ISSUE 392 + agent independence discipline#45
Open
isaacsight wants to merge 3 commits into
Open
Conversation
A critical pass exposed that this session had one model author an issue
and then "audit" its own design, language, and imagery and report it all
passed — the audits were never independent. This adds the discipline to
prevent reviewer theater.
INDEPENDENCE.md (new) is the governing doc: the one test ("what does this
agent know, see, or optimize that the author didn't?"), the five
independence sources (different-model, evidence, adversarial, different
criteria, human-gate), the autocorrelation trap (a same-model "team" is
one model agreeing with itself N times, not N reviewers), a
Mechanical-vs-Judgment taxonomy, a mandatory "Independence source:"
header for every agent, and the heterogeneity policy: exploit kbot's
multi-provider BYOK to run judgment and specialist review agents on a
different provider's model than the author (e.g. Japanese review on a
JP-native model), routed via per-agent config rather than hardcoded.
japanese-editor.md is rewritten as the worked example: it declares its
independence source, adds an evidence-grounding protocol step (usage,
corpus, dictionary, and precedent checks instead of parametric opinion),
and changes its verdict to PASS-TO-HUMAN / NEEDS-REVISION — it escalates
to a native reader, it never certifies.
designer.md and reviewer.md get the Independence-source header as the
start of rollout: mechanical findings are trustworthy through tools (cite
the tool or number), but taste and judgment findings are flagged as
non-independent on the author's own model unless run on a different model
or staged for a human. Self-agreement is not assurance.
https://claude.ai/code/session_01X9H4TPZ1aT8Ep4TCkiDzfv
Completes the independence discipline started in the prior commit. Rollout: the "Independence source" header is now on all 53 agents (previously only designer/reviewer/japanese-editor). Each is classified and declares its source honestly: - mechanical — tool is the independence (qa, debugger, deployer, devops, ship, github, environment, hardware, performance, pulse, sync, autopoiesis, stream-auditor); cite the tool, don't assert from memory. - mixed/adversarial — security, hacker, immune: tools + hunt-to-reject, prefer a different model for novel-attack reasoning. - mixed/evidence — documenter, obsidian-sync, rival-intel, admin, replit, speed-of-light: ground claims in artifacts. - judgment — taste/strategy/creative agents (pixel-artist, product, playtester, the strategy set, video/creative set, architects): NOT independent on the author's model; run on a different-provider model via kbot BYOK or stage for a human; never report self-agreement as assurance. - authoring — magazine-editor: independence comes from downstream review plus a human, not self-check. Proof: re-ran ISSUE 391's Japanese through a genuinely different model (GPT-OSS-20B, different provider/weights). It caught five strings the original same-model self-review had passed — demonstrating the autocorrelation the doctrine predicts. It also false-flagged an intentional house phrase, confirming the converse: a decorrelated model surfaces candidates, but a native human is the final authority. 391's Japanese is unchanged; the candidates are staged for human adjudication, exactly as japanese-editor.md now prescribes. https://claude.ai/code/session_01X9H4TPZ1aT8Ep4TCkiDzfv
A deliberate restoration of range. 388–391 ran four issues deep on agentic AI and the magazine was drifting toward a trade newsletter; 392 returns to the culture/craft format the publication was built on (360–366) without cutting away from the thread. 391 argued that taste and the hand are what survive the machine; 392 stops arguing and goes and looks — at the analog objects a coder keeps beside the screen (notebook, pen, keyboard, cup), and at choosing them by hand as the same scarce discernment the work itself now turns on. Continuity in the argument, not the topic. Identity: essay (field-piece, 369 template), kraft stock + classic layout (breaks the asymmetric-left streak of 389–391) + coffee accent + a BY HAND seal. Dossier reads as an inventory of the six inches around the screen. Pull quote: "The screen is rented. The desk is yours." Registered in index.ts; typecheck + build clean. Back cover (public/back-covers/392-desk.jpg): a Flux/Pollinations still-life of the desk objects on kraft — the issue's subject made literal — passing the designer.md acceptance bar. The Japanese was run through a genuinely different model (GPT-OSS-20B) before shipping, per the independence discipline. Its findings were adjudicated, not blindly applied: most alternatives were wrong for house register and the JP is kept as written. One signal is escalated for a native-human decision — it flags the house construction 〜の用 (also in 369) as unidiomatic, which legitimately questions canon. Staged, not auto-changed. https://claude.ai/code/session_01X9H4TPZ1aT8Ep4TCkiDzfv
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Brings the branch's post-#44 work into main so ISSUE 392 (already deployed to gh-pages) is permanent and main stays in sync with the live site.
Contents
Agent independence discipline (
82c3f31,ecf5ccd)INDEPENDENCE.md— the governing doc: the one test, the five independence sources, the autocorrelation trap, Mechanical/Judgment taxonomy, mandatoryIndependence sourceheader, and the kbot-BYOK heterogeneity policy.japanese-editor.mdrewritten as the worked example (evidence-grounded, escalates-to-human, PASS-TO-HUMAN verdict).Independence sourceheader rolled out to all 53 agents.ISSUE 392 — THE DESK THAT SURVIVES (
ee28215)〜の用) staged for a native human.Verification
npx tsc --noEmit+npm run buildclean.Open item (not blocking): the
〜の用construction in 369 + 392 awaits a native-speaker decision.https://claude.ai/code/session_01X9H4TPZ1aT8Ep4TCkiDzfv
Generated by Claude Code