Governance pile

Measured: 2026-07-27 · Platform: Claude Code · Engine: claude-opus-5 · Target: the always-injected governance pile, pinned by byte snapshot (it is not a released artifact) Detail + raw: results/2026-07-27-governance-pile-claude-code.md · Pre-registered before any adherence data: PREREGISTRATION.md

Tests Anthropic's Steering Claude Code (2026-06-18) claim — appending the system prompt has diminishing returns for adherence, worse when instructions contradict; keep CLAUDE.md under 200 lines — against our pile.

TL;DR

Claim	Verdict
"Keep CLAUDE.md under 200 lines"	FALSE unit for us. AGENTS.md passes at 188 lines while carrying 7.0× the content of a compliant 200-line file (324.9 chars/line vs 43.6 conventional). Gate on tokens; a line count reads green on a file 7× over.
"...particularly if any contradict"	Mechanism real, cost not observed. One HARD live self-contradiction verified at source (hooks-safety.md:104 retires the test :122 still states). But under 5 deliberate contradictions, workers detected, enumerated and disclosed them 3/3 runs rather than degrading.
"more instructions → less strictly followed"	NOT ANSWERED — and that is the result. Adherence sat at ceiling in every cell, so nothing could be ranked. A ceiling proves the level, never the slope.

The pile, measured

	Chars	Est. tokens	Lines
Always-injected governance	198,129	~79,281	1,510
All seven plugins' always-on (platform's own counter)	—	5,697	—

The plugins are 7.2% of the pile. The pile is the cost centre.

• Token figures are CALIBRATED ESTIMATES, not a tokenizer count. chars/4 undercounts by 1.539× (n=13 skill bodies vs claude plugin details; band 1.399–1.586). Char counts are measured; token counts are derived.
• The working baseline of 158,754 chars reconciles exactly but omits the machine-global files, the auto-memory index and the tombstone ledger — all injected. The real pile was 20.5% larger.
• The pile is not stable. It grew +6,875 net chars in ~19 minutes of concurrent editing during this benchmark.

Adherence — 3 conditions × 3 rounds, 9 workers

Ten probes whose correct answer a pile rule dictates; five of them PILE-SPECIFIC (the generic answer differs). Cells: control · +30 consistent dispatch instructions · +30 with 5 contradicting pairs.

90/90 probe items PASS · 45/45 pile-specific · 30/30 dispatch canaries · 0 INDETERMINATE · spread 0 pp in every cell, every round.

Zero-event bounds (Rule of Three, 95%): pile-specific failure rate < 6.7% · all probes < 3.3% · canary adherence under live contradictions < 10%.

The sharper finding is qualitative: in 3/3 conflict runs the worker named the contradicting pairs by number and stated its tie-break — "mutually exclusive, so the responder followed the lower-numbered rule of each pair." At this dose the predicted failure mode appeared as detection and disclosure, not silent degradation.

What this does NOT show

• A ceiling cannot measure a slope. H1 and H2 are neither confirmed nor refuted at their own margin — nothing degraded enough to compare.
• The pile itself was never varied. Editing the rule files was out of scope and the platform injects the full stack at spawn regardless. The dose rode the DISPATCH channel, not the system-prompt channel Anthropic's sentence is about. Never quote this as a replication of their setup.
• The workers are not decorrelated — all nine got the same governance stack at spawn, and every one of them said so unprompted. This measures that the pile is internally sufficient and self-consistent on ten surfaces and survives 30 competing instructions. It does not measure whether the pile is TRUE.
• n=3 per cell, 5 items per headline; the pile moved mid-run; the P9 scorer predicate was corrected mid-run and all nine cells re-scored (both scorings published).

Dated 2026-07-27; a 9-run, 10-probe sample against a byte-pinned snapshot — not a guarantee.

Next arm (not run)

The ceiling is the binding limit, so the next arm must push until something breaks: escalate dose (30 → 100 → 300 instructions) and conflict density (5 → 25 pairs) until canary compliance leaves 5/5, then bisect to find the knee. Until an arm actually produces a failure, "our pile is fine" rests on a ceiling result and must say so. A true H1 test needs the pile itself varied — that needs ownership of the rule files and a window with no concurrent editor.