Domain adaptation — modes 3/4/5 beyond software

Mirrored from the skill package. This is the canonical English text that Claude actually loads, published here so it can be read next to its Turkish mirror instead of only on GitHub. Edit the source above, not this page; the Turkish mirror is docs/tr/alan-uyarlama.md.

Mizan Domain Adaptation — Modes 3/4/5 Beyond Software

Modes 3, 4, and 5 are software costumes on three universal patterns:

• Mode 3 (audit): the claim-bearing artifact is never the behavior-bearing artifact — verify every hop between them.
• Mode 4 (anomaly registry): anomaly → mechanism hypothesis → rival hypotheses → discriminating test. Never close on the first story that fits.
• Mode 5 (commitment gate): any forward resource commitment carries value/cost/dependency claims — tier them, lock thresholds and a kill condition BEFORE committing, force alternatives incl. the null.

The adaptation recipe (use this for any domain not listed)

Answer six questions; the answers ARE the domain module. Question 0 is the one that decides how much of the software-verification loop actually survives the move:

1. Arbiter: who returns the verdict on a locked threshold in this domain — a deterministic executor, an independent instrument, a third party, or the claim’s own author? Software’s rigor is borrowed from the runtime, not from the fact that the artifact is code; most domains inherit the protocol and lose the judge. Answer honestly per entry (schema field hypothesis.arbiter, rule R8): author caps the claim at a permanent [KKE], none means the threshold is decorative and the entry stays [S]. A domain where every arbiter is author is not thereby un-auditable — it just cannot produce [K], and saying so is the whole point.

2. Hop map: where do claims live vs. where does evidence live? (In code: comment vs. implementation. In sales: CRM field vs. activity log.)
3. Instruments: what produces the numbers, and what are its known distortions?
4. Confound catalog: what boring explanation produces the same result in this domain?
5. Ground-truth latency: how fast and how cheaply can a claim be refuted? (Seconds in code; quarters in marketing.) Set expectations: slow domains make “underpowered” the NORM, not the exception.
6. Prior art: which existing discipline in this domain is already a loose version of this? (Name it in entries; Mizan formalizes, it rarely invents.)

Universal rules that transfer unchanged: append-only history; refuted entries never deleted; surprising positives wait for a symmetric control; where a symmetric control is unethical or impossible (human-subject domains), the claim carries a PERMANENT [KKE] as an honesty label, not a defect. DC-001 applies wherever individual hit rates could be weaponized (sales forecasts, analyst calls, hiring decisions): individual scores stay with the individual; management sees aggregates only.

Domain catalog

Every entry answers question 0 and question 4 on one line, Arbiter / latency, because those two set the entry’s ceiling before any threshold is written: an author arbiter caps at a permanent [KKE] (R8), and a latency of quarters or years makes “underpowered” the expected outcome. Where a domain has two arbiters for two kinds of claim, both are named — the split is usually the most useful thing the line says.

1. Data & Analytics

• Hop map: metric label / dashboard title / report narrative ↔ the actual query (SQL/pipeline) ↔ the raw data. “Active users” on the chart vs. what the WHERE clause counts.
• Mode 4 = metric-anomaly investigation: rival hypotheses for any drop/spike — tracking change, seasonality, mix shift (Simpson’s paradox), real effect.
• Confounds: instrumentation changes, backfills, timezone/window definitions, survivorship in cohorts.
• Prior art: data-quality testing (dbt tests ≈ assertions on claims).
• Arbiter / latency: runtime for “does the query count what the label says” — re-run it, [K] is reachable; instrument for an anomaly’s cause. Minutes to days.

2. Marketing & Growth

• Hop map: campaign brief / landing-page promise ↔ product’s actual behavior; audience claims ↔ survey/telemetry evidence.
• Mode 5 = campaign gate: CAC/conversion threshold locked pre-launch; spend cap = kill condition; null alternative (“don’t run it”) priced.
• Confounds: novelty effect, seasonality, cannibalization of sibling channels, peeking (early stopping = R1 violation).
• Prior art: proper A/B methodology IS preregistration; [Y] was practically invented for marketing copy.
• Arbiter / latency: instrument (the experiment platform, analytics) for conversion claims; author for positioning and brand claims, which therefore carry a permanent [KKE]. Weeks to quarters — attribution windows make most closures underpowered.

3. Sales & CRM

• Hop map: pipeline-stage field / “champion identified” claim ↔ activity log, buyer-side artifacts (emails, meetings).
• Mode 5 = deal-qualification gate (MEDDIC/BANT = loose prior art); disqualification criteria = kill condition; close-probability = preregistered prediction.
• Mode 4 = loss analysis: seller’s post-hoc story vs. buyer’s stated reasons vs. rival mechanisms.
• Confounds: quarter-end pressure, discount effects, single-threaded contact masquerading as consensus. DC-001 is culturally hardest here.
• Arbiter / latency: instrument (won/lost in the CRM) for a forecast; the seller’s own loss story is author, and only buyer-side evidence (third_party) lifts it. One sales cycle — weeks to months.

4. Academic / Scientific Research

• Native habitat — Modes 1/2 came from here. The genuinely new transfer is Mode 3 on literature: abstract claims ↔ methods/results evidence (abstract-inflation is documented tier drift), citation claims ↔ what the cited paper actually shows.
• Confounds: publication bias, the garden of forking paths, small-sample effects that regress on replication, a review cited in place of the primary study it summarises.
• Prior art: preregistration, registered reports, PRISMA.
• Arbiter / latency: for a citation hop the instrument is reading the cited paper — anyone can, so [K] is reachable in hours; for a finding, third_party replication, in years.

5. Finance & Investment Decisions

• Hop map: investment thesis ↔ position; “we believe X because Y” ↔ the data Y actually shows.
• Mode 5 = position gate: entry thesis with refutation condition (thesis-invalidation ≠ price stop-loss — record both), sizing as cost claim.
• Confounds: market beta dressed as alpha, regime luck, survivorship in backtests, overfitting to history.
• Prior art: investment memos + pre-mortems; trading journals are informal Mode 4 registries.
• Note: Mizan structures the reasoning; it is not financial advice machinery and does not pick trades.
• Arbiter / latency: instrument (reported financials) for the thesis’s operating milestones. Price is NOT the thesis’s arbiter: it moves for reasons the thesis does not claim, so a price move alone neither confirms nor refutes it. Quarters for milestones; a full regime for a strategy claim.

6. Operations / Manufacturing / Logistics

• Mode 4 = root-cause analysis formalized: 5-Whys chains are mechanism hypotheses that usually skip rival hypotheses and discriminating tests — Mizan adds exactly those. “Fix worked” after a process change follows the revert-check rule where feasible.
• Mode 3: SOP/work-instruction claims ↔ what the floor actually does.
• Confounds: Hawthorne effect (observation changes behavior), concurrent changes, demand mix.
• Prior art: A3/8D reports, Six Sigma DMAIC.
• Arbiter / latency: instrument (defect counts, cycle times); a feasible revert-check makes the verdict nearly runtime. A shift to a few weeks.

7. Security & Incident Response

• Mode 4 is nearly isomorphic to IR: symptom (alert/IOC) → intrusion hypothesis → rival hypotheses (misconfig? scanner noise? true compromise?) → discriminating evidence. Post-incident reports are HARKing magnets — timeline claims need artifact citations.
• Mode 3: security-posture claims (docs, compliance answers) ↔ actual configs and controls.
• Prior art: blameless postmortems (= DC-001’s ancestor), ATT&CK hypothesis hunting.
• Arbiter / latency: runtime for a configuration claim — read the config, run the control. For “no compromise” there is no arbiter that can return a pass: silence caps at [KKE] (R26). For an asset-centred pass rather than an incident, use Mode 7 (security-probe.md). Minutes for config; intrusion dwell time is measured in months.

8. Hiring & People Decisions

• Hop map: job-description and scorecard claims (“strong systems thinker”) ↔ the evidence that produced them (work sample, the interview answer itself, references).
• Mode 5 = hire gate: role’s problem claim, success metric at 90 days locked before the offer, kill condition for the ROLE (not the person) if premises fail.
• Interview signals = hypotheses; preregistered predictions per interviewer make calibration measurable over many hires.
• Confounds: halo effect, market conditions, onboarding quality confounded with selection quality.
• HARD constraint: DC-001 fully applies; per-interviewer hit rates never become performance weapons. Human-subject symmetric controls are mostly impossible → permanent [KKE] labels are normal here.
• Prior art: structured interviewing, work-sample tests.
• Arbiter / latency: almost always author or a manager as third_party; a 90-day metric is rarely an instrument → [K] is close to unreachable, say so. 90 days to a year, with small n.

9. Procurement & Vendor Selection

• Mode 3: vendor claims (SLA, benchmark decks, “enterprise-ready”) ↔ contract terms ↔ measured behavior in POC. Vendor benchmarks are [Y] until reproduced.
• Mode 5 = RFP gate: requirements as tiered claims; dependency claims (“integrates with our stack”) verified BEFORE signing; exit/switch cost recorded as kill-condition economics.
• Confounds: demo-environment vs. production, reference-customer selection bias.
• Prior art: POC bake-offs, weighted scoring matrices.
• Arbiter / latency: instrument (POC measurement) only if the POC was designed before signing; a vendor’s own benchmark is the vendor as author. Weeks for a POC; months for real SLA behaviour.

10. Legal / Contracts / Compliance

• Mode 3: policy/compliance claims (“we are GDPR-compliant”) ↔ actual clauses, actual data flows; marketing promises ↔ contractual obligations (a promise not in the contract is [H] at best).
• Mode 4: dispute analysis — each side’s narrative as rival hypotheses against the documentary record.
• Confounds: selective document production, hindsight shaping each side’s narrative, a clause read without the definitions it depends on.
• Prior art: legal due diligence; control testing in compliance audits (SOC 2, ISO 27001).
• Arbiter / latency: reading the clauses against the actual data flows is instrument-like for Mode 3; the real third_party (court, regulator, auditor) usually arrives only with a dispute. Days for the document check; years for a dispute.
• Note: structures the evidence; not legal advice.

11. Product / UX Research

• Mode 3: “users want X” claims ↔ interview transcripts (what was actually said vs. the summary — summarization drift is tier drift), usability-report claims ↔ session recordings.
• Confounds: leading questions, sample skew toward vocal users, say-do gap (stated preference vs. behavior).
• Prior art: continuous-discovery practices, evidence-based design.
• Arbiter / latency: instrument (task success, behavioural telemetry) or moderated sessions as third_party; the researcher’s synthesis is author. Days for sessions; weeks for telemetry. For auditing an application’s UX rather than research claims, the sibling skill ux-mizan carries the full procedure.

12. Content / Journalism / Technical Writing

• Mode 3: headline ↔ body ↔ source (headline inflation = tier drift); every factual claim’s hop to a primary source.
• Mode 5 = story/content gate: audience-value claim, distribution dependency claims, evergreen-vs-decay expectation as preregistered prediction.
• Confounds: two outlets citing the same wire story counted as two sources, survivorship in “what performed”, a platform algorithm change read as audience preference.
• Prior art: fact-checking desks; [Y] and [KKE] map directly onto editorial standards.
• Arbiter / latency: the primary source as third_party for a factual claim — [K] reachable within hours; instrument (analytics) for a content-value prediction, over weeks.

13. Policy / Program Evaluation (public sector, NGO)

• Hop map: program reports (activities delivered — outputs) ↔ the outcome data those activities were meant to move. Outputs are not outcomes; a report that counts the first is silent on the second.
• Mode 5 = program gate: theory-of-change as a chain of tiered claims; sunset clause = kill condition (rare in practice, transformative when preregistered).
• Confounds: selection into programs, secular trends, regression to the mean in targeted populations.
• Arbiter / latency: an independent evaluator as third_party, instrument where administrative data exists; the implementing body judging itself is author. Years — underpowered is the norm; say so in every entry.
• Prior art: RCT evaluation culture, logic models.

14. Personal Experimentation (health, fitness, productivity)

• Hop map: the felt improvement or diary entry ↔ the measured value (wearable, lab result, logged output) — the gap between them is the expectancy confound made visible.
• Mode 4 on n=1: symptom → mechanism hypothesis → the cheapest discriminating change, ONE variable at a time; washout periods as the personal symmetric control.
• Confounds: placebo/expectancy, regression to the mean (you start interventions at your worst), season/sleep/stress co-movement.
• Permanent honesty labels: n=1 means most closures cap at [H]; that is the honest ceiling, not failure.
• Prior art: quantified-self practice, n-of-1 trials in clinical research.
• Arbiter / latency: the subject is also the author, so closures sit at [H] or [KKE]; an instrument (wearable, lab test) raises what it measures, not the felt claim. Days to weeks per cycle, washout included.
• Note: structures self-observation; medical decisions belong with clinicians.

Anti-pattern for this file

Do not force all five gate fields onto domains where they parody themselves. If a domain entry’s kill condition or metric feels theatrical, record the honest version: “no credible instrument exists → claim stays [S]” is a legitimate, useful outcome.