test(wave-4): echo-chamber cohort #1 — consensus-with-dissent (W4.1, #196)

[Dewinator]

Summary

Fourth fixture row of the W4.1 anti-echo corpus (#196). Adds the echo-chamber cohort — the §10.4 diversity-filter trigger boundary case (over_concentration = 4/5 = 0.8), distinct from sybil-flood's saturation case (M=10, over_concentration=1.0).

The cohort is heterogeneous: 4 echoers + 1 visible dissenter. The load-bearing claim is that the dissenter's presence does NOT save the consensus lesson from the firebreak — §10.4 is purely a function of over_concentration, not of cohort heterogeneity. This is the case that defends against organic-looking groupthink where one might otherwise argue "but there's a dissenter, the cohort isn't really converged".

What's in the box

• mcp-server/scripts/build-anti-echo-fixtures.mjs — Step 1d (write-once cohort-keys-echo-chamber.json, 5× Ed25519) + Step 6 (echo-chamber cohort fixture builder). Same write-once discipline as the plagiarism / sybil cohorts; distinct key file so a future refactor of either neighbouring cohort cannot silently shift the §10.4 boundary this row asserts on.
• mcp-server/src/__tests__/fixtures/anti-echo/cohort-keys-echo-chamber.json — fixture-only Ed25519 keys for the 5 cohort peers.
• mcp-server/src/__tests__/fixtures/anti-echo/echo-chamber/consensus-with-dissent-cohort.json — 5 signature-valid envelopes; envelopes[0..3] are echoers (shared content + evidence_root, near-duplicate embeddings via rampPerturbed(i+200), signed within a 3-minute window), envelopes[4] is the dissenter (heterodox content, distinct evidence_root, reversed-ramp embedding ≈ cosine 0.5 vs every echoer).
• mcp-server/src/__tests__/anti-echo-echo-chamber.test.ts — 11 assertions:
  1. Metadata declares broadcast_suppressed, negative trust delta, §10.4 owner.
  2. cohort_size == envelopes.length and over_concentration ≥ 0.8 spec threshold.
  3. Every envelope is signature-valid against its cohort key.
  4. Distinct origin_node_id per envelope (5 peers, not one peer 5 times).
  5. Every envelope passes validateWireRecord (§10.4 is post-admission).
  6. Pairwise cosine > 0.95 within the echoer subset (§10.4 p_duplicate_cosine).
  7. Dissenter cosine < 0.95 vs every echoer (real heterodox peer, not a sock-puppet).
  8. signed_at spread ≤ §10.4 p_signed_at_window_d (7 days).
  9. Echoers share evidence_root; dissenter has its own.
  10. scoreFinding(prev=1.0, over=0.8) clamps new_weight=0.2 strictly below MIN_LOCAL_WEIGHT_FOR_BROADCAST (0.3) — even with a visible dissenter.
  11. Padding topic_cohort_size=100, near_dup_origin_count=80 (same 0.8 ratio) does NOT relax the multiplier — proves §10.4 has no count-dependent escape hatch.

Status of the W4.1 corpus after this PR

#	Category	Mechanism	PR	Status
1	forgery	§3.7 + §5 r.1	#197	merged
2	plagiarism	§10.4	#198	merged
3	sybil-flood	§10.2 + §10.4	#201	merged
4	echo-chamber	§10.4	this	open
5	slow-poisoning	§10.5 REM self-audit	—	open
6	truth-by-repetition	§10.4 + §10.6	—	open
7	polarity-inversion-pair	§10.3 contradicts	—	open
8	chain-rewrite	§5 r.19	—	open

Test plan

• node --test dist/__tests__/anti-echo-echo-chamber.test.js — 11/11 pass locally
• node mcp-server/scripts/build-anti-echo-fixtures.mjs re-run is a no-op (write-once + identical fixture bytes)
• npm run build clean (no new TS errors)

🤖 Generated with Claude Code

[Dewinator]

Local validation (189th-tick fresh-clone walk, queue-bound, no-new-PR)

Walked PR #212 on a clean rebuild from agent/wave-4-echo-chamber-fixture (head da18c43). Echo-chamber fixture is self-contained — no sibling regression in the W4.1 anti-echo corpus.

#212 (this PR — W4.1 row 4 echo-chamber cohort, branch agent/wave-4-echo-chamber-fixture)

rm -rf dist && npm run build                                         → clean
node --test dist/__tests__/anti-echo-echo-chamber.test.js            → 11/11 pass (459ms)

Sibling-cohort regression (forgery / plagiarism / sybil-flood unchanged)

node --test dist/__tests__/anti-echo-forgery.test.js \
            dist/__tests__/anti-echo-plagiarism.test.js \
            dist/__tests__/anti-echo-sybil-flood.test.js             → 24/24 pass (474ms)

The write-once discipline holds: cohort-keys-echo-chamber.json is a fresh 5-key file (35 lines, distinct from cohort-keys-plagiarism.json and cohort-keys-sybil-flood.json). The cohort fixture itself is 3937 lines (5 signature-valid envelopes — 4 echoers + 1 dissenter). No file outside the four declared in the PR is touched.

What this PR uniquely asserts (re-read after running)

The two boundary-distinguishing assertions are the load-bearing ones:

• Test 10 — scoreFinding at over_concentration=0.8, M=4 of 5 clamps below the broadcast threshold. This is the §10.4 trigger boundary (vs sybil-flood's over_concentration=1.0 saturation).
• Test 11 — scoreFinding does NOT relax the firebreak when topic_cohort_size grows but over_concentration stays at 0.8. This is the dissenter-doesn't-save-it claim — §10.4 is purely a function of concentration, not heterogeneity. Without this, organic-looking groupthink with one visible dissenter would slip past the firebreak.

Together with row 1 (forgery, signature integrity), row 2 (plagiarism, citation echo), and row 3 (sybil-flood, saturation), row 4 closes the §10.4 boundary case. The remaining four W4.1 rows (per docs/wave-4-w4.1-row-triage.md) remain as separate fixtures.

#212 is independent of the DbClient stack (#209 → #210 → #211) and of #208 (sidecar-bundle). It can land in any order relative to those four; the file paths don't overlap.

No queue growth from this validation tick.

🤖 Generated with Claude Code

[Dewinator]

Empirical merge-rebase validation (201st tick, 2026-05-04)

PR #212 validated locally end-to-end against current main (HEAD 180c1e9, 12 docs commits ahead of PR base ec8a088).

Method: fetched pull/212/head, git rebase origin/main (clean, no conflicts — confirms the GitHub mergeStateStatus = CLEAN), rm -rf dist && npm run build clean, then ran node --test against built artifacts.

Results on the rebased state:

• Full mcp-server suite: 1017 / 1018 pass, 0 fail, 1 skip (7.8 s).
• W4.1-specific (anti-echo-forgery.test.js + anti-echo-plagiarism.test.js + anti-echo-sybil-flood.test.js + anti-echo-echo-chamber.test.js): 35 / 35 pass (5 forgery, 9 plagiarism, 14 sybil-flood, 11 echo-chamber).
• The PR body's load-bearing claim — "even with a visible dissenter, scoreFinding(prev=1.0, over=0.8) clamps new_weight=0.2 strictly below MIN_LOCAL_WEIGHT_FOR_BROADCAST (0.3)" — fires green in the rebased state. §10.4 has no count-dependent escape hatch in this code path either.
• Build-script idempotency: node scripts/build-anti-echo-fixtures.mjs re-run on top of an already-populated tree → exits 0, git status empty (write-once + identical fixture bytes hold).

Implication for merge order: PR #212 is base = ec8a088 and the 12 commits since then are docs-only (wave-3, native-dbclient-tracker, wave-2, waves, claude.md, gitignore — zero overlap with mcp-server/src/__tests__/anti-echo-*.test.ts, mcp-server/src/__tests__/fixtures/anti-echo/**, or mcp-server/scripts/build-anti-echo-fixtures.mjs). The PR remains independent of #208 (sidecar bundling) and the #209/#210/#211 DbClient stack — Reed can land it in any order.

Tracker grounding (sanity-check ride-along): re-ran the docs/native-dbclient-migration-tracker.md .from(/.rpc( methodology against current main. Counts for all 22 listed services + the lesson-chain helper match the tracker's tabulated values exactly — Stand 2026-05-04 (195. tick recount) is still empirically grounded.

🤖 Posted by autonomous tick (mycelium 201st)

[Dewinator]

Local empirical validation 2026-05-05 (213th-tick) — 11/11 tests pass + spec/SQL cross-check

Fresh pr-212 checkout in mcp-server/, ran the full PR test file standalone:

npx tsx --test src/__tests__/anti-echo-echo-chamber.test.ts
ℹ tests 11
ℹ pass 11
ℹ fail 0

All 11 assertions green in 926ms — metadata, wire-signature (5×), wire-validator (5×), pairwise cosine (echoer subset > 0.95, dissenter ≪ 0.95), signed_at window, evidence_root structure, scoreFinding boundary, and the padded-cohort invariant.

Cross-check the load-bearing claim against the actual SQL gate

The PR claims over_concentration = 4/5 = 0.8 is the §10.4 trigger boundary. The spec text in docs/SWARM_SPEC.md:741 says >80% (strict greater-than), which would not fire at exactly 0.8. The PR's claim only holds if the SQL relaxes that to >=.

supabase/migrations/082_swarm_lesson_diversity.sql:199 confirms it does:

WHERE tc.topic_cohort_size >= p_min_cohort_size
  AND (nd.near_dup_origin_count::FLOAT / NULLIF(tc.topic_cohort_size, 0)) >= p_over_concentration;

And the migration's own header comment (lines 116–129) acknowledges the deliberate spec-text→implementation relaxation: "default 0.8 matches the issue's '5 peers, 4 hold the same' acceptance test, even though §10.4 spec text says [>80%]". Boundary-at-0.8 is intentional, not accidental — PR #212's fixture exercises exactly the case the migration was tuned for.

Why the padded-cohort test (M=80/N=100) is the strongest assertion

The second scoreFinding test is the one that closes a real attack vector: an adversary can't gather extra dissenters to dilute the concentration ratio downward, because §10.4's multiplier is (1 − over_concentration) — purely ratio-driven, with no N-dependent escape. Pinning that invariant in the row-4 fixture (rather than only at the M=4 boundary) is what makes this row a defense and not just a regression test.

Hands match feet across the §10 chain

• scoreFinding(prev=1.0, over=0.8) → new_weight = 0.2 (verified mathematically: factor = 1 - 0.8 = 0.2, candidate = 1.0 * 0.2 = 0.2, max(0, min(1.0, 0.2)) = 0.2).
• MIN_LOCAL_WEIGHT_FOR_BROADCAST = 0.3 in swarm/endpoints/lesson-feed.js.
• 0.2 < 0.3 → broadcast firebreak fires, lesson stays Tier-B. Matches the expected_outcome: "broadcast_suppressed" in the fixture metadata.

Mergeable from my side. No bugs surfaced. The boundary-vs-saturation split with sybil-flood (row 3) gives W4.1 two complementary §10.4 cases, and the row-1 rule-numbering correction (§5 rule 1 → §5 rule 18, commit d69570a on main) does not affect this row — owns_mechanism here is §10.4 only.

🤖 Generated with Claude Code