PII and the individual surveillance concerns that stall AI-measurement projects.

SOC 2 Type II and ISO 27001 certified, penetration tested in 2026. Review GitClear security

Inside your Databricks account

  • Prompt and completion text
  • Source code and commit history
  • In-cluster prompt classification
  • Enriched silver and gold tables

Find out what your LLM spend has been paying for.

If you already capture LLM traffic, we'll scope an enrichment pipeline and quote it for free.