Skip to content

title: "Challenge — "What could go wrong with our conclusions?"" source: "tasks/TFW-46__evidence_layer/research/iter2/4_challenge.md"


Challenge — "What could go wrong with our conclusions?"

Mindset: Critic. Your job is to break what Extract built. Find the failure modes. Test: "Did I find a reason our surviving configuration might FAIL in practice?" Parent: HL-TFW-46

Consistency Checks

C1: Step 10.5 + Convention + Reviewer (A1+C4) — Stress Test

Challenge: Does adding Step 11 (Evidence Collection) break the handoff flow's cognitive rhythm?

Current flow:

Implement → Test → Build gate → [Evidence] → Pre-RF Gate → Write RF

Test scenarios:

Scenario What happens at new Step 11 Breaks flow?
Web app task, Playwright available Navigate to deployed URL, take screenshot. 2-5 min ❌ Natural pause between "code works" and "write about it"
API backend task, curl available Run curl commands, capture output. 1-3 min ❌ Trivial, almost part of existing verification
Trivial task (fix typo in docs) TS has no Evidence fields → skip step entirely ❌ Proportional scope (D3) handles this
Complex multi-device task (Android + web + API) Need to deploy, set up scrcpy/adb, take screenshots across devices. 15-30 min ⚠️ Significant time, but this is EXACTLY where evidence matters most
No tools available (no Playwright, no deploy env) Executor marks DEFERRED for all evidence items ❌ Honest — DEFERRED is by design (D2)
Content task (blog post) Navigate to published URL, screenshot. Verify citations via curl. 5-10 min ❌ Manageable, catches real issues (TFW-36 citation fabrication)

Verdict: Step 11 doesn't break cognitive flow. It EXTENDS it naturally: "code works → actually works → document it." The proportional scope mechanism (coordinator can leave Evidence fields empty for trivial tasks) prevents bureaucratic overhead. The only pressure case is complex multi-device tasks — but those are precisely the tasks where evidence is most valuable.

Counter-argument considered: "Step 11 will be skipped because agents optimize for speed." Response: Same argument applies to Step 9 (tests) and Step 10 (build gate). Convention + reviewer gate enforcement is the same mechanism that enforces testing. If the reviewer checks for evidence completeness (Judge check #7), agents can't skip it without reviewer catching it.

C2: Coordinator Prediction (H2) — What if the coordinator gets the Evidence Plan wrong?

Failure modes:

  1. Coordinator specifies wrong tool — e.g., writes "Playwright" but app needs native mobile testing
  2. Fallback: Executor can deviate from TS Evidence Plan with justification in RF (same as Technical Guidance §6: "Executor MAY deviate with justification")
  3. Structural fix: Evidence field says what to verify, not mandating how. "Evidence: Verify login page renders correctly" is better than "Evidence: Playwright screenshot of login page"

  4. Coordinator misses an AC that needs evidence — e.g., forgets to add Evidence for the most critical AC

  5. Fallback: Executor proactively collects evidence even when Evidence field is empty (proactive tooling principle, HL §7.6)
  6. Structural fix: This is why the empty Evidence field means "executor decides" — not "no evidence needed"

  7. Coordinator over-specifies evidence for trivial task — unnecessary bureaucracy

  8. Fallback: Executor can mark N/A with reason in RF
  9. Risk: Low — the proportional scope principle (D3) guides coordinators

  10. Coordinator can't predict environment at TS time — e.g., deploy target unknown until execution

  11. Fallback: Evidence field: "DEFERRED — deploy target unknown at TS time. Executor determines environment."
  12. This is expected and handled by design

Verdict: H2 confirmed with qualification — coordinator prediction works as a STARTING POINT, not a contract. The executor has freedom to adapt. The Evidence field is guidance, not mandate (parallels Technical Guidance §6 "MAY deviate").

Important design implication: The Evidence field instruction in TS template should say: "What to verify in real environment. Executor MAY adapt tool and environment with justification." NOT: "Executor MUST follow exactly."

C3: Tooling Coverage (H4) — Is 60-65% realistic?

Challenge to the 60-65% automation estimate from Gather G2:

Domain G2 Estimate Counter-evidence Revised
Web UI 90% Complex multi-step flows, 2FA, OAuth redirects — Playwright handles most but not all 85%
API/Backend 95% External service dependencies, rate limits, auth tokens 90%
Database 90% Schema migrations, data integrity across services 85%
Documents 70% Print layout, fonts, encoding — Playwright renders but can't verify typographic quality 60%
Blog/CMS 80% Social previews, SEO crawlers, citation accuracy (TFW-36!) 65%
Mobile/Android 10% Project scan: no scrcpy. adb+logcat text evidence for data-plane. Visual = human-only 40-50% (data-plane 70%, visual 15%)
HR/Tenders 20% Portal rendering, published listing — mostly manual 20%
Analytics 85% Dashboard visual check — Playwright can screenshot dashboards 80%

Revised overall estimate: ~60-70% (weighted). H4's 70% threshold is borderline achievable when Android evidence is properly split into data-plane (automatable) vs visual (human).

Key insight from a mobile testing project: Evidence medium is domain-dependent — browser evidence is VISUAL (Playwright screenshots), Android evidence is DATA-PLANE (adb text output). Same anti-self-deception contract, different artifact type. TFW's Evidence table handles both: Artifact column accepts file paths (PNG) or inline text (adb output).

H4 verdict: 🟡 Borderline — 60-70% with proper domain decomposition. The original 70% is achievable if we count data-plane evidence for non-visual domains. The right framing: evidence coverage depends on available tooling PER PROJECT, not per framework.

C4: Renumbering Risk — Could it break existing projects?

Challenge: 22 file updates sounds manageable, but what about existing projects that forked TFW and have RF files with §5 Observations, §6 FC, etc.?

Analysis: - .tfw/ files: all updated in the Evidence Layer task — no drift - Existing RF files in tasks/: these are HISTORICAL artifacts. They reference old section numbers. This is expected — CHANGELOG.md already has historical §5 references from pre-TFW-25 renumbering - Knowledge topic files: reference RF §6, RF §7 — would need updating if they cite specific section numbers - compilable_contract.md: reference format resolver needs to handle both old and new numbering for historical compatibility

Verdict: Renumbering is a ONE-TIME mechanical operation. Historical artifacts keep their numbering (they're traces). Only .tfw/ framework files and active templates need updating. This has been done before (KNOWLEDGE.md §5→§4 in TFW-25, TS §4→§5 in TFW-41). Not a blocker.

C5: Domain-Agnosticity (H1) — Final Validation

Test: Does the Evidence table work for ALL domain types from HL §7.4?

Domain AC example Evidence field Status Artifact type
Code (web) Login page renders Navigate to localhost:3000, screenshot VERIFIED evidence/login.png
Code (API) Endpoint returns 200 curl -v /api/health, paste output VERIFIED inline
Analytics Report shows correct totals DB MCP query, paste result VERIFIED inline
Blog Post renders without broken styles Playwright screenshot of published URL VERIFIED evidence/blog_render.png
HR Job listing published on portal Manual check on portal — DEFERRED (needs user login) DEFERRED
Design Mockup matches specification Visual comparison — DEFERRED (needs human judgment) DEFERRED
Document PDF opens without encoding issues Open PDF, screenshot first page VERIFIED evidence/doc_page1.png
Mobile app App launches, main screen loads scrcpy screenshot via adb (if available) — or DEFERRED VERIFIED/DEFERRED evidence/app_screen.png or —

Verdict: H1 ✅ confirmed — the Evidence table + 4-status vocabulary works across all domains. The key is that DEFERRED is not a failure — it's honest reporting. Domain-specific differences are in evidence TYPE (screenshot vs query vs curl output), not in evidence STRUCTURE (the table is universal).

C6: "Evidence becomes another checkbox" (HL DoF-1) — Stress Test

The most critical failure mode from HL §6.

How it could happen: 1. Agent reads Evidence table, writes "VERIFIED" for all items without actually doing anything 2. Agent takes a screenshot of localhost but the app isn't actually running (empty page) 3. Agent writes plausible evidence descriptions from memory without executing

How the design prevents it:

Prevention Mechanism Where
Artifact requirement VERIFIED must reference a file path or inline output Anti-pattern R1 in §14
Temporal ordering Evidence collected BEFORE RF writing (Step 11 before Step 12-13) Handoff.md flow
Reviewer verification Reviewer checks that artifacts exist and match claims Review verify.md, Judge check #7
DEFERRED honesty DEFERRED is structurally easier than fabricating evidence 4-status vocabulary (D2)
Proportional scope Trivial tasks can skip evidence entirely (empty Evidence fields) TS design, D3

Verdict: The design makes fabrication HARDER than honest reporting. Fabricating a plausible screenshot requires deploying the app (which IS the evidence). Writing DEFERRED takes 5 seconds. The incentive structure favors honesty.

Remaining risk: Agent takes real screenshot but of wrong thing (e.g., screenshots the code instead of the running app). This is a quality-of-evidence issue, not a structural issue. The reviewer catches it during Evidence Audit.


OODA Loop Summary

Loop 1: Stress-tested A1+C4 (Step 10.5 + convention + reviewer) against 6 scenarios. Holds. Key: proportional scope prevents bureaucracy, reviewer gate prevents skipping.

Loop 2: Challenged coordinator prediction (H2). Confirmed with qualification: Evidence field is guidance, not contract. Executor MAY adapt. Important template instruction implication.

Loop 3: Challenged tooling coverage (H4). Refuted 70% target, reframed: any automation > 0% is a win. DEFERRED/BLOCKED handles the rest. Pending: Android scrcpy/adb data from a mobile testing project scan.

Checkpoint

Sufficiency Verdict

  • [x] External source used? — Compliance contemporaneous documentation principle, industry "shift-left verification" pattern
  • [x] Briefing gap closed? — All surviving configurations stress-tested
  • [x] Hypothesis tested? — H1 (✅ confirmed), H2 (✅ confirmed with qualification), H4 (❌ refuted as stated, ✅ reframed)
  • [x] Counter-evidence sought? — Yes: coordinator prediction failures, DoF-1 checkbox risk, renumbering compatibility, tooling automation limits

Metacognitive Check

The most valuable NEW insight from Challenge is the reframing of H4. The original question "Can 70% of evidence be automated?" is the WRONG question. The right question is "Does the design handle the gap between automated and manual evidence honestly?" And it does — via DEFERRED/BLOCKED. This reframing eliminates a whole class of objections ("but you can't automate evidence for HR documents!") — true, and that's fine, because the design accounts for it structurally.

The second new insight is that the Evidence field in TS should be phrased as GUIDANCE, not mandate — paralleling Technical Guidance §6's "MAY deviate" clause. This prevents the coordinator from over-constraining the executor.

Pending: Mobile testing project Android evidence scan results (scrcpy/adb patterns). Will integrate into RES when available.


Stage complete: YES