v1 principles audit¶

This report tests the repo against its own principles P1 to P12 before the v1.0.0 tag.

Date: 2026-10-19.
Base: origin/main at commit 8e16299 (after PR #89).
Sources: the framework files, .github/workflows/, .cornerstone/, stone/src/, the 61 merged PRs, the open issues, and the GitHub repo settings.
This report changes no code and fixes no gap.

Result¶

The repo follows 1 principle, partly follows 9, and has a gap in 2 (P3 and P5).

Principle	Verdict	Reason
P1 Checks decide	Partly	CI runs the checks, but K7, K14, and K17 only warn, and no check is a required status check.
P2 Writer and judge	Partly	K6 blocks implementer commits to protected paths. It trusts the `Role:` trailer and does not limit the tester.
P3 Judgment into inputs	Gap	No issue has all Definition of Ready fields, and no deterministic G1 check exists.
P4 Context budget	Partly	The files are small, but no check in this repo tests the instruction file or the skill size.
P5 Each component earns its place	Gap	All 10 manifest entries have `models_tested: []`. Few components have an eval. No job finds a past review date.
P6 Enforce what you can state	Partly	Many rules are checks, but some precise rules in `AGENTS.md` and the repo config have no check.
P7 Ceremony scales with risk	Follows	`stone check` rejects a profile without a plan form or valid gates, and compares them with `gates.md`.
P8 State in files	Partly	Decisions, the roadmap, and eval results are in files. No decision-log file was ever committed.
P9 Bounded autonomy	Partly	The workflows have minimum permissions and pinned actions. The `main` ruleset is disabled, and a direct push occurred.
P10 Volume is a cost	Partly	The limit is 600 lines. 11 of 36 PRs after K7 were over it, and K7 rejects 3 of their waivers.
P11 Make the work visible	Partly	K14 runs on each PR in warn mode. 16 of 44 PRs after K14 (PR #18) fail it.
P12 Neutral core	Partly	`stone check` passes, but D40 names two spec tools in the core, and `render.py` names a harness folder.

P1. Checks decide¶

Verdict: Partly.

Evidence	Finding
`.github/workflows/cornerstone-checks.yml:32`, `:49`	K7 and K14 run with `--mode warn`.
`.github/workflows/stone.yml:39`, `:47`	K17 runs with `--mode warn`. The roadmap moves it to block on 2026-10-31 (issue #62).
`.github/workflows/cornerstone-checks.yml:56`	The "All checks" step runs `stone check-writing` with no `--mode`. The default is `warn` (`stone/src/stone/cli.py:48`).
GitHub ruleset `main` (id 25320579)	The ruleset is `disabled`. It has only the `deletion` and `non_fast_forward` rules. No check is a required status check.
PR #89	K7 reported a warning (1149 lines, no waiver). The PR merged.

What works:

G3 runs in CI, outside the agent session (cornerstone-checks.yml, stone.yml).
K6, the unit tests, and stone check fail the job.
Each merged PR from #2 to #89 has only successful or skipped check runs. PR #1 has no check runs.

When the warn checks block:

Check	Where	Block date in the roadmap	Issue
K17	`stone.yml`	2026-10-31	#62
K14	`stone.yml`	2 weeks after the K14 step is in that workflow. The step is not in `stone.yml` today.	#63
K14, K7, K15	G3 template	2 weeks after stone 0.7 (PR #23), K7 (PR #26), and K15 (PR #27). All merged on 2026-10-18, so the earliest date is 2026-11-01.	#63
K4, K5, K8, K10, K11, K12	G3 step templates (`continue-on-error: true`)	2 weeks after PR #30	#63
K18, K19	`stone check`	2 weeks after their merge	#64

The v1 exit criteria (framework/README.md:194) need K6, K7, K14, K15, and K16 to block. Thus the tag cannot be on time before 2026-11-01.

Other findings:

The G3 workflow installs stone from the tag v0.5.0 (cornerstone-checks.yml:37). The framework_version in .cornerstone/config.yaml:2 is 0.5.0, but stone is 0.9.0. K14 in G3 runs the PR #18 code.
The single-ticket profile lists K1, K2, and K9 (framework/profiles/single-ticket.md:17). This repo has no lint, type check, or complexity command (.cornerstone/config.yaml:10). So G3 does not run them, and waivers: [] records no reason.

P2. Separate the writer from the judge¶

Verdict: Partly.

Evidence	Finding
`.cornerstone/checks/protected_paths.py:26-29`	K6 reads the last `Role:` trailer. Any value except `implementer` skips the commit. A false trailer passes (issue #81).
`framework/checks.md:66` and `protected_paths.py:28`	The doc and the code disagree. Rule 1 says a check is read-only for the implementer and the tester. K6 limits only the implementer. A `Role: tester` commit can change `.cornerstone/checks/`.
`framework/README.md:102` (D39)	The hook does not see a shell write. K6 is the only control in CI, and only for implementer commits.

Commit trailers on main (206 commits that are not merges):

Trailer	Count
`Role: implementer`	86
`Role: tester`	54
`Role: generator`	22
`Role: maintainer`	2
No `Role:` trailer	42
`Actor: agent`	162
`Actor: human`	2

Each commit that changes stone/tests/ has Role: tester or no trailer. No implementer commit changes a test.
8 commits with no trailer change a K6 path. Two are squash merges (PRs #51 and #52), so the Role: trailers of the PR commits are lost on main. The other six are from PR #7 or earlier, before K6 ran in CI.
The trailers do not show a different actor. Each tester and implementer commit has Actor: agent, and no trailer names a session.
P2 says the pipeline MUST start the tester and the reviewer (framework/principles.md:39). The only pipeline in this repo is the label workflow. It starts the shaper only (.github/workflows/cornerstone-single-ticket-label.yml:48). A human or a parent session starts the tester and the reviewer.

P3. Judgment goes into the inputs¶

Verdict: Gap.

Evidence	Finding
Issues #38 to #85 (28 issues)	0 of 28 have all the Definition of Ready fields. 23 have acceptance criteria. 2 have non-goals. None has a test oracle or a `Personas:` line.
`framework/gates.md:32`, `.github/`	Single-ticket sets G1 to `check`. No deterministic G1 check exists. The repo has no issue template.
`cornerstone-single-ticket-label.yml:66-71`	The label workflow asks a model (the shaper) to list the missing fields. That is a report, not a check.

The Definition of Ready is in framework/workflow-model.md:92. Non-goals and the test oracle are on lines 98 and 99.

P4. Context is a budget¶

Verdict: Partly.

Evidence	Finding
`AGENTS.md` (85 lines), `CLAUDE.md` (1 line)	The instruction file is under the 300-line limit.
`plugins/workflow/skills/*/SKILL.md`	The largest skill has 130 lines (`dev-pr`). Each description has 474 characters or fewer. The model profile limit is 500 lines.
`framework/principles.md:64`	The Enforce item needs a size check and a skill lint. K15 is not rendered here, because single-ticket does not list K15. `stone check` has no skill lint.

AGENTS.md repeats facts that are also in the repo, for example the commands in .cornerstone/config.yaml:11-13. The rule says that the file MUST contain only facts that the agent cannot find. This is a small cost at 85 lines.

P5. Each component earns its place¶

Verdict: Gap.

Evidence	Finding
`.cornerstone/manifest.yaml:12` and the 9 other entries	Each entry has `models_tested: []`. The `role.tester` eval ran on a model (`evals/results/20261018-085807-role.tester-claude-opus-5-5.md`), but its entry is empty.
`framework/adapters/claude-code/components.yaml`	17 component entries. Eval folders exist for 3 of them (`role.tester`, `role.planner`, `hook.protect-paths`). `hook.settings` shares the hook folder. Only the 2 role evals have a result.
`framework/principles.md:74`	No scheduled CI job opens an issue for a component past its review date. The first `review_by` is 2027-04-17.

Other findings:

The task names 22 components. I count 17 entries in components.yaml. With the 6 G3 steps, the count is 23. The owner can confirm the correct base.
Each manifest entry has an owner (brawblan), an eval key, and a review_by date. stone check tests the eval key and the date (D43). It does not test owner or models_tested.
framework/writing.md:64 says the manifest has the entry rule.writing-standard. The manifest and components.yaml have no such entry. The doc and the code disagree.
The plugin skills have no manifest entry. AGENTS.md asks for an eval plan and a review date for each new skill. Only cornerstone-setup has one, with the status "Planned. No eval has run."

P6. Enforce what you can state¶

Verdict: Partly.

These rules are checks: K6, K7, K14, K17, the profile and gate rules, the D34 branch-name rule, and the D43 component rule.

These precise rules have no check:

Rule	Where	Possible check
Pin each action to a commit SHA	`AGENTS.md:66`	A grep for `uses:` without a 40-character SHA. All pins pass today.
Pass event text through `env`	`AGENTS.md:67`	A scan of `run:` for `github.event`. All steps pass today.
Commit format `type(scope): summary`	`AGENTS.md:74`	A commit-message check
Increase the plugin version when a skill changes	`AGENTS.md:60`	A diff check
Never push to `main`	`AGENTS.md:73`	The disabled ruleset (see P9)
PR sections Summary to Rollback plan	`AGENTS.md:76`	K14 tests 4 other sections only
`function_lines: 60`, `cyclomatic: 10`	`.cornerstone/config.yaml:21-22`	K9. An approximate AST count finds 1 function over 60 lines and about 21 over a complexity of 10 in `stone/src/`.

The manifest has no instruction entry, so the "why not a check" field (framework/principles.md:83) has no entry to test.

P7. Ceremony scales with risk¶

Verdict: Follows.

Evidence	Finding
`stone/src/stone/framework_check.py:86-116`	`stone check` fails a profile without `plan_form` or `gates`, with a wrong gate value, or with a gate that differs from `gates.md`.
`framework_check.py:107`	A profile that builds code for a PR must set G3 to `required`.
`stone check` on this branch	0 findings

Note: render and check do not read overrides or waivers from the repo config (issue #78). A repo can record a waiver to remove a gate, but no tool reads it.

P8. Small units, fresh sessions, state in files¶

Verdict: Partly.

Evidence	Finding
`framework/README.md:60-106`, `:112-198`	Decisions D1 to D43 and the roadmap are in a versioned file.
`evals/results/`	The eval results are committed files (D36).
`framework/adapters/claude-code/templates/agents/implementer.md:25`	The implementer writes the decision log at `docs/plans/<ticket-id>.decisions.md`. No such file was ever committed. The decision logs are in the PR bodies only.

The 86 implementer commits did not write a decision-log file. Either the sessions did not use the role file, or the file had no effect. I did not read the transcripts, so I cannot say which.

P9. Bounded autonomy¶

Verdict: Partly.

Evidence	Finding
Ruleset `main` (`enforcement: disabled`)	No PR requirement and no required check on `main`. Commit `94c5ebc` and its revert `2a5b2e7` went to `main` directly on 2026-10-17, after PR #2.
`.claude/agents/reviewer.md:4`, `.claude/agents/shaper.md:4`	Both roles have the shell tool. The reviewer "changes no file", but the shell can write (D39).
`cornerstone-single-ticket-label.yml:54`, `:77`, `:101`	The trigger role has the model secret, the untrusted issue text, and a public report. D32 limits the reads, and D38 frames the report.

What works:

Each workflow sets top-level permissions to contents: read or {}. Write scopes are on single jobs only.
Each uses: in .github/workflows/ and in the templates is pinned to a 40-character SHA.
Event text reaches each run: step only through env.
The trigger role has a turn limit and a job time limit (D38).
Platform secret scanning and push protection are on. They cover K13.

P10. Volume is a cost¶

Verdict: Partly.

Evidence	Finding
`.cornerstone/config.yaml:20`, `.cornerstone/checks/change_size.py:16`	The limit is 600 changed lines.
PRs #27 to #89	11 of the 36 PRs after K7 (PR #26) are over 600 lines.
`change_size.py:47`	K7 accepts only a heading that starts with "Waiver (P10)". PRs #35 and #57 use "Size waiver (P10)". PR #89 has the waiver in a list item. K7 accepts none of the three.

Over-limit PRs before K7: #1, #16, #18, #21, #22, #24, and #25. Their waivers do not use the K7 format, because K7 did not exist.

PR #89 is the example from the task. K7 warned: "The diff has 1149 changed lines. The limit for single-ticket is 600." The body has no Profile: line. Its waiver is a list item, and the PR merged.

If K7 blocks with the current parser, it fails PRs that a human accepted. Fix the parser or the waiver format first.

P11. Make the work visible¶

Verdict: Partly.

Evidence	Finding
`cornerstone-checks.yml:44-49`	K14 runs on each PR in warn mode.
`stone check-evidence --mode block` on each merged PR body	16 of the 44 PRs from #19 to #89 fail. Examples: #68 to #73, #75, #76, #83, #84.
PRs #86, #87, #88, #89	The 4 latest PRs pass K14.

Most failures are docs-only PRs (the guides). This data can help the owner decide whether K14 needs a docs exception before it blocks.

P12. Neutral core, thin adapters¶

Verdict: Partly.

Evidence	Finding
`stone check`	0 findings. The check covers `*.md` in `framework/` outside `adapters/`.
`framework/README.md:103` (D40)	The core names two spec tools (Spec Kit and BMAD). The name list in `stone/src/stone/framework_check.py:29-30` does not have them.
`stone/src/stone/render.py:32`	The generator hard-codes the harness folder `.claude/`. The adapter does not give it.

Other findings:

The check reads only *.md files (framework_check.py:135). No other file type is in the core today.
The name list has no model family names. A text search of the core found none.
framework/generator/design.md:154 assumes git and GitHub. P12 does not name the CI host, so this is not a break of the rule.

Dossier citations¶

I picked 5 IDs at random (seed 20261019) from the IDs that the principles, decisions, and checks cite.

ID	Cited by	Exists	Does the source say it?
E1	P8	Yes	Not verifiable. S2 is talk notes with no public text.
B15	P2	Yes	Not verifiable. S1 is talk notes with no public text.
C3	`framework/checks.md:62`	Yes	Not read. S21 is a book. The text agrees with the book's known definition of a fitness function.
F8	D31, D38, the label workflow	Yes	Yes. S35 and S36 say to pass untrusted input through an environment variable. They name the issue title and the PR title. Neither names an issue label.
H2	P3	Yes	In part. The S12 page returned HTTP 403. Search results confirm "Humans steer. Agents execute." I did not confirm the second sentence.

Note: the dossier IDs D1 to D11 and the decisions D1 to D43 use the same letter. P3 cites dossier D1, which a reader can confuse with decision D1.

Gaps for v1¶

Ranked by severity. Each line gives a suggested issue title.

High. The main ruleset has the status disabled, so no check can block a merge. Issue: chore(repo): enable the main ruleset with PR-only merges and required G3 checks.
High. K7, K14, K15, and K17 only warn. The v1 exit criteria need them to block. Issues #62 and #63 exist. The earliest date is 2026-11-01.
High. K7 rejects the waiver forms in PRs #35, #57, and #89. Issue: fix(stone): make K7 accept the waiver forms of merged PRs, or document one exact form.
Medium. models_tested is empty in all 10 manifest entries, and D43 does not test it. Issue: fix(stone): write models_tested from eval results and test owner and models_tested (P5).
Medium. K6 does not limit the tester, but framework/checks.md:66 says it does. Issue: fix(framework): decide if K6 limits the tester (checks.md rule 1 vs protected_paths.py).
Medium. Single-ticket sets G1 to check, but no G1 check exists. No open issue meets the Definition of Ready. Issue: feat(stone): add a G1 check for the Definition of Ready fields.
Low. D40 names spec tools in the core, and the name list misses them. Issue: fix(framework): move tool names out of D40 and extend the P12 name list.
Low. framework/writing.md:64 names a manifest entry that does not exist. Issue: fix(framework): add rule.writing-standard to the manifest, or change writing.md.

Gaps after v1¶

No scheduled job finds a past review_by (P5). Do it before 2027-04-17. Issue: feat(stone): render a scheduled job that opens an issue for each past review date.
13 of 17 components have no eval folder. The hook eval has no model run (roadmap 4g).
This repo has no K1, K2, or K9 command. About 21 functions in stone/src/ are over the complexity limit (approximate count).
No skill lint, and no K15 in this repo (P4).
No decision-log file in docs/plans/ (P8). The owner decides: the file or the PR body.
The pipeline does not start the tester and the reviewer (P2).
K6 trusts the Role: trailer (issue #81). The hook does not see a shell write (D39).
The reviewer and the shaper have the shell tool (P9).
Squash merges lose the Role: trailers on main (PRs #51, #52).
G3 installs stone from v0.5.0. framework_version did not move with stone 0.9.0.
The P12 check reads only Markdown and has no model family names.
The dossier IDs D1 to D11 and the decision IDs share a letter.
Sources S1 and S2 are talk notes with no public text. B15 and E1 rest only on them.

What I did not check¶

Session transcripts. I cannot say if each session did one unit of work or if a tester ran in a fresh session.
The reviews on each PR, and who approved G4.
The eval runner, evals/, and stone/tests/. Another session works on them.
An eval run. I ran no model.
The full dossier. I tested 5 of 74 cited IDs.
The text of S1, S2, and S21, and the full text of S12.
Branch rules for branches other than main, and the GitHub App settings.
The other 9 profiles in a target repo. This repo renders single-ticket only.
Whether the guides in docs/guides/ match the code.