Skip to content

docs: alpha.55 accuracy audit fixes (Codex + Cursor audit, lead-verified) - #114

Merged
AlbSar merged 10 commits into
mainfrom
docs/a55-audit
Oct 24, 2026
Merged

docs: alpha.55 accuracy audit fixes (Codex + Cursor audit, lead-verified)#114
AlbSar merged 10 commits into
mainfrom
docs/a55-audit

Conversation

@AlbSar

@AlbSar AlbSar commented Oct 24, 2026

Copy link
Copy Markdown
Contributor

Summary

Docs-only follow-up to alpha.55. The owner ordered an accuracy audit after the release, and this PR applies its verified fixes. Two read-only auditors did the audit, Codex (gpt-6.1-sol) and Cursor. The lead then checked every finding against the code. Four fix agents changed only the Markdown in their own disjoint file groups.

  • Release record: alpha.55 is marked released (2026-10-24, PR alpha.55: pre-beta core layers (memory, recovery, model roles, agents, signers, headless Brain, quota-ask) #113, live 09:36Z, N1 09:38Z) in CHANGELOG, PLAN and COMPLETED-PLAN. The 2026-10-24 owner decisions are added in the owner's own words.
  • Current behaviour now described correctly:
    • Policy upgrades have two paths. A template policy uses policy upgrade --template current. A hand-written policy uses init policy --upgrade --person.
    • Shadow mode applies nothing, but memory verification and recovery shadow do call their model.
    • The Brain signs in with models login, then models connect, and keeps its login in a Deckent-owned directory.
    • The policy-turn record is written at most once per turn, on a best-effort basis.
    • A known gap is stated: alpha.55 has no command to promote a Nervous rule.
  • Twins and readability: each EN/TR pair now says the same thing. The README CLI list and the glossary gained the new core layers. The architecture overview is stamped alpha.55.
  • Preserved text: the 28 replaced PLAN/ARCHITECTURE lines are kept word for word in the plan archive.

Evidence

  • Audit and verification files are kept outside the repo, in proof/DOCS-AUDIT-ALPHA55/: codex-report, codex-verified, cursor-report, cursor-verified and FIX-BRIEF.
  • Checks passed: lint-docs and lint-arch with 0 issues, and check-doc-preservation --ref f1aba6e28 found all 28 removed lines. The docs-script-references and readme-commands tests passed 5/5.
  • Review: Astra asked for revisions twice (REVISE 2748: the glossary's policy-turn guarantee and the PASS scope versus the migration proof). Both are fixed, and Astra passed review 2750 on this exact SHA.

Risk

Markdown only: no source, test, config or generated file changes. This is a two-way door, and there is no release or live switch; the docs reach the live install with the next release.

Code items the audit found are left out of this PR and tracked separately:

  • a failed policy-turn write gives no signal;
  • the Turkish glossary does not match the catalogs (an owner choice);
  • the i18n key remember.db_not_found is dead;
  • some source comments are stale.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob

AlbSar and others added 10 commits October 24, 2026 13:28
@AlbSar @claude
docs(alpha.55): operator reference and known limits match the alpha.5…
b254780 
…5 source (G3 audit rows)

Applies the CONFIRMED-DOC rows of the alpha.55 docs audit for the operator reference and known limits,
EN and TR twins with the same meaning: policy template v15 with both upgrade paths (template-installed
via `policy upgrade --template current`, hand-written via `init policy --upgrade --person`), the
claude-cli-subscription kind and its login/connect commands and Deckent-owned login directory, model
precedence, subscription Brain availability, acceptance wording, mouse selection, `/runs` `a`, switched
layers and the shadow wording (memory-verification and recovery shadow call their model), quota rewaits,
image tagging, remote model servers with a registered tariff, MCP approve on a terminal, acceptance
profiles not scored, current ledger v70, the Nervous promotion gap (no command in alpha.55, recovery
shadow yields no data), shadow verification egress, per-Run cost report, and stale/shadow suggestions.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
docs(alpha.55): G2 architecture module docs match the alpha.55 code (…
f8aa685 
…docs audit)

Apply the lead-verified CONFIRMED-DOC rows for approval-and-delivery,
model-invocation-billing and the architecture modules: shadow applies
nothing but memory-verification and recovery shadow call their model;
both template upgrade paths (policy upgrade / init policy --upgrade
--person); policy-turn record is at most one per turn, best effort;
Nervous has no promotion command in alpha.55 (card in alpha.56); MCP
default realm sandbox-net with the project read-only; secret set needs a
terminal; quota wait/ask numbers; selected Brain template v4; CI job
counts and the 2026-10-18 landing rule; terminal idle stop, local-zero
budget, /runs answer key, typed-argument exceptions, alpha.54/55 additions;
pointer-style notes for superseded dated paragraphs; spaces restored in
the cancelled Qwen pilot record. No source change.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
docs(alpha.55): G1 docs audit fixes — release record, shadow egress, …
01e6fe9 
…policy-turn bound, Nervous promotion gap

Applies the CONFIRMED-DOC rows of proof/DOCS-AUDIT-ALPHA55/codex-verified.md for the root and history files plus
the lead decisions of FIX-BRIEF.md. ARCHITECTURE: first fence written by the live Nervous broker but no
rule-promotion command in alpha.55; memory states candidate → verified → retired; shadow applies nothing yet
memory-verification and recovery shadow call their model; at most one best-effort policy-turn record per deciding
turn; KARAR 12 closure pointer; Codex client-reported evidence; partial discovery notice; loopback zero-tariff
exemption; 23 tracked skills. PLAN: alpha.55 release record and gate 5, R1/migration closed, alpha.54 items live,
K-B standing list, 19 plan-bound areas, KARAR 12 closed, stale known-limits notes dropped, 2026-10-24 index line.
CHANGELOG: alpha.55 released (2026-10-24, f1aba6e, PR #113, live/N1, switch, ledger v70), both template
upgrade paths, shadow/policy-turn wording, Nervous gap, CI shard-1 reserve. COMPLETED-PLAN: alpha.55 entry.
owner-decisions: 2026-10-24 section (Brain login dir s3c, telemetry direction, Hub pointer, merge+switch,
no-MVP owner wording). decisions.md: 2026-10-24 s3c and core layers rows, 2026-10-22 LINT-ARCH-DYNAMIC row.
Replaced PLAN/ARCHITECTURE lines kept verbatim in the plan archive (check-doc-preservation --ref f1aba6e: 28/28).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
docs(alpha.55): README, architecture overview and glossary (EN/TR) ma…
91a9b46 
…tch alpha.55 source (G4, Cursor audit verified)

- Full access: switched on for the session or saved as start mode (`/mode start full-access`); README table and glossary.
- README everyday CLI names `deckent memory`, `deckent agents`, `deckent models roles show`.
- Core layers: every switch (not every layer) starts off; shadow applies nothing but memory-verification and recovery shadow
  still call their model; no Nervous rule can be promoted in alpha.55, so automatic recovery cannot act yet; the new rights
  come from `policy upgrade --template current` or, for a hand-written policy, `init policy --upgrade --person`.
- Architecture overview stamp: 2026-10-24, package 1.0.0-alpha.55.
- Glossary: short Core layers section (Brain, WorkClass, assignment, memory journal, Auditor, Nervous, recovery, agent catalog).

Verification table: proof/DOCS-AUDIT-ALPHA55/cursor-verified.md (outside the repository).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
merge docs/a55-g1
6efc842 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
merge docs/a55-g2
b2646d6 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
merge docs/a55-g3
dc35934 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
merge docs/a55-g4
587912d 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
docs: REVISE 2748 — glossary policy-turn is at most one best-effort r…
c768a78 
…ecord (EN/TR); PASS 2744 scope kept separate from the lane migration report

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
@AlbSar @claude
docs: exact wording for the post-switch keyed audit verify
3fb636f 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JrXPpzy6N7zFwchFM6Adob
claude[bot]
claude Bot reviewed Oct 24, 2026
View reviewed changes

@claude claude Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

Tip: disable this comment in your organization's Code Review settings.

@AlbSar
AlbSar merged commit 2400b57 into main Oct 24, 2026
37 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@claude claude[bot] claude[bot] left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@AlbSar