Diffract assumes that reviewers — human and AI — will fail. Not out of malice, but because of cognitive fatigue, confirmation bias, time pressure, and social dynamics. The framework makes dishonesty structurally difficult through 8 concrete mechanisms, each adapted from a high-stakes industry where inspection failures have life-or-death consequences.
Source: Aviation maintenance — RFID scanning forces physical presence at the inspection site before a task can be signed off.
In Diffract: Every lens must produce file:line evidence or an explicit
list of what was checked. A claim of "all clean" without evidence is invalid.
Source: Japanese railways and nuclear control rooms — operators must physically point at each indicator and verbally call out its state. This engages motor, visual, auditory, and verbal processing simultaneously, breaking the "looking without seeing" failure mode.
In Diffract: When a lens produces no findings, the reviewer must describe what a finding would look like. This proves they understand the lens and actually examined the code, not just waved at it.
### 🎯 Variety
Checked: all switch/match statements on HTTP status codes
A finding would look like: a status code reaching the default branch
that should have explicit handling (e.g., 503 for retry logic).
No findings matching this pattern."No findings" without cognitive anchoring is not allowed.
Source: Philosophy of science (Karl Popper) — a claim is only valid if it can be disproven.
In Diffract: Every finding must be grounded in an objective, measurable criterion. "This variable is never read" is falsifiable (grep the codebase). "This variable name feels wrong" is not falsifiable and should be discarded unless the reviewer can explain what it should be called and why.
Source: Metrology (measurement science) + Radiology (dual-reading) — two independent radiologists read the same images. The final report only proceeds when both findings are reconciled.
In Diffract: A different reviewer applying the same lenses to the same code should reach the same conclusions. This is not aspirational — it's testable.
Calibration Test: After a review is complete, a second reviewer (human or AI at the same capability level) independently applies the same lenses. If they produce zero new findings, the review is calibrated. If they find issues the first reviewer missed, the review is incomplete.
Source: Three independent domains:
- UXO clearance — inert munitions secretly buried in grids before sweep teams arrive. If a team declares a grid "clear" but misses a planted seed, the entire grid is resurveyed.
- Radiology — slides with known abnormalities mixed into daily screening queues. If a pathologist misses one, thresholds are recalibrated.
- Legal e-discovery — pre-coded documents seeded into review pools. If reviewers miss them, the algorithm is retrained.
In Diffract: After a "nothing found" round, ask: "If I deliberately introduced a bug in this lens's domain, would my process have caught it?" If not, the process failed — not the code.
Source: Aviation Crew Resource Management (CRM) — the monitoring pilot reads the challenge ("Landing Gear"), the flying pilot must physically verify and verbally respond ("Down and Green"). This creates mutual verification that nullifies single-person dominance.
Also related: the Sterile Cockpit Rule — below 10,000 feet, all non-essential conversation is banned to protect cognitive bandwidth during critical phases.
In Diffract: In panel reviews, the monitoring panel must actively challenge each lens's conclusions. Not "I agree" — but "I see your evidence; I challenge on X." Passive agreement is not allowed.
Source: Aviation — Required Inspection Items (RII). The mechanic who performs a repair is legally barred from signing off on the inspection. An independent inspector, reporting to a separate management chain, must validate the work.
In Diffract: Lenses find issues. Governors decide whether to fix them. The reviewer who identifies a finding does not unilaterally decide its disposition — the governors (Compass, Cobra) make that determination.
Source: Continuous improvement (Deming, Toyota Production System)
In Diffract: After every review cycle, ask:
- What did the framework miss?
- Were we honest?
- Should the framework itself be updated?
The framework evolves through its own PDCA cycle.
| # | Mechanism | Source Industry | What It Prevents |
|---|---|---|---|
| 1 | Evidence | Aviation (RFID) | Claims without observation |
| 2 | Cognitive Anchoring | Railways (shisa kanko) | Looking without seeing |
| 3 | Falsifiability | Philosophy (Popper) | Opinion disguised as fact |
| 4 | Calibration | Metrology + Radiology (dual-reading) | Reviewer-dependent outcomes |
| 5 | Nothing-Found Verification | UXO / Radiology / Law | False negatives |
| 6 | Challenge-Response | Aviation (CRM) | Passive agreement |
| 7 | Finder/Decider Separation | Aviation (RII) | Conflict of interest |
| 8 | Retro | Manufacturing (Deming) | Framework stagnation |