Skip to content

docs(adr): cut-choice contract (0034) + judge tier (0033) - #298

Merged
thejefflarson merged 3 commits into
mainfrom
cut-choice-contract-adr
Jul 29, 2026
Merged

docs(adr): cut-choice contract (0034) + judge tier (0033)#298
thejefflarson merged 3 commits into
mainfrom
cut-choice-contract-adr

Conversation

@thejefflarson

Copy link
Copy Markdown
Owner

Two paired ADRs for the cut-choice work:

  • ADR-0034 — the cut-choice contract: the model names the compromised on-path nodes ({assessment, reason, contain:[node-key…]}); determinism resolves each to its narrowest reversible cut. Target-choice, not a mechanism-menu.
  • ADR-0033 — the judge tier (T2b/JEF-568): the deployed qwen3:1.7b holds the contract — assessment 8/8, all refute/cut traps, and the cut-set with a tightened exact-set output instruction (validated on the deployed pod). No escalation to 4B. Records the methodology finding that the cut-choice bench must run on deployed hardware (local backend diverges on the cut-set).

Implementation: the incident/ module landed in #296 (JEF-609); the validated bench + prompt in #297 (JEF-568). JEF-570 wires them into the engine next.

🤖 Generated with Claude Code

thejefflarson and others added 3 commits July 27, 2026 22:29
…ism-menu

Settle the ADR-0032 minimality fork: the model names the compromised on-path
nodes ({assessment, reason, contain:[node-key]}); determinism resolves each to
its narrowest legal cut. Per-target minimality is monotone/deterministic (no
judgment to delegate), and a node-key is what a 1.7B judge can reliably copy
from attended evidence — so this is the 1.7b-viable contract (JEF-568 re-scope).
Collapses the 4-value verdict to a 3-value assessment. Menu is advisory input;
guards downgrade to Uncertain never Refuted; skeptic default inert both ways;
journal v2 stores resolved cut_signature+fingerprint so replay can't repoint.
Refines ADR-0032 §3; + the idea brief. No code yet.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VtjoJttCvBY4dzCoE4f9vP
…exact-set prompt (JEF-568)

Records the T2b measurement: the deployed qwen3:1.7b judge lands the ADR-0034 cut-choice
contract (assessment 8/8, all refute/cut traps, and the cut-set with a tightened output
instruction that pins contain to exactly the evidence-bearing set). No escalation to 4B.
Includes the methodology finding — the cut-choice bench must run on the deployed pod
(local arm64 backend diverges from deployed amd64 on the cut-set).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VtjoJttCvBY4dzCoE4f9vP
@thejefflarson
thejefflarson merged commit 8a2ca97 into main Jul 29, 2026
5 checks passed
@thejefflarson
thejefflarson deleted the cut-choice-contract-adr branch July 29, 2026 04:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant