open/voyage
1
0
Fork 0
Code Issues Pull requests Releases Wiki Activity Actions

docs(voyage): S14 — devil's-advocate audit results (Dynamic Workflow)

Browse source 
Cold adversarial audit of Voyage via Workflow tool (13 agents, 6 attack
dimensions -> rebuttal -> synthesis). Verdict: ship-worthy machine, but
docs need a truth-pass. Verified MAJORs: README sells 'cheap Sonnet'
swarms while all 24 agents are model: opus; stale counts (109 vs 683
tests, 23 vs 24 agents, 5 vs 7 hooks); NW2 bake-off raw data uncommitted;
brief framing enforcement bypassable via brief_version 2.1. Audit only --
no Voyage code/docs changed; acting on findings needs fresh go-ahead.
This commit is contained in:
Kjell Tore Guttormsen 2026-06-18 18:49:01 +02:00
commit 0e84d926e2
1 changed files with 130 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

130
docs/devils-advocate-results.md Normal file
View file

@ -0,0 +1,130 @@
# Devil's-advocate results — adversarial cold audit of Voyage
**Run:** S14, 2026-06-18. **Method:** Dynamic Workflow (`Workflow` tool), 13 agents @ `effort: high`, Opus 4.8.
**Shape:** 6 adversarial dimensions (D1–D6), each `attack` (strongest case AGAINST) → `rebut` (steelman defense → STANDS/WEAKENED/REFUTED) → 1 synthesis. ~905k subagent tokens, 9.3 min.
**Plan:** `docs/devils-advocate-plan.md`. **Scope:** audit only — findings + recommendations, NOT an implementation session. Run ID `wf_4af87c2f-d4e`.
> **Read this caveat first.** This is a *cold, hostile* audit that was instructed to prove Voyage is over-engineered/wrong/unmaintainable, then to rebut itself honestly. Verdicts are the rebuttal pass's, not gospel. The audit caught a factual error in one of its OWN findings during rebuttal (D5 "0 refs" → REFUTED), which is a reminder that some cited `file:line` counts from the attack agents may be imprecise. I (main context) **independently re-verified** the load-bearing MAJOR claims — see the **Verification log** at the bottom. Everything else is reported as the audit found it.
---
## Bottom line (overall verdict)
**Voyage survives as a functioning, disciplined, but over-documented and under-measured plugin.** It is not fraudulent or broken. But its self-description is materially inaccurate, and its flagship value proposition is unproven where it was actually tested. The strongest honest charges are concentrated and cheaply fixable:
1. The README sells "cheap Sonnet swarms" while **every one of the 24 agents ships as Opus** — on a section marketed as "observable in the code, not aspirations."
2. The headline counts are **stale by up to 6×** (109 vs 683 tests, 23 vs 24 agents, 5 vs 7 hooks) and unguarded by the doc-consistency test that exists to prevent exactly this.
3. The **NW2 "POSITIVE" verdict** rests on a single un-archived bake-off whose raw per-run data was never committed and whose *primary* fidelity metric actually failed and was re-derived post-hoc.
4. The **brief PUBLIC-CONTRACT framing defense is bypassable** by writing a smaller version number (`brief_version: 2.1`).
Against that, the plugin shows genuine engineering hygiene the attack under-credited: the dormant synthesis-agent is a clean *measured-and-declined prune-in-place* (the opposite of accretion); the EVALUATE-parked topology decisions are honestly staged, not frozen on a dead premise; the auto-mode classifier risk was *discovered-and-priced*, not ignored; and the framing BLOCKER **does** hard-fail even in soft mode.
**The deepest problem is not over-engineering — it is an unaudited scoreboard.** A plugin whose entire pitch is contract-conformance and filesystem-validated rigor cannot keep its own test count, agent count, model claims, and version references straight — and, tellingly, its author did not dogfood it on its own upgrade. **Ship-worthy machine; the docs need a truth-pass and the central context-engineering claim needs either a real measurement or a humbler sentence.** Crucially, the audit never tested whether the happy path produces *good plans* — the question that matters most remains open (see Blind spots).
---
## Survivors — ranked by severity × actionability
Critiques that **STAND** or are **WEAKENED-but-real** after the rebuttal pass. (Severity = residual after best defense.)
### MAJOR × HIGH actionability — fix first
| # | Finding | Dim | Verdict | Evidence (verified ✓) |
|---|---------|-----|---------|----------|
| 1 | **Cost claim is false by default.** README sells "Sonnet runs exploration/review swarms / front-loads cheap Sonnet work" in 6 places, but all 24 agents hardcode `model: opus` and `trekplan.md` concedes sub-agents are "still pinned to opus." Sold under "observable in the code — not aspirations." | D6/D1 | STANDS | README.md:195,223,266,785,804,857; all 24 `agents/*.md` `model: opus` ✓; trekplan.md:922,968 |
| 2 | **README hard counts stale & unguarded.** Architecture block: 23 agents (actual **24**), 5 hooks (actual **7** scripts / 6 events), **109** node:test cases (actual **683**, a 6× miss on the line that pitches "npm test is the fork-readiness gate"). | D6 | WEAKENED | README.md:804,807,809 ✓; `npm test` = 683/681/2 ✓ |
| 3 | **Bake-off data never committed.** NW2 "POSITIVE" cites a 6-run token/wall/finding table + jaccard ladder (0.41/0.71/0.86/1.00), but no per-run JSON (`a1.json`..`b3.json`) exists in repo or git history. `bakeoff-fidelity.mjs` requires them as input → the numbers cannot be regenerated, audited, or falsified. | D2 | STANDS | scripts/bakeoff-fidelity.mjs:17-19,98-101; docs/T2-bakeoff-results.md:131-135; f7c8aa4 stat |
| 4 | **Self-reported test count wrong.** STATE.md:43 + the audit plan claim 697 (695 pass); runner reports **683 (681 pass, 2 skip)**. Same unaudited-self-description class the plugin's conformance discipline exists to catch. | D1/D6 | STANDS | STATE.md:43; `npm test` ✓ |
| 5 | **Three reference-only orchestrator agents ship with an `Agent` tool grant** while their bodies declare "reference, not a runnable sub-agent" and no command invokes them. A harness could dispatch an agent the plugin says must never run. Trivial, untaken fix. | D4 | STANDS | agents/{planning,research,review}-orchestrator.md; review-orchestrator.md:226-227; trekreview.md:33 |
| 6 | **`plan-critic` "9 dimensions" wrong in 3 places** (README.md:223,266; CLAUDE.md:46) — agent defines **10** numbered dimensions, the 10th a hard gate (manifest quality). Unpinned. | D6 | STANDS | agents/plan-critic.md:170 |
| 7 | **README self-contradicts brief-reviewer dimension count:** README.md:158 says "five dimensions" while README.md:15 headlines the **6th** (memory alignment) as the v5.5 flagship; agent + CLAUDE.md say six. Exactly the consistency defect brief-reviewer exists to flag. | D6 | STANDS | README.md:15,158; agents/brief-reviewer.md:39,155; CLAUDE.md:42 |
| 8 | **Stale model string in a live contract:** trekplan Phase 8 inline-sealing rationale cites "Opus 4.7" while the plugin elsewhere states opus = Opus 4.8. Sealing logic is sound; only the parenthetical is wrong. | D4 | STANDS | trekplan.md:574; CLAUDE.md:56 |
| 9 | **PUBLIC-CONTRACT versioning cites a non-existent release:** HANDOVER-CONTRACTS.md:46,99 + CLAUDE.md:9 say "v5.4 froze the brief_version 2.1 baseline," but v5.2–5.4 were never shipped (CHANGELOG jumps 5.5.0→5.1.1). No consumer stranded (compat keys off `brief_version`), but a doc trading on contractual precision cites a release nobody can install. | D2 | WEAKENED | CHANGELOG.md:7-9,63; HANDOVER-CONTRACTS.md:46,99; CLAUDE.md:9 |
### MAJOR × MEDIUM/LOW — structural, need design not just edits
| # | Finding | Dim | Verdict | Evidence |
|---|---------|-----|---------|----------|
| 10 | **Framing enforcement is producer-elective.** All three `brief_version 2.2` framing layers are wrapped in `atLeast22`; there is **no** minimum-version gate (grep confirms); 2.0/2.1 briefs are blessed as fully valid. Any producer ships an intent-misaligned brief by declaring `2.1` and the entire framing defense evaporates with no warning. The contract advertises producer-agnostic enforcement but enforces only against the one producer (`/trekbrief`) that hardcodes 2.2. | D3 | STANDS | brief-validator.mjs:123-140; HANDOVER-CONTRACTS.md:101 |
| 11 | **Memory-alignment gate is a no-op by default.** Score-5-on-no-context is indistinguishable to the Phase-4e gate from score-5-on-verified-alignment; whether memory is gathered is an unbacked LLM-judgment step with no validator backstop. In the common (memory-less / external-producer) case the gate always passes. Defensible as *portability* (fail-open), but a passing score is not evidence of alignment. | D3/D1 | WEAKENED | brief-reviewer.md:168-169; trekbrief.md:492-497,541-544 |
| 12 | **Layer 1 framing value is self-attested, enum-validated only.** No cross-check against reality; its triangulation partner (Layer 2 memory) is inert in the no-memory case → framing is an unchecked assertion carried as pipeline source-of-truth. The "no default / unskippable" design blunts the tired-one-click vector but doesn't make the self-label true. | D3 | WEAKENED | trekbrief.md:114-130; brief-validator.mjs:110-116 |
| 13 | **NW2 "POSITIVE" = +4.4% tokens, +54% wall-time, and the one cited upside ("frees main context") never measured.** The bake-off captured subagent tokens (the axis a coordinator-adding Workflow can only tie-or-lose); Δ main-context (the decision-relevant quantity) was left unmeasured. The +15% POSITIVE bar **was** pre-registered (so not goalpost-moved), but the win rests on unquantified benefit. | D1/D2 | WEAKENED | T2-bakeoff-results.md:152-170 |
| 14 | **NW2 fidelity verdict salvaged post-hoc.** Pre-registered PRIMARY metric (triplet-level finding-set equivalence) scored **0/9** strict-equivalent, 0.41 jaccard; POSITIVE reached by reporting the looser (file,rule_key)=0.71, labeling the 0/9 a "metric-calibration artifact," then recommending future metrics use the coarser granularity that cleared. Within-arm control makes "artifact" plausible, but it was decided after seeing the result. | D2 | WEAKENED | T2-bakeoff-results.md:131-135,172-185; T2-cc27-workflow-substrate.md:170 |
| 15 | **Context-Engineering flagship unproven where tested.** The one experiment (NW3) measured Δ main-context ≈ 0 and shipped dormant. The Δ=0 is real but narrowly scoped to delegating only the Phase-7 read; the actual fan-out relief — the claim that matters — was never measured. The swarm topology's central justification has no positive measurement behind it, only a structural argument + a declined experiment. | D1 | WEAKENED | T1-synthesis-poc-results.md:16-30,88-103 |
| 16 | **Voyage was not used to build Voyage.** Much of the v5.5 arc was research/meta-work the pipeline doesn't target (blunts it), but genuinely pipeline-shaped code sub-tasks (armB script, synthesis schema, fidelity libs) were also hand-orchestrated. Revealed-preference signal for those is real. | D1 | WEAKENED | devils-advocate-plan.md:8; W1/T1/T2 docs |
### MINOR — cheap correctness / leanness nits
| # | Finding | Dim | Verdict | Evidence |
|---|---------|-----|---------|----------|
| 17 | **Prose-pin bloat.** `doc-consistency.test.mjs` (67 tests, ~9.5% of suite) is predominantly `.includes/.match` on free prose, churned across 26 commits, flips red on benign rewording. The structural sub-tests (row-count == file-count, command coverage) are genuine; the headline test-count over-credits coverage. | D5/D1 | WEAKENED | tests/lib/doc-consistency.test.mjs:1-6 |
| 18 | **Chartered auto/bypass classifier check skipped.** "Classifier interference: 0" was reported satisfied, but the auto/bypass-mode re-run that W1 made an explicit guard was skipped on a technicality ("mode not settable in-session") and footnoted rather than gating. The mode that matters for headless trekreview remains untested. | D2 | STANDS | T2-bakeoff-results.md:146-149; W1-narrow-wins-plan.md:29-32 |
| 19 | **NW3 decline dressed in pseudo-quant scaffolding.** The faithful Δ≈0 is genuinely *structural* (holds with no tokenizer), but it is surrounded by a sweep table + break-even + a "15% adopt-floor" that contradicts the plan's 30% bar, over a chars/4 estimate with a swept (never API-measured) BASE. Conclusion sound, scaffolding inconsistent. | D2 | WEAKENED | synthesis-measure.mjs:41-42; T1-synthesis-poc-results.md:12-14,94; W1-narrow-wins-plan.md:17 |
| 20 | **Experiment scaffolding outlived its experiment.** 5 of 7 `scripts/` (~89KB, incl. `synthesis-measure.mjs` for the declined agent) are one-off PoC/measurement code on no command path, each referenced only by its own test, carrying 27 live tests. Partly defensible as reproducibility artifacts; the global continuity convention says such history belongs in git. | D5 | WEAKENED | scripts/synthesis-measure.mjs:2-3; tests/scripts/synthesis-measure.test.mjs |
| 21 | **CHANGELOG cross-doc wart.** CHANGELOG.md:440 advertises "711 pass … +5 Playwright e2e" while README.md:595 documents Playwright e2e as removed in v5.0.0. (The "no number matches reality" framing was refuted — historical snapshots legitimately diverge from HEAD.) | D6 | WEAKENED | CHANGELOG.md:440; README.md:595 |
---
## REFUTED — critiques that did NOT survive the rebuttal
These are recorded so they are not re-litigated. Each was a plausible attack that the defense (and, where noted, my own check) genuinely defeated.
1. **"8 of 14 docs are dead scratch; the decision-matrix has 0 inbound refs" (D5).** *False on its lead exhibit.* `cc-upgrade-2.1.181-decision-matrix.md` has **9** inbound refs incl. three production command files (trekplan, trekbrief, trekresearch) + three orchestrator agents + CHANGELOG — the *most*-referenced of the eight. Only `devils-advocate-plan.md` (the just-created S14 plan) cleanly matches "0 inbound." A much smaller "a couple of decision docs could move to git" survives at MINOR.
2. **"Orchestrator descriptions assert the FALSE premise as STANDING rationale" (D4).** The descriptions frame it as *discharged history* in the same sentence-group ("Historically… As of CC 2.1.172 sub-agents CAN spawn…, so a redesign is under evaluation") — exactly what CC-01 prescribed (replace with truth + forward pointer). A dispatching model is steered *away* from spawning, not toward the falsehood. The duplication is a real but MINOR doc-style critique (→ became survivor #5's narrower form).
3. **"The dormant synthesis-agent is ceremony spent on sand" (D1/D5).** It is the strongest *counter*-example: built → measured (Δ≈0) → found NEGATIVE → deliberately not wired, with a stated re-activation trigger and a schema-backed, tested contract. Disciplined negative-result hygiene. The residual (it still ships in the distributable, inflating counts) is a MINOR leanness nit.
4. **"The auto-mode classifier is a NEW liability the plugin is trading into" (D4).** Inverts the evidence: the plugin *discovered and logged* the risk via a cheap probe in a path it has **not** adopted, priced it as a cost against delegation, and declined the trade. Verification discipline, not a hidden liability.
5. **"Soft-mode read means framing/TL;DR add no enforcement at the boundary" (D3).** Refuted *for framing specifically*: `BRIEF_MISSING_FRAMING` is pushed **unconditionally** to `errors[]` (no strict/soft branch), and soft mode means "warnings don't block, errors DO" — so a 2.2 brief missing framing hard-fails even in soft read. (The broader "external 2.0/2.1 producer gets nothing" point survives as survivor #10.)
6. **"No published test count matches reality" (D6/D2).** Half-refuted: a CHANGELOG is an append-only point-in-time log; historical snapshots (503/516/518/608/711) *correctly* diverge from current 683. The genuine residual is the narrow Playwright cross-doc contradiction (→ #21).
7. **"NW2's +15% bar was set AFTER the run to make a tie pass" (D2).** The +15%/+30% thresholds were pre-registered before the run, which predicted token cost "≈ wash." Not a moved goalpost. What survives is the distinct, separately-filed point that the cited upside was never *quantified* (→ #13).
---
## Top changes — prioritized
1. **FIX THE COST CLAIM** (MAJOR × HIGH). Either flip exploration/review agent frontmatter to `model: sonnet` (makes the README true by default — the cheaper fix a cost-conscious user expects) **or** rewrite the 6 README claims to state Opus-by-default and Sonnet requires `--profile economy` / `phase_signals.model`. Add a doc-consistency pin that fails if README says "Sonnet exploration" while agent files say opus.
2. **FIX THE HARD COUNTS** (MAJOR × HIGH). README → 24 agents / 7 hooks / 683 tests; correct STATE.md 697→683 and the plan's 697. Add doc-consistency pins for agent-count, hook-script-count, and a tolerance-banded or dynamically-computed test-count so the Architecture block cannot silently drift.
3. **COMMIT OR DOWNGRADE THE BAKE-OFF DATA** (MAJOR × HIGH). Commit the raw arm outputs (`a1.json`..`b3.json`) under `tests/fixtures/bakeoff-rich/runs/` + a test that re-derives the medians/jaccards — or relabel "POSITIVE" to "defensible as opt-in, single un-archived run."
4. **ADD A MIN-VERSION GATE OR DOCUMENT THE HOLE** (MAJOR × MEDIUM). Optional `--min-brief-version` at the `/trekplan` + `/trekresearch` boundary (warn when an older brief sidesteps framing); document in HANDOVER-CONTRACTS.md that pre-2.2 briefs receive zero framing enforcement.
5. **MAKE THE MEMORY GATE HONEST** (MAJOR × MEDIUM). Emit a `memory_alignment.status` distinguishing "N/A (no memory)" from "verified aligned"; document that in memory-less environments the wrong-premise defense is Layer 1 + Layer 3 only.
6. **STOP OVERSELLING THE FLAGSHIP** (MAJOR × MEDIUM). Either measure the fan-out's real main-context relief with a tokenizer/API run, or soften README/CLAUDE.md to "asserted, not yet measured." Relabel NW2 from POSITIVE to opt-in-defensible and disclose its primary fidelity metric failed and was re-derived post-hoc.
7. **CHEAP CORRECTNESS NITS** (MINOR × HIGH). Drop the `Agent` grant from the three reference-only orchestrator agents; `plan-critic` 9→10 dimensions; README brief-reviewer 5→6 dimensions; trekplan "Opus 4.7"→"4.8"; replace the non-existent "v5.4" contract-freeze references with "v5.5.0."
8. **PRUNE PROSE-PIN BLOAT** (MINOR × MEDIUM). Keep the structural cross-file invariant tests; prune pure phrase-pins that break on reword; report behavior-test count separately from prose-pin count.
9. **RUN THE CHARTERED AUTO/BYPASS CLASSIFIER CHECK** (MINOR × MEDIUM) before treating classifier interference as retired — or move it from "metric satisfied" to "open residual."
---
## What this audit might have missed (completeness critic)
The audit was instructed to critique its own plan. It did:
1. **The 6 dimensions are introspection-biased.** D1/D2/D4 all interrogate the v5.5 self-measurement arc (synthesis, bake-off, decision matrix) and **overlap massively** — three dimensions largely re-attack the same NW2/NW3/CC-26/27 docs. This over-weights recent meta-work and **under-weights the product surface a user actually touches.** There is no dimension auditing whether the **core happy path** (`/trekbrief → /trekplan → /trekexecute` on a real feature) produces correct, useful plans/code. *The audit proved the docs are inconsistent and the experiments were narrow — but never tested whether the pipeline WORKS end-to-end. That is the question a prospective user cares about most, and it remains open.*
2. **No security / safety dimension.** This plugin spawns autonomous agents, has `--gates` autonomy primitives, OTLP export with stated SSRF mitigation, and hooks that gate Bash/Write. None was audited for actual safety (does the SSRF mitigation work? do the pre-bash/pre-write hooks actually block? can a malicious brief drive headless execution?). For an autonomous-execution plugin, the **highest-consequence blind spot.**
3. **No empirical cost measurement.** The entire cost debate (D1/D6) was argued from frontmatter strings and README prose, never from the `trek*-stats.jsonl` the plugin itself publishes. The single most decision-relevant number (what one `/trekplan` run actually costs on Opus vs Sonnet) was never pulled from real run data — so the cost critiques remain *inferential*.
4. **The audit accepted the plugin's own metrics** (Δ main-context, jaccard, verdict-match) as the terms of debate, then critiqued within them. It never asked: are these the right metrics at all? **Plan quality** (does the adversarial review catch real bugs?) was never measured by anyone — the whole value proposition is *assumed* by both attack and defense.
5. **Trust in unverified citations.** Refuted/weakened verdicts were taken largely on trust where the agents couldn't independently re-verify every line. The D5 "0 refs" error that was caught suggests other cited counts may also be imprecise. (Main context independently re-verified the load-bearing MAJORs — see below — but not all ~1672 lines of trekexecute or every cited doc line.)
6. **No user-facing ergonomics / onboarding dimension.** 7 commands, 24 agents, 14 docs, a profile system, `--gates`, brief_version axes — the cognitive load on a new operator is itself a product risk that revealed-preference (the author used bespoke docs) hints at, but the plan never made first-class.
7. **The audit conflates "documentation defect" with "product defect" throughout.** Most MAJOR survivors are doc-accuracy issues — real and cheap to fix, but the plan never separated "the docs lie" from "the machine is broken," leaving the genuinely important question (is the runtime sound?) unanswered.
---
## Verification log (Verifiseringsplikt)
**Independently re-verified by main context this session (not taken on the agents' word):**
| Claim | Method | Result |
|-------|--------|--------|
| Test count is 683/681/2, not 697/695/2 | `npm test` (full run) | ✓ **683 tests, 681 pass, 0 fail, 2 skipped** — STATE.md:43 + plan are wrong |
| All exploration/review agents are Opus, none Sonnet | `grep -h "^model:" agents/*.md \| sort \| uniq -c` | ✓ **24 `model: opus`, 0 `sonnet`** |
| Agent file count is 24 (README says 23) | `ls agents/*.md \| wc -l` | ✓ **24** |
| README sells "Sonnet" exploration/review | `grep -n Sonnet README.md` | ✓ lines **195, 223, 266, 785, 857** + 804 "(sonnet for exploration + review…)" |
| README hard counts stale | `grep -nE "specialized agents\|node:test\|hooks" README.md` | ✓ **804** "23 specialized agents", **807** "5 hooks", **809** "109 node:test cases" |
**Taken from the workflow agents, spot-checked but NOT exhaustively re-read by main context:** the internal `file:line` citations in `docs/T2-bakeoff-results.md`, `docs/T1-*`, `lib/validators/brief-validator.mjs` (the `atLeast22` guard + unconditional `BRIEF_MISSING_FRAMING`), `agents/brief-reviewer.md`, `commands/trekbrief.md`, and the absence of `a1.json`..`b3.json` in git history. These are reported as the audit found them; they are consistent with the verified anchors above but were not each independently confirmed line-by-line.
**Known imprecision:** the attack agent for D5 asserted `cc-upgrade-2.1.181-decision-matrix.md` has "0 durable refs"; the rebuttal pass found **9** and REFUTED it. Treat raw counts inside individual attack findings with that caution.
---
## Scope note
This is an **audit** — findings + recommendations only. Per the plan's scope guard and the operator's run mode, **no Voyage code or docs were changed** on the basis of these results (the one exception: STATE.md's stale `697` test count is corrected at session end, since STATE is continuity state, not plugin surface, and propagating a now-verified-wrong number would be its own defect). Acting on the top changes above requires a fresh operator go-ahead.