Publication record
Corrections ledger
Every claim needs a receipt. Every correction does too. This is the public record of what Bitter Frontier got wrong, what changed, and the primary source that settled it.
Entries
Newest first
-
reviewer availability to embedders ('The CLI's classifier does not come with the crate')
feature 2026-08-23-codex-the-body-not-the-pencil
- Before
- embedders were told the automatic-review path is not part of the embeddable surface
- Corrected
- App Server exposes approvalsReviewer: auto_review ('a carefully prompted subagent'; legacy alias guardian_subagent); the host owns routing, consent, and the audit record
Verdict: wrong as published; the protocol schema names the reviewer route
- Event date
- 2026-08-20
- Reported by
- editor-commissioned outside review
- Record
- run 2026-08-23-codex-study-2025-04-16_2026-08-23-frontier-v0
-
default approval routing ('A model is already in your approval path')
feature 2026-08-23-codex-the-body-not-the-pencil
- Before
- guardian_approval default_enabled:true read as a model sitting in the approval path of every default install
- Corrected
- the capability is stable and enabled; App Server approval routing 'Defaults to `user`'; a model reviews only where auto_review / Approve for me is selected or managed policy requires it
Verdict: capability-vs-routing conflation; schema description is explicit
- Event date
- 2026-08-20
- Reported by
- editor-commissioned outside review
- Record
- run 2026-08-23-codex-study-2025-04-16_2026-08-23-frontier-v0
-
existence and status of the canary/npm pre-release channel
profile paperclip (avoid_for, watch_next, and an open question)
- Before
- Anyone who needs a pre-release channel to track: there is none, the npm/canary line has been dead since 2026-03-12; referred to twice more as the dormant npm line
- Corrected
- the canary line has published continuously since 2026-03-17 -- 1,058 tags through 2026-08-18 and 1,110 npm publishes on the 2026.x line with no gap -- and on 2026-08-17 it was formalised as the first stage of a canary/nightly/beta/stable train
Verdict: wrong-fact, and wrong in the direction that costs a reader something: an operator who wanted a pre-release channel was told none existed. Cause recorded because it will recur -- the canary tags live under slash-prefixed namespaces (refs/tags/canary/...), which the flat repository tag listing does not surface, so an endpoint that was structurally incapable of showing them was read as evidence they were absent. Confirmed by git/matching-refs/tags/canary/ returning 1,058 refs from v2026.3.17-canary.3 to v2026.818.0-canary.7, and by the npm packument.
- Event date
- 2026-03-17
- Reported by
- adversarial-verify
- Record
- run 2026-08-17-weekly-digest-2026-08-10_2026-08-17-frontier-v0
-
completeness of the social sweep, and four claims that outran their receipts
digest 2026-07-02_2026-07-27-weekly (research-method description and four claims)
- Before
- 'we swept public X discussion across the whole watchlist ... 314 public claims, each carrying a post URL and a date, each adjudicated against the primary record', plus 'Every release that hardened something was posted within minutes', 'they are precisely the ones that loosened a control', 'our thesis leans on the idea that platform-hosted operating environments are where durable authority should live', and 'Claude Code carries the most rows in our evidence not because it is the worst offender but because Anthropic writes the most down'
- Corrected
- The sweep covered thirteen of fourteen projects: the eve run failed with an API timeout and returned nothing, and flue's twenty-six claims were never adjudicated. Hardening releases were posted at lags from six minutes to about a day. Three of the five unannounced releases can be characterized as loosening a control, not all five. The platform-governance claim was this publication's own one-window reading in the prior issue, not a standing thesis, and the link implying otherwise is removed. The documentation comparison is stated as undecidable, which is what the issue's own argument requires.
Verdict: overclaimed against our own evidence -> corrected. Found by a cold read against the house bar the same day the issue published. The method claim is the serious one: an issue arguing that stated controls do not match executing behavior described its own research as more complete than it was.
- Event date
- 2026-07-27
- Reported by
- adversarial-verify
- Record
- run 2026-07-27-weekly-digest-2026-07-02_2026-07-27-frontier-v0
-
the comparative reading of platform-hosted versus self-hosted governance
digest 2026-06-24_2026-07-01-weekly (eve as the governance counterexample)
- Before
- Where the self-hostable harnesses shipped their governance fixes to main-but-not-tagged, the platform-hosted one shipped containment to its runtime ... This window, the platform side was the one whose controls you could actually run.
- Corrected
- The reading overstated what shipping a control to a runtime guarantees. Across 2026-07-02..2026-07-27, four human gates inside eve did not bind: 0.21.0 guaranteed a delegation tree could never outspend its root budget and 0.27.1 fixed that declining a child let the parent retry it against a fresh quota share, fifteen days later; 0.26.1 made the developer interface show rejected tool approvals as denied instead of successful; 0.25.0 introduced the guarantee that a stale approval never authorizes an earlier call, which means it previously could; and 0.24.2 fixed approvals that left the tool unrun. Platform-hosted governance is not a substitute for testing the gate.
Verdict: overstated reading -> corrected (the contrast between platform and self-hosted governance was drawn too cleanly. Evidence complicating it existed in our own record when the reading was published: a comparable eve instance dated 2026-07-01 and an approval-surface caveat flagged 2026-06-17. The 2026-07-02..2026-07-27 digest carries the reversal in the body rather than a footnote, and the eve profile now opens by naming it.)
- Event date
- 2026-07-22
- Reported by
- adversarial-verify
- Record
- run 2026-07-27-weekly-digest-2026-07-02_2026-07-27-frontier-v0
-
like-for-like Microsoft comparison
finding 2026-07-02-agent-flywheel-arxiv-solo-author-velocity and profile agent-flywheel
- Before
- 110 for dicklesworthstone versus 52 for Microsoft, combining counts from different tables
- Corrected
- 110 for dicklesworthstone versus 97 for Microsoft in the same top-35 figure
Verdict: apples-to-oranges comparison -> corrected to counts from the same figure
- Event date
- 2026-06-05
- Reported by
- adversarial-verify
- Record
- run 2026-07-12-exemplar-pass-agent-flywheel-frontier-v0
-
v0.7.0 safe-mode agent-confirmation boundary
- Before
- safe mode keeps standard agent confirmations
- Corrected
- safe mode still installs shared cc, cod, agy, and gmi shortcuts that select dangerous or always-proceed behavior
Verdict: implementation mismatch -> corrected against tagged shell config
- Event date
- 2026-06-26
- Reported by
- adversarial-verify
- Record
- run 2026-07-12-exemplar-pass-agent-flywheel-frontier-v0
-
v0.7.0 safe-mode passwordless-sudo behavior
- Before
- safe mode omits or removes passwordless sudo
- Corrected
- on a fresh host, safe mode does not create ACFS's own NOPASSWD sudoers file; this branch does not revoke a rule left by an earlier vibe installation or configured elsewhere
Verdict: overbroad safety claim -> narrowed to the tagged installer's actual branch behavior
- Event date
- 2026-06-26
- Reported by
- adversarial-verify
- Record
- run 2026-07-12-exemplar-pass-agent-flywheel-frontier-v0
-
event date and eligibility for the 2026-07-01..2026-07-02 window
signal 2026-07-02-agent-flywheel-assembly-layer-dangerous-defaults
- Before
- counted as a July 2 signal and described as the live-window case
- Corrected
- withdrawn from the accepted signal count; ACFS v0.7.0 shipped 2026-06-26 and is retained only as a pre-window intake baseline
Verdict: out-of-window -> withdrawn (source intake is not an upstream product event)
- Event date
- 2026-06-26
- Reported by
- adversarial-verify
- Record
- run 2026-07-12-exemplar-pass-agent-flywheel-frontier-v0
-
lifecycle status of consumer Gemini CLI
digest 2026-06-16_2026-06-23-weekly (Gemini Antigravity succession)
- Before
- framed as prospective -- Gemini CLI 'may be entering a managed succession,' with the Antigravity funnel and banner in stable
- Corrected
- the consumer service was already discontinued. Per Google's announcement, Gemini CLI stopped serving requests on 2026-06-18 for AI Pro/Ultra, free individual Code Assist, and new GitHub-org installs; enterprise Code Assist retained access and the OSS repo continues. The 06-16..06-23 window called an executed shutdown prospective.
Verdict: under-called -> corrected (the shutdown executed inside the 06-16..06-23 window; recorded as accomplished in the 2026-07-01 digest, with Antigravity registered as a separate source)
- Event date
- 2026-06-18
- Reported by
- adversarial-verify
- Record
- run 2026-07-01-weekly-digest-2026-06-24_2026-07-01-frontier-v0
-
scope of the flash-to-3.5 routing change
signal 2026-06-10-gemini-flash-ga-routing
- Before
- framed as an unconditional silent switch of flash workloads to gemini-3.5-flash
- Corrected
- gated behind an experiment flag plus auth-type access logic; stable v0.46.0 ships a placeholder flag ID (the real-flag-ID commit is not in v0.46.0)
Verdict: out-of-window-as-stated -> corrected (the change is real and in-window; the overstatement was scope, not date)
- Event date
- 2026-06-10
- Reported by
- adversarial-verify
- Record
- run 2026-06-16-weekly-digest-2026-06-04_2026-06-16-frontier-v0
-
channel / scope of "Antigravity migration commands + built-in skill"
digest 2026-06-04_2026-06-16-weekly (Gemini Antigravity migration)
- Before
- framed as shipping (migration commands and an Antigravity skill present)
- Corrected
- the transition banner is in stable (v0.45.2, uncapped); the migration commands and Antigravity skill are preview-only (v0.47.0-preview.0)
Verdict: channel-mismatch (resolved by git ancestry: commit in v0.47.0-preview.0 only, not in any stable tag)
- Event date
- 2026-06-05
- Reported by
- adversarial-verify
- Record
- run 2026-06-16-weekly-digest-2026-06-04_2026-06-16-frontier-v0
Last updated 2026-08-24. Entries are stored with the publication, not in a separate editorial database.