Borehole · Risk register · free tier

unslothai/unsloth

Free reading: not for reliance

7f32241 · 11,953 commits · 9,151 files
Issued 2026-10-23

AGPL-3.0 since 2026-03-12: open source; offering it over a network counts as distributing it, so a hosted product's changes must be published.

103 checks ran across 6 assessable dimensions. 3 findings are worth checking first. 18 findings, each with evidence. One dimension cannot be assessed without an acquirer.

What matters most

05.60 AGPL-3.0 since 2026-03-12: open source; offering it over a network counts as distributing it, so a hosted product's changes must be published

What the repository says about itself

03.60 The project says one person maintains it README.md:476
05.60 Relicensed: Apache-2.0 → AGPL-3.0 (2026-03-12) LICENSE

Worth checking first

Borehole's proprietary scoring is calibrated across hundreds of repositories. Each finding here reads significant, but its check has not yet held up often enough in that calibration to rate significant on its own. If one holds when someone checks it, it is significant.

05.20 27 SQL statements built from strings
05.26 15 locked dependencies have published vulnerabilities, 5 high or critical in what ships
05.8 Manifest says Apache-2.0, licence file says AGPL-3.0

Published under today's rules. The verification code attests to this reading as it was stored. This page differs from it as follows.

3 findings from checks Borehole's calibration found wrong more often than right are no longer published.
The file and line of 1 credential-shaped match is withheld. File and line are shown to whoever GitHub says controls this repository, signed in, and to a buyer of this report.

	Dimension	Rating
01	Architecture & Codebase — including AI provenance	adequate
02	Engineering Process & SDLC Maturity	adequate
03	Organization & Key-Person Risk	adequate
04	Product & Engineering Maturity	strong
05	Security & Reliability Posture	adequate
06	Observability & Operations	strong
07	Fit with the Acquirer	not assessable

strong adequate significant findings not assessable from a repository

Provenance disclosure does not affect any rating

1% of added lines arrived in commits declaring an AI agent as co-author

The surviving-lines measurement was not run on this repository.

1% of lines added in the last 12 months (since 2025-10-23) arrived in AI-declared commits: 116/8,685 commits, confidence moderate. Also a floor.

116/11,374 commits declaring an agent · Claude, Claude Fable 5, Claude Fable 5.1 and 15 more

confidence: moderate

39% of non-merge commits carry any trailer at all
low trailer coverage: the convention may not be in consistent use
no widespread author/committer date divergence

69,540 lines excluded as lockfiles, vendored or generated.

This is a lower bound. It counts what the repository declares. Undeclared AI authorship is invisible here, and a single rebase removes the trailers this rests on. Absence of a declaration is absence of disclosure, not absence of AI authorship.

Architecture & Codebase — including AI provenance

01 · adequate · ratings ↑

01.25

2 base images float: no version, or `latest`

adequate

2 of 4 `FROM` lines in product Dockerfiles name an image without a version, or with `latest`. Each build pulls whatever the registry serves that day, so the image that ships can change without a commit, and last month's build cannot be reproduced. Pin a version, or a digest.

Evidence

docker/Dockerfile	`docker/Dockerfile:368` — base image without a version
docker/Dockerfile	`docker/Dockerfile:397` — base image without a version

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

An owner's response is removed only when it is unlawful, impersonates someone, publishes personal data or a credential, or is not about the finding. Disagreeing with the report is never a reason. To report a response, or to reach us about one, write to takedown@borehole.dev.

Notes that do not affect this rating

01.6

studio/backend/core/inference/diffusion.py is 14,543 lines, 85× this codebase's median file

adequate

The median source file here is 170 lines. 159 files exceed 1,700 lines, the largest being 14,543. Size is not a defect on its own — some problems genuinely live in one place — but these are where merge conflicts, review fatigue and single-owner knowledge concentrate, and they are the first thing that slows an inheriting engineer down.

A note that does not affect this rating: measured against its own codebase's median, it is below the 85th percentile of 382 surveyed projects' largest files, where this check begins to count.

Evidence

line-count	`studio/backend/core/inference/diffusion.py:1` — 14,543 lines (85× median)
line-count	`studio/install_python_stack.py:1` — 13,744 lines (80× median)
line-count	`studio/backend/core/inference/video.py:1` — 13,662 lines (80× median)
line-count	`unsloth/import_fixes.py:1` — 12,906 lines (75× median)
line-count	`studio/install_llama_prebuilt.py:1` — 12,567 lines (73× median)

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

01.15

1,843 errors caught and discarded in product code

adequate

1,843 places catch an error and do nothing with it, across 432 files (1.0 per thousand lines). Each is a failure the system will not report. They surface later as problems nobody can explain, and they are cheap to fix now.

A note that does not affect this rating: in Borehole's calibration, this check's notes were usually true but overstated.

Evidence

docker/jupyter/unsloth_branding.py	`docker/jupyter/unsloth_branding.py:225` — error caught and discarded
docker/jupyter/unsloth_branding.py	`docker/jupyter/unsloth_branding.py:230` — error caught and discarded
docker/unsloth_nb_strip_colab.py	`docker/unsloth_nb_strip_colab.py:51` — error caught and discarded

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

Engineering Process & SDLC Maturity

02 · adequate · ratings ↑

02.24

5% of recent commits change more than 2,000 lines

adequate

512 of 10119 commits in the last 24 months changed over 2,000 lines of code, not counting reformatting, renames, vendored or generated code. A change that size is rarely reviewed line by line. Ask how they were reviewed, and what made them that large.

Evidence

git log --numstat

512 commits over 2,000 lines

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

Organization & Key-Person Risk

03 · adequate · ratings ↑

03.1

Bus factor 1: recent work is concentrated in one contributor

adequate

In the last 24 months, one contributor authored more than half of 10,730 non-merge commits among 429 authors; across all history, more than half of 11,263. This is the single most expensive thing to discover after closing, and the repository cannot tell you whether that person is part of the transaction.

Evidence

git shortlog -sn

last 24 months: top author more than half of 10,730 commits; bus factor 1; 429 authors

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

03.60

The project says one person maintains it

adequate

Its own files say so. Whatever the commit counts show, the project describes itself as depending on one person.

Evidence

README.md README.md:476 — says it has a single maintainer

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

Product & Engineering Maturity

04 · strong · ratings ↑

No finding. 19 checks ran and none fired; the rejection log below says why each did not.

Security & Reliability Posture

05 · adequate · ratings ↑

05.1

1 credential-shaped string committed to the tree

adequate

Matches for JWT appear in tracked files outside test and fixture paths. They stay in the history even after they leave the tree, so what matters is whether each was rotated. The code cannot tell whether any of them is live, and that decides whether this matters: none sits in production configuration, and none is in a format its provider keeps for live keys. Is any of these live? Show that each was rotated after it was committed, and when.

Evidence

JWT	JWT pattern matched
	File and line are shown to whoever GitHub says controls this repository, signed in, and to a buyer of this report.

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.6

1 Dockerfile runs the application as root

adequate

No USER instruction switches away from root, so the process runs with full privileges inside the container. It is a one-line fix and an ordinary expectation, which is why its absence is worth asking about: it usually says more about whether anyone has reviewed the deployment than about this container specifically.

Evidence

no-USER-instruction docker/Dockerfile — runs as root

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.20

27 SQL statements built from strings

adequate worth checking first

27 places in product code build SQL by formatting or joining strings rather than passing parameters, and in 13 of them a value, not a table or column name, goes into the string. Where any of those values comes from a request, it is an injection point. Each site needs a look; the fix, parameters, is small.

The check that found this reads significant, but it has not yet held up often enough in Borehole's calibration to rate a finding significant on its own. Opening studio/backend/auth/storage.py:597 and the 1 other file it cites settles it. If it holds when someone checks it, this finding is significant.

Evidence

studio/backend/auth/storage.py	`studio/backend/auth/storage.py:597` — SQL built from a string
studio/backend/auth/storage.py	`studio/backend/auth/storage.py:601` — SQL built from a string
studio/backend/core/rag/store.py	`studio/backend/core/rag/store.py:577` — SQL built from a string

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.26

15 locked dependencies have published vulnerabilities, 5 high or critical in what ships

adequate worth checking first

15 of 1,764 locked dependencies match 26 published advisories in OSV, 6 of them rated high or critical. A lockfile says which version is installed, not whether the flawed code is reachable from this product, so each needs a look. Most are fixed by an upgrade, and the advisory names the version. This lockfile pins development tools alongside what ships and does not say which is which, so some of these may never reach a user.

The check that found this reads significant, but it has not yet held up often enough in Borehole's calibration to rate a finding significant on its own. Opening studio/frontend/package-lock.json settles it. If it holds when someone checks it, this finding is significant.

Evidence

studio/frontend/package-lock.json	`studio/frontend/package-lock.json` — dompurify 3.4.16: GHSA-cg2g-xh64-v4q7 (critical), fixed in 3.4.17
studio/frontend/package-lock.json	`studio/frontend/package-lock.json` — @modelcontextprotocol/sdk 1.31.0: GHSA-22jm-h49p-29qw (high), fixed in 1.32.0
studio/frontend/package-lock.json	`studio/frontend/package-lock.json` — node-forge 1.4.0: GHSA-86w9-cpqp-85rv (high)

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.8

Manifest says Apache-2.0, licence file says AGPL-3.0

adequate worth checking first

pyproject.toml declares Apache-2.0. COPYING contains the text of AGPL-3.0. Package managers and downstream users read the first; auditors and acquirers read the second. One of them is wrong, and until it is settled the terms this code is offered under are genuinely ambiguous.

The check that found this reads significant, but it has not yet held up often enough in Borehole's calibration to rate a finding significant on its own. Opening pyproject.toml:16 and the 1 other file it cites settles it. If it holds when someone checks it, this finding is significant.

Evidence

pyproject.toml	`pyproject.toml:16` — declares Apache-2.0
COPYING	`COPYING` — text of AGPL-3.0

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.60

Relicensed: Apache-2.0 → AGPL-3.0 (2026-03-12)

adequate

The root licence changed 1 time(s). The latest terms restrict what users, hosts or a buyer may do in ways the earlier ones did not. Code taken under the old terms stays under them, so both matter in diligence.

Evidence

1e2ba1b1d	`LICENSE` — Apache-2.0 from 2023-11-29
c26aa1a1e	`COPYING` — AGPL-3.0 from 2026-03-12

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

Notes that do not affect this rating

05.27

The build runs third-party code it does not pin

adequate

24 steps pipe a download straight into a shell. Tags have been moved to malicious commits in real supply-chain attacks, and a piped script is whatever the server returns that minute. Pinning to a commit is the fix, and it is small.

A note that does not affect this rating: notes like it are common in well-run projects.

Evidence

.github/scripts/agent-guides-install.sh .github/scripts/agent-guides-install.sh:63 — download piped into a shell

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.21

14 shell commands built from strings

adequate

14 places run a command through a shell (`shell=True`, `os.system`, `exec` with a template) rather than passing arguments. Where any part comes from outside, it is command injection. Ask where each one's input comes from.

A note that does not affect this rating: in Borehole's calibration, this check's notes were usually true but overstated.

Evidence

studio/backend/core/inference/tool_path_approval.py	`studio/backend/core/inference/tool_path_approval.py:2426` — command run through a shell
studio/backend/core/inference/tool_path_approval.py	`studio/backend/core/inference/tool_path_approval.py:3184` — command run through a shell
unsloth/_gpu_init.py	`unsloth/_gpu_init.py:714` — command run through a shell

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.22

2 deserialisers that can run code

adequate

2 places read data with a deserialiser that executes what it reads (pickle, `yaml.load` without a safe loader). That is safe on data the system wrote itself and remote code execution on anything else. Ask where each one's input comes from.

A note that does not affect this rating: in Borehole's calibration, this check's notes were usually true but overstated.

Evidence

studio/backend/core/inference/tool_path_approval.py	`studio/backend/core/inference/tool_path_approval.py:3407` — code-executing deserialiser
unsloth_cli/commands/eval.py	`unsloth_cli/commands/eval.py:309` — code-executing deserialiser

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.48

6 places insert HTML from a variable without sanitising it

adequate

6 places set HTML from a value that is neither a literal nor passed through a sanitiser. Where the value can come from a user, that is stored cross-site scripting, and it is the first thing a buyer's security review looks for in a frontend.

A note that does not affect this rating: in Borehole's calibration, this check's notes were usually true but overstated.

Evidence

studio/backend/assets/docs_ui/redoc.standalone.js	`studio/backend/assets/docs_ui/redoc.standalone.js:2` — HTML from a variable
studio/backend/assets/docs_ui/redoc.standalone.js	`studio/backend/assets/docs_ui/redoc.standalone.js:2` — HTML from a variable
studio/backend/assets/docs_ui/redoc.standalone.js	`studio/backend/assets/docs_ui/redoc.standalone.js:752` — HTML from a variable

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

05.65

2 advisories against this project in the last 24 months, 1 high or critical

adequate

The newest was published on 2026-10-19. A project that publishes advisories is telling its users about its fixes, and the count and pace say how much security work it carries. None of them affects the surveyed version.

A note that does not affect this rating: it describes the project's advisory record, and no advisory in it is known to affect the surveyed commit.

Evidence

CVE-2026-112075	2026-10-19: CVE-2026-112075 (low): A weakness has been identified in unslothai unsloth up to 0.1.501-beta. This affects the function FastSentenceTransformer._is_transformer_module_ref of the file
CVE-2026-93348	2026-09-28: CVE-2026-93348 (high): Unsloth Zoo versions 2025.9.9 before 2026.8.14, as implemented in Unsloth 2025.9.9 through 2026.8.19, contains a code injection vulnerability in the model-loadi

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

Observability & Operations

06 · strong · ratings ↑

A note that does not affect this rating

06.9

No error reporter is wired in

adequate

Nothing in the service initialises an error reporter (Sentry, Rollbar, Bugsnag, Honeybadger or an APM agent). An exception in production reaches a person only if someone is reading the logs, which usually means a customer reports it first.

A note that does not affect this rating: notes like it are common in well-run projects.

Evidence

error reporting

no error reporter initialised

Respond to this finding

Your response is published beside the finding, not instead of it. A finding you can argue with is the point — and a wrong one is calibration data we want.

Fit with the Acquirer

07 · not assessable · ratings ↑

Not rated: structurally not assessable from a repository. What a person adds here.

Check it on the running system

7 checks against the running system follow from the findings above: what a buyer's engineer will try, how to run it first, and what to have ready. A paid survey of this repository includes them. See the tiers.

A paid survey of a public repository is also reviewed: a language model reads each finding against the code and withdraws only findings it judges the code contradicts, each listed with its reason. The review is a language model's reading of the findings against the code, and it can be wrong. It never adds a finding, never raises a band, and never removes a finding from a proven check or an advisory that affects the surveyed version.

What a person adds

The survey reads what the repository can show. The code cannot settle these questions, and on them this report states nothing. Each says where its answer lives: your own engineer, the company, or an assessor you choose can settle it. Questions marked measured open with something this repository shows.

Architecture & Codebase — including AI provenance

01 · full coverage

measured No architecture or design document is in the tree, across 3,733 product files. Where does the design live, and who holds it? Answer lives in: The team01.38
Is the architecture right for the next three years of the product? Answer lives in: Your architects, with the product roadmap01.39
Were the home-grown components reasonable choices when they were made, and are they now? Answer lives in: The team's reasoning at the time01.40

Engineering Process & SDLC Maturity

02 · high coverage

Is review real or a rubber stamp? Comment depth and time to approve need the pull request history. Answer lives in: The code host's pull request data02.32
Is delivery planned and measured, or heroic? Answer lives in: The issue tracker: backlog, sprint history, estimate accuracy02.38
How are incidents run, reviewed and learned from? Answer lives in: Tickets, chat and postmortems02.37

Organization & Key-Person Risk

03 · partial coverage

measured One contributor authored more than half of the commits in this history. Is that person part of the transaction, and for how long after it? Answer lives in: The deal terms03.29
Who mentors whom, and would that survive a departure? Answer lives in: The team; review history only suggests candidates03.28
Can the team grow, and who stands behind the top three contributors? Answer lives in: Recruiting data and the org chart03.30
Is the seniority mix sound, and will pay survive your salary bands? Answer lives in: HR data03.31
Is the pace sustainable? Commit times cannot say: git keeps no reliable local time. Answer lives in: Interviews and on-call history03.14
Does the team want to stay after close? Answer lives in: Interviews03.32
How are decisions made and disagreements resolved? Answer lives in: Time with the team03.33

Product & Engineering Maturity

04 · minimal coverage

measured 8,574 commits landed in the last year, against 2,150 the year before (+299%). What drove it: new people, a deadline, or code produced faster than before? Answer lives in: The team04.11
measured Experiment tracking is configured (tensorboard, wandb). Can the model that ships be tied to the run, data and code that produced it? Answer lives in: The ML team and its run history04.34
measured No model card is in the tree. What is the model's intended use, what was it trained on, and where does it fail? Answer lives in: The ML team04.35
Can this team deliver the stated roadmap on the stated timeline? Answer lives in: The data room and the team's record of estimates04.32
Does leadership know about the debt, and is time budgeted to repay it? Answer lives in: Budget and roadmap decisions04.31
Is the engineering shape consistent with the commercial story? Answer lives in: Revenue, customer and churn data04.33

Security & Reliability Posture

05 · partial coverage

measured 05.1 and 05.5 found credential-shaped strings in this repository. Was each one revoked at the provider? Deleting it from the code revokes nothing; only the provider's audit log shows that it was. Answer lives in: The provider's console and audit log05.38
measured No threat model is in the tree of what looks like a running service. Does one exist elsewhere, and who owns it? Answer lives in: The security owner05.43
What has actually broken, how often, and how badly? Answer lives in: Ticketing and chat05.41
What did external testing find, and was it fixed? Answer lives in: Reports in the data room05.42
What is the SOC 2, ISO 27001 or GDPR status, and where is the evidence? Answer lives in: Audit reports; policy files in the tree are not evidence05.44

Observability & Operations

06 · partial coverage

measured Nothing in the tree backs up or restores data. How is data backed up, and has a restore ever been rehearsed? Answer lives in: The operations team06.19
measured No alert rule is defined in the repository. Where are alerts defined, who receives them, and when did one last fire? Answer lives in: The alerting tool and the on-call rota06.10
Are the dashboards opened, the alerts acknowledged and the pages answered? Answer lives in: Pager history and dashboard logs06.29
Are the service-level objectives actually met? Answer lives in: Production metrics06.30
Who carries the pager, and how often does it go off? Answer lives in: Pager exports and interviews06.31
What does it cost to run, and how much headroom is there? Answer lives in: Cloud billing and capacity plans06.32

Fit with the Acquirer

07 · none coverage

Not rated: structurally not assessable from a repository.

This is primarily a TypeScript codebase, with Python alongside (32% of source files). Does your engineering organisation already run TypeScript in production, or would this be the first? 07.1
The bulk of this system sits in studio/, unsloth/, unsloth_cli/. Which of your existing systems would each of those have to absorb or replace? 07.10
The repository carries deployment configuration (docker/Dockerfile and 3 more). Does that match your deployment and on-call model, or does it need rehosting first? 07.2
What is your tolerance for the findings above remaining unaddressed for the first two quarters after close, given everything else that will compete for this team's time?
Which of these systems duplicate yours, and which get retired? Answer lives in: Your architecture07.11
What will integration cost, and in what order? Answer lives in: The engagement07.12
Where does the team land in your organisation, and who reports to whom? Answer lives in: Your org design07.13
Are there earn-outs, retention packages or contractor IP to assign? Answer lives in: Deal terms07.14
Do customer contracts limit hosting, regions or vendors? Answer lives in: Customer contracts07.15

Request a human assessment Download JSON Badge PDF

A human assessment is Borehole's own paid service, €12,000+ (total price is determined per engagement; travel billed at cost). It is not independent of this report: the same business produced both. It is a separate contract, with its own scope, terms and reliance.

Embed the badge

Tied to this commit and this date, and it fades after 90 days — an honest badge is one that expires. A repository with significant findings gets a neutral "read" badge rather than a red one: the library is where the range is visible, not someone else's README.

[![Borehole](/badge/unslothai/unsloth.svg "Borehole free reading: not for reliance. See the full report.")](/r/unslothai/unsloth/7f32241)

If you embed it: link it to the report, remove it if the report is delisted, and do not present it as a certification or an endorsement.

Rejection log — 85 signals that ran and did not fire

Appendix: rejection log — 85 signals that ran and did not fire

Published on purpose. A signal that looked and decided not to fire has told you something, and this is the calibration data behind every threshold above.

Looked, and found nothing — 65

Signal	What it found
04.hotspot_concentration	change is spread across the codebase rather than pooling in a few filesfiles 10,224 · top5 share 0.048
04.debt_markers	debt markers are sparsemarkers 27 · per kloc 0.01 · lines read 1,904,025
04.dependency_sprawl	direct dependency counts are unremarkable
04.churn_trend	rework is not rising against new worklate delete per insert 0.272 · early delete per insert 0.447
04.fixes_with_tests	fixes usually come with a testfixes 1,640 · share 0.452 · with tests 742
04.fix_share_trend	fixing is not crowding out building
04.compat_layers	little code is kept for the old wayshare 0.0 · code files 3,334 · compat files 1
04.dependency_redundancy	one library per job, mostly
06.telemetry_presence	a telemetry or error-reporting library is referencednote presence only; whether anyone watches it is not visible here · matches 6
06.structured_logging	a structured logging library is referencedmatches 61
06.healthcheck_presence	a health or readiness check is defined
06.correlation_ids	a request can be followed through the logstracing False · request id sites 62
06.usage_reporting_default	the product's code names no reporting hostsettings on 1 · settings off 1
06.environment_branching	environment differences live in configurationsites 0
06.release_markers	releases are markedtags last year 88 · release commits 11 · changelog commits 21
06.migrate_on_start	migrations do not run at container start
03.active_contributors	more than one contributor active in the trailing yearall time 439 · active 12mo 352
03.knowledge_silos	every substantial directory has been touched by more than one personauthors 439 · directories 6
03.drive_by_contributors	an open contributor base, where one-commit contributors are normalauthors 439 · one off share 0.663
03.contributor_attrition	the dominant contributor is still activeauthors 439 · top share more than half · top contributor last committed within the last three months
03.critical_path_ownership	more than one person changes the critical pathsauthors 429
03.second_contributor_attrition	the second and third contributors are still active
03.bus_factor_trend	the team has not narrowed to one person
03.joiners_and_leavers	the team is not shrinkingleft 7 · joined 80 · regular last year 89
02.merge_discipline	a meaningful share of recent change carries evidence of reviewmerges 559 · window last 24 months · with tracked 353
02.ci_presence	CI configuration is present in the working treecount 63
02.test_presence	tests are present in a normal proportion to sourcetests 5,138 · source 8,414 · test ratio 0.611
02.release_cadence	tags are presenttags 107
02.conventional_commits	commit subjects generally carry contentsampled 11,374 · low content share 0.021
02.branch_protection_hints	process scaffolding is present in the repository
02.ci_runs_tests	CI invokes a test runnerconfigs 63 · with tests 30
02.skipped_tests	few tests are switched offfocused 0 · skipped 0 · test files 5,162
02.ci_test_gating	no test run in CI is written so that it cannot failconfigs 63
02.revert_share	change is rarely undoneshare 0.005 · reverts 53 · recent commits 10,730
02.dependency_update_cadence	dependencies move with the codelockfiles 5 · commits since 120 · days since lockfile changed 0
02.flaky_test_markers	tests are not retried until they passtest retries 0 · suite retries 0
02.lint_in_build	a linter or formatter runs in the buildwired 11 · ci configs 63 · product files 3,334
02.long_lived_branches	the typical branch reaches the trunk quicklyshare 0.105 · merges 38 · long lived 4
01.burst_import	the largest burst was merged pull requests, each developed and reviewed on its own branch (100% of it), which explains the rate: translations arrive from a vendor in bulk, dependencies and generated code are committed whole, a merged pull request is dated when it merged, not when it was written, a branch merged by a merge commit was reviewed as one piece, and an export or mirror tool dates history when it replays itlargest commits f6260a9..899d7ae · bursts explained 10 · largest net added 7,665
01.initial_dump	first commit is not a disproportionate share of the codebaseinsertions 363 · share of insertions 0.0
01.message_uniformity	commit subjects vary in wording and lengthlength stdev 29.47 · repeat share 0.209
01.revision_depth	most files were revisited after they were first writtenfiles seen 3,695 · write once share 0.272
01.duplicated_blocks	little shipping code is repeated verbatim across filesshare 0.0069 · product files 3,313 · distinct blocks 1,143,005
01.unfinished_migration	no language migration is stalled halfway
01.weights_in_git	no model weights are committedfiles 0
05.security_guardrails	at least one standard guard rail is configured
05.dependency_pinning	a lockfile is committed, so builds are reproduciblelockfile True
05.license_present	a licence file is presentfound True
05.secrets_management_pattern	no environment file is committed outside tests and examplesenv files 0
05.outbound_timeouts	outbound calls set timeouts, or are fewcalls without timeout 0
05.privileged_workloads	no workload runs privileged or in the host's namespacessites 0 · behind a setting or built here 0
05.credentials_in_deploy_config	deployment configuration references secrets, never holds themfiles 0 · sites 0
05.public_env_secret	only publishable values carry a public prefixnames 0 · sites 0 · not secrets 0
05.source_maps_shipped	production builds do not publish source mapssites 0
05.signing_material	no signing keystore or profile is committedfiles 0
05.dataset_provenance	little data is committeddata files 0 · licence or card files 0
05.license_copyleft_inherited	the project is itself AGPL-3.0, so copyleft components carry no obligation it has not already acceptedown license AGPL-3.0 · copyleft components 3
05.license_vendored_unlicensed	all 6 vendored components state their terms, in a licence file, their own headers or a notices filecomponents 6 · stated in headers or notices 0
05.license_grant_unclear	the licence is recognisable as AGPL-3.0class network-copyleft · license AGPL-3.0
05.committed_native_libraries	no widely used compiled library is committed to the productlibraries 0
04.repository_status	the README's first screen states no end, move or freezestatements 0
02.development_elsewhere	nothing says development happens elsewhereno pr statements 0 · mirror statements 0 · replayed commit share 0.0
05.paid_edition	no directory carries a separate paid-edition licencedirectories 0
04.releases_stopped	releases keep pace with developmentmedian gap days 2.0 · days since last tag 0 · commits last 90 days 5,607
04.release_branch	the newest release was cut from the branch surveyedelsewhere 0 · on branch 76 · tags in window 76

Could not conclude, or did not apply — 20

Not a clean result. Each says why it could not answer here.

Signal	Why
04.test_growth	tests are kept inside the files they test, which a count by file name cannot see
04.eol_runtime	no pinned runtime version to check
04.schema_churn	no schema migrations to read
06.print_logging	a frontend: console output is the browser's, not the server's
06.alert_rules	no alert rules in the repository; they may live in a console
06.deployed_image_floats	no deployment descriptor names an image
06.probes_identical	no Kubernetes probes
06.single_replica	no Kubernetes deployments
06.feature_flags	no feature-flag service; flags may live elsewhere
01.notebook_outputs	too few notebooks
05.migration_reversibility	too few migrations to judge
05.resource_limits	no Kubernetes workload manifests
05.cleartext_traffic	not a mobile app
04.flags_accumulating	the history scanned covers 2 months, less than the 12 this question needs
04.api_versions_coexisting	the history scanned covers 2 months, less than the 18 this question needs
04.deprecations_lingering	the history scanned covers 2 months, less than the 24 this question needs
04.dependency_churn	the history scanned covers 2 months, less than the 24 this question needs
04.ci_platforms_dropped	the history scanned covers 2 months, less than the 24 this question needs
05.project_advisories	record incomplete: no advisory found in the sources that answered; OSV PyPI unsloth (an empty record for a package published here) did not
05.project_exploited	no affecting advisory is on CISA's Known Exploited list

Anyone can check this report against the authoritative record — click the stamp, or enter HVQQ-GM9Q-D9QE at borehole.dev/verify.

The code commits to the 106 checks this run performed, the bands they produced, and the commit they were read at. It does not cover checks written since.

Put the mark in your README

The image is wrapped in a link to the verification page, so a reader can check it rather than take it on trust. If you embed it: link it to the report, remove it if the report is delisted, and do not present it as a certification or an endorsement.

[![Borehole reading, not a certification](https://borehole.dev/badge/unslothai/unsloth/7f32241.svg "Borehole free reading: not for reliance. See the full report.")](https://borehole.dev/verify/HVQQGM9QD9QE?via=badge)

Where this scan ran

This scan ran on ordinary cloud infrastructure (cloud-run), not in a hardware-isolated confidential VM. Nothing about it is attested by Google.

What an attested survey proves, and what it does not.

What this scan did with the repository — 7 steps

Written by the steps themselves as they ran, not described afterwards. A hand-written account of a data flow drifts the first time somebody changes the flow and not the account.

At	Step	Detail
0.0s	read repository metadata from the GitHub API	size mb: 609.2authenticated: False
95.8s	cloned the repository to a temporary directory	depth: full
150.8s	read the git history	commits: 11953tracked files: 9151
721.4s	ran the deterministic checks	blame pass: Falseread working tree: False
721.5s	recorded where this scan ran	attested: Falseenvironment: cloud-run
1163.2s	kept the facts, never the source, for re-reading under new checks
1163.2s	deleting the working copy

Total 1163.2s ·source content sent off this machine: none · calls to a language model: 0 · the clone was deleted when the survey ended

An automated reading of a repository's code and its history up to one commit, with public advisory records. It may be wrong, it is not legal, financial, investment or security advice, and it does not replace diligence by a qualified person. It is provided for information: nobody may rely on it (Borehole's terms, section 3).

BOREHOLE 0.7.15 DETERMINISTIC LAYER · NO MODEL CALLSREAD ON BOREHOLE'S SERVER · CLONE DELETED TAKEDOWNCONTRIBUTED TO THIS REPOSITORY? YOUR RIGHTS