Skip to content

Rule text: converged candidates by multi-angle convergence, models only through a converged harness - #1049

Merged
seathatflowsinourveins merged 6 commits into
mainfrom
foundation/converged-candidates-rule-20261012
Oct 13, 2026
Merged

seathatflowsinourveins merged 6 commits into
mainfrom
foundation/converged-candidates-rule-20261012

Conversation

@seathatflowsinourveins

@seathatflowsinourveins seathatflowsinourveins commented Oct 12, 2026 •

Copy link
Copy Markdown
Collaborator

Scope

  • What this PR changes, in one or two sentences: on all six always-loaded rule surfaces, the "Decide by evidence" bullet gains two foundation sentences: "Use only converged candidates" (trial after the layer's screen, its requirements checked at a pin; a default only when the layer's multi-angle convergence agrees; no single angle makes a candidate a default) and "Use models only through a converged harness with its full layers". Its not-evidence clause now reads "alone are not evidence", as does the long form in docs/harness-defaults.md:9. Recorded in docs/decisions/2026-10-12-converged-candidates-only.md. No tool, hook, check or runtime behavior changes.
  • Base commit: 703d588c5660d698a87b622efe0e3bd8c22f7e29 (rebased from 7151664c9, a registry-only rebase: only manifests/evidence.json was resolved, by re-registering each commit's files from the merged bytes).
  • Lane: lane:foundation
  • Owned paths touched: AGENTS.md, examples/claude-native/CLAUDE.md, adoption/templates/codex.AGENTS.template.md, adoption/scaffold/AGENTS.md, adoption/new-wsl/claude-user-instructions.md and adoption/new-wsl/codex-user-instructions.md (both generated), docs/harness-defaults.md, tests/test_install_claude_profile.py, tests/test_codex_worker_lane.py, manifests/evidence.json, the new decision record.
  • Commits (rebased ids): e18efdcc (first sentence), d945f4ce (harness sentence), e4d8f694 (trial/default split), 37577653 (multi-angle wording and "alone"), 14ea7e1b (record only: how the rule applies to every layer, the freeze, revisits and the north star), 67f9e826 (review round 1 for both reads: private digests removed, licence never a disqualifier, the protocol's reopen rule quoted as written, the harness-defaults clause, the screen defined and "no single angle makes a candidate a default"). Each later commit is a fast-forward; the earlier wordings never reached main and are listed in the record.

SOTA sources

Evidence-class table

Claim Evidence class Command / receipt
All six layers and three renderers carry both sentences and the "alone" clause once, in place local_integration presence test failed first at each of 5 steps (9 subtests, rc 1; last 22:27:23Z), passes after
The long-form clause reads "alone" local_integration test_the_long_form_rule_says_alone_in_its_not_evidence_clause: the old line counts 0 new and 1 old
Previous pin and ceilings reject the new text local_integration 63-test controls, rc 1 (decision record, Checks)
Carriers equal their sources local_integration --write-blocks then cmp
Startup bytes 8,076/8,177, ceilings 8,480/8,586 local_integration startup_files(); decision record table
Codex default project-doc limit is 32 KiB source_review openai/codex config_toml.rs:75 at the mirror pin
Agents follow the rule not measured see the record's Limits

Local commands run

On head 67f9e826 (the tree as committed), 22:31:08Z to 22:34:57Z:

$ systemd-run --user --scope -q -p CPUWeight=idle env TMPDIR=/var/tmp/cc-converged-rule-tmp python3 -B -m unittest tests.test_install_claude_profile tests.test_managed_block tests.test_new_wsl_client_config tests.test_validate tests.test_codex_worker_lane tests.test_scaffold_repo -v
Ran 627 tests in 228.404s
OK (skipped=13)
$ systemd-run --user --scope -q -p CPUWeight=idle env TMPDIR=/var/tmp/cc-converged-rule-tmp python3 -B -m unittest tests.test_adoption_docs_consistency tests.test_ecosystem_manifest
Ran 127 tests in 100.453s
OK (skipped=1)
$ python3 scripts/validate.py
{"components": 70, "hashed_files": 11855, "profiles": 4, "receipts": 239, "status": "passed"}
$ gitleaks git --staged --no-banner --redact .
no leaks found

tests.test_codex_worker_lane and tests.test_scaffold_repo are added to the brief's four modules because this change re-pins TOP_RULE_SHA256 and moves the scaffold block with the Codex template. The two modules in the second run read docs/harness-defaults.md.

Decision record

docs/decisions/2026-10-12-converged-candidates-only.md: the rule, its basis, what it changes, how it applies to every layer and the freeze, alternatives, the comparison that would overturn it. Supersession: no decision record is superseded; the not-evidence clause is amended in place.

Host evidence

Not applicable: no file under evidence/hosts/ changes.

Checklist

  • New/changed GitHub Actions are pinned to a full commit SHA with a
    version comment (no floating tags). (none changed)
  • New/changed workflows declare top-level permissions: {} and grant each job
    only what it needs (contents: read, or an explicitly justified addition). (none changed)
  • No secrets are printed, logged or committed; no new required secret was
    added without a documented owner.
  • No new paid hosting, subscription or billing surface was introduced.
  • Peer-owned untracked files and worktrees were preserved (not deleted,
    moved or overwritten).

🤖 Generated with Claude Code

@seathatflowsinourveins seathatflowsinourveins added the lane:foundation Foundation lane: Claude/Codex setup, hosts, memory, RAG, research, workers label Oct 12, 2026
@seathatflowsinourveins seathatflowsinourveins changed the title Rule text: use only converged candidates (foundation rule) Rule text: converged candidates by multi-angle convergence, models only through a converged harness Oct 12, 2026
Append the owner's 2026-10-12 foundation rule to the end of the "Decide
by evidence" bullet on all six always-loaded rule surfaces: AGENTS.md,
the portable Claude block, the Codex template, the scaffold and the two
new-WSL carriers, regenerated with new_wsl_client_config.py
--write-blocks. A tool, runtime, model or skill is installed, run or
made a default only after the layer's own convergence passes it;
popularity, recency and vendor claims make it a lead, never an
admission.

- tests: a fail-first test that every layer and native renderer ends
  the bullet with the sentence once (9 failing subtests before the
  edit); CORE updated; the Codex top-rule pin re-derived with
  template_segments() (2,607 bytes); the startup byte ceilings raised by
  a dated comparison to 7,548/7,654.
- docs/decisions/2026-10-12-converged-candidates-only.md: the rule, its
  basis (owner direction, CC ruling addendum-01 with correction #192,
  the existing evidence rule), what it changes, surfaces, budget
  comparison, checks and overturn conditions. Supersession: none.
- manifests/evidence.json: the changed files registered with
  register_file.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Append the owner's second 2026-10-12 foundation rule (20:15Z direction)
right after the converged-candidates sentence, at the end of the
"Decide by evidence" bullet, on all six rule surfaces; the two new-WSL
carriers regenerated with new_wsl_client_config.py --write-blocks. A
model runs only inside a converged harness with its full layers; a bare
API call is for measurement probes and smokes only, recorded as such.

- tests: the presence test now covers both sentences, in order (9
  failing subtests before this edit); CORE updated; the Codex top-rule
  pin re-derived with template_segments() (2,860 bytes); the startup
  ceilings raised by the dated comparison to 8,079/8,185.
- docs/decisions/2026-10-12-converged-candidates-only.md: the second
  sentence, its basis (owner direction 20:15Z, CC runtime-roadmap
  ruling), its change, the budget table and checks for both steps.
- manifests/evidence.json: the changed files re-registered.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Tighten the converged-candidates sentence on all six rule surfaces, on
the CC's direction: a candidate is installed or run for trial only
after the layer's own screen passes it, and becomes a default only
after a measured win on quality at no higher cost. The Stage A screen
checks documented capability at a pin and runs nothing, so a pass is
not quality evidence. The harness sentence after it is unchanged; the
two new-WSL carriers regenerated with new_wsl_client_config.py
--write-blocks.

- tests: CONVERGED (and with it CORE) carries the new wording; the
  presence test failed first (9 subtests); the Codex top-rule pin
  re-derived with template_segments() (2,864 bytes); the startup
  ceilings re-derived to 8,088/8,194.
- docs/decisions/2026-10-12-converged-candidates-only.md: the new
  wording, the earlier wording it replaces before landing, why a screen
  pass is trial-only (CC runtime-profiles addendum-01, Stage A and B),
  the budget table and checks.
- manifests/evidence.json: the changed files re-registered.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Final wording on the owner's 20:29Z direction (CC multi-angle
convergence ruling): on all six rule surfaces, the converged-candidates
sentence now admits a trial after the layer's screen and a default only
when the layer's multi-angle convergence agrees (capability at the pin,
upstream engineering quality, community consensus, net usage, published
evaluations and our own measured runs); where angles disagree, the
disputed dimension is measured first-hand, and no single angle admits a
candidate. The bullet's not-evidence clause reads "alone are not
evidence". The harness sentence is unchanged; the two new-WSL carriers
regenerated with new_wsl_client_config.py --write-blocks.

- tests: CONVERGED and CORE carry the final wording; the presence test
  also checks the "alone" clause in place and the old clause absent
  (9 failing subtests before the edit); PROCEDURE_PHRASES updated; the
  Codex top-rule pin re-derived (3,017 bytes); the startup ceilings
  re-derived to 8,409/8,515.
- docs/decisions/2026-10-12-converged-candidates-only.md: rewritten for
  the final text, with the branch's earlier wordings, the multi-angle
  basis and supersession, the budget comparison and every check.
- manifests/evidence.json: the changed files re-registered.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
… star

Add "How the rule applies: every layer, the freeze and the north star"
to docs/decisions/2026-10-12-converged-candidates-only.md, on the CC's
direction: the six angles and per-layer weights, the scope across both
repositories, the wave cycle, the freeze in final-stack.json (current
counts: converged 0, reopened 30, no-winner 2, unswept 2; no angle
weights), what scripts/final_stack.py and CI check today, the revisit
triggers, the SOTA-sources pin rule and the trading north star. Facts
cited path:line at main 489b263, byte-identical at the base. No rule
surface, pin or ceiling changes.

- manifests/evidence.json: the record re-registered.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
@seathatflowsinourveins
seathatflowsinourveins force-pushed the foundation/converged-candidates-rule-20261012 branch from ce11fbd to 14ea7e1 Compare October 12, 2026 22:26
…open rule as written

One commit for both reads of #1049 at ce11fbd (the GPT verdict and the
Claude read), after the registry-only rebase onto main 703d588.

- GPT P1 (#161): the decision record cites the CC's private rulings by
  name only, with no path and no digest; the PR body does the same.
- GPT P1 (licence): the hard disqualifiers are a missing required
  capability (an entitlement or terms conflict counts as one) and an
  abandoned project. Licence is a compliance note (SPDX id, source,
  read time), per the CC ruling ruling-licence-not-a-disqualifier-20261012.
  No rule surface named licence.
- Claude P2-1: the protocol's reopen rule is quoted as written (a FINAL
  layer reopens only when a sweep detects a change, bound to that
  sweep). The angle sources become sweep inputs, a named follow-up to
  recipes/saturation-sweep.md; review contradictions go through the
  corrections rule.
- Claude P2-2: docs/harness-defaults.md:9 reads "alone are not
  evidence", pinned by a new test that fails on the old line.
- P3s: on all six surfaces "the layer's screen (its requirements
  checked at a pin)" and "no single angle makes a candidate a default",
  carriers regenerated with --write-blocks; the presence test failed
  first (9 subtests); TOP_RULE_SHA256 re-derived (3,051 bytes); the
  startup ceilings re-derived to 8,480/8,586 with the Codex
  project_doc_max_bytes primary source (openai/codex rust-v0.162.1,
  config_toml.rs:75). CI checks angle records only once the freeze-wave
  PR lands; plans and host reports are labelled as such.
- manifests/evidence.json: the changed files re-registered.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

lane:foundation Foundation lane: Claude/Codex setup, hosts, memory, RAG, research, workers

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant