Forge
markdowne8ad0934
1# Epistemic methods and calibration v1
2
3Тела секций epistemic/calibration/hpmor из `agent-notes.md`. L0 `epistemic-default-distrust-v1` — выше `public-cut`. Extended: `knowledge/epistemic-default-distrust-extended-v1.md`.
4
5---
6<!-- section:calibration-playbook-14d-v1 -->
7## Calibration Playbook 14D (под нас) v1
8
9Статус: active
10Период: 14 дней
11Режим: lightweight, без бюрократии.
12
13### Day 0: Baseline
141) Зафиксировать стартовые значения:
15- TTUC
16- decision latency
17- execution start lag
18- recovery SLA
19- rework rate
20- agency gain (0-10)
21
222) Определить 3 типовых сценария:
23- рабочий стрессовый,
24- творческий/исследовательский,
25- low-resource день.
26
27### Day 1-7: Instrument only
28- Ничего радикально не менять.
29- Просто собирать метрики и наблюдать паттерны.
30- Ловить top-3 повторяющихся friction points.
31
32Daily log (2-3 минуты):
33- today_top_friction:
34- action_now:
35- TTUC:
36- recovery_time:
37- agency_0_10:
38
39### Day 8-14: Targeted intervention
40- Взять top-3 friction points.
41- На каждый — 1 минимальный фикс (не big-bang).
42- Для сложных задач обязательно representative-model-scaling.
43- Для причин — N-Why с dual stop.
44
45### Review (end of day 14)
461) Что ускорилось устойчиво?
472) Где был ложный прогресс?
483) Какие ритуалы реально работают?
494) Что закрепляем как постоянный стандарт?
50
51### Exit criteria
52- >=20% улучшение минимум по 2 core metrics ИЛИ
53- заметное снижение recovery SLA + rework rate при сохранении safety.
54
55### If no improvement
56- Пересобрать модель (не винить исполнителя).
57- Проверить level collapse и boundary blindness.
58- Запустить ещё один 7-дневный микроцикл с другим primary-подмиром.
59<!-- /section:calibration-playbook-14d-v1 -->
60
61<!-- section:epistemic-cognitive-failure-modes-v1 -->
62## Epistemic Cognitive Failure Modes v1
63- Failure mode: motivated reasoning.
64 - Early signal: evidence selection favors preferred outcome.
65 - Countermeasure: require strongest counter-argument before final decision.
66- Failure mode: planning fallacy.
67 - Early signal: repeated optimistic deadlines with low variance.
68 - Countermeasure: reference-class forecasting from past similar tasks.
69- Failure mode: confirmation bias.
70 - Early signal: searches/tests only support current hypothesis.
71 - Countermeasure: mandatory disconfirming test.
72- Failure mode: halo effect in tool/vendor choice.
73 - Early signal: global positive impression replaces requirement fit.
74 - Countermeasure: score options by explicit constraints/SLO/cost.
75- Failure mode: narrative overfitting.
76 - Early signal: "beautiful" explanation with weak predictive power.
77 - Countermeasure: prefer simpler model with better falsifiability.
78- Failure mode: ambiguity denial.
79 - Early signal: binary answers for uncertain state.
80 - Countermeasure: use confidence intervals and reversible decisions where possible.
81- Failure mode: status-quo inertia.
82 - Early signal: known-bad path persists due to migration fear.
83 - Countermeasure: bounded experiment with rollback safety.
84<!-- /section:epistemic-cognitive-failure-modes-v1 -->
85
86<!-- section:epistemic-daily-checklist-v1 -->
87## Epistemic Daily Checklist v1
88- What do we believe right now, and with what confidence?
89- What is our prior before seeing new evidence?
90- What would change our mind today (explicit disconfirming evidence)?
91- Did we write observations separately from interpretation?
92- Did we test mechanism-level causality (not just narrative fit)?
93- What is the global objective and success metric (before tactics)?
94- Are we using proxy metrics without divergence monitoring?
95- Are we stuck in a fake dichotomy (is there a third option)?
96- Did we check base rates before trusting anecdotes?
97- Which assumption is most likely wrong and most expensive if wrong?
98- What single observation can most reduce uncertainty today?
99- Are we calibrated (confidence aligned with recent hit rate)?
100- Is there a reversible experiment before irreversible commitment?
101- Did we run a pre-mortem for high-impact decision?
102- Did we model incentives, not just stated intentions?
103- Did we account for second-order effects?
104- Did we red-team critical plan assumptions?
105- Is decision traceable to evidence and constraints?
106- Is rollback path defined if assumption breaks?
107- Did we complete full validation chain before closure?
108- Did KPI improvements preserve guard metrics (quality/cost/reliability)?
109- Is process quality sound independent from outcome luck?
110- Is this learning transferable and reproducible by another reviewer?
111- Did we update memory with distilled learning after outcome?
112<!-- /section:epistemic-daily-checklist-v1 -->
113
114<!-- section:epistemic-hpmor-heuristics-v1 -->
115## Epistemic HPMOR Heuristics v1
116- Heuristic: Belief as probability, not identity.
117 - Prevents: dogmatic lock-in and defensive reasoning.
118 - Operational check: explicitly state confidence range before decision.
119 - Revisit trigger: new evidence with high likelihood ratio appears.
120- Heuristic: Confusion is a signal, not a failure.
121 - Prevents: fake certainty and narrative patching.
122 - Operational check: if model feels inconsistent, pause and list unknowns.
123 - Revisit trigger: mismatch between prediction and observed outcome.
124- Heuristic: Make predictions that can fail.
125 - Prevents: unfalsifiable explanations.
126 - Operational check: attach at least one measurable expected outcome.
127 - Revisit trigger: outcome not observed in expected window.
128- Heuristic: Avoid affect heuristic in technical decisions.
129 - Prevents: preference-driven architecture choices.
130 - Operational check: separate "I like it" from "it meets constraints".
131 - Revisit trigger: tradeoff analysis cannot be reproduced by another reviewer.
132- Heuristic: Scope uncertainty explicitly.
133 - Prevents: hidden assumptions and accidental overreach.
134 - Operational check: mark assumptions and their confidence.
135 - Revisit trigger: assumption invalidated by runtime/production evidence.
136- Heuristic: Distinguish map from territory.
137 - Prevents: treating documentation/model as reality.
138 - Operational check: verify with instrumentation and direct signals.
139 - Revisit trigger: docs/model disagree with live behavior.
140- Heuristic: Update fast, not defensively.
141 - Prevents: sunk-cost continuation of bad decisions.
142 - Operational check: define in advance what evidence will trigger change.
143 - Revisit trigger: trigger condition met.
144<!-- /section:epistemic-hpmor-heuristics-v1 -->
145
146<!-- section:epistemic-methodology-v1 -->
147## Epistemic Methodology v1
148- Scope: general cognition methodology (outside pure IT stack), used to improve reasoning quality in any domain.
149- Current source anchor: HPMOR (in progress, targeted for completion as active track).
150- Working principles:
151 - separate observation, inference, and hypothesis confidence;
152 - prefer falsifiable claims over rhetorical certainty;
153 - track update triggers: what evidence should change current belief;
154 - minimize identity lock-in to previous conclusions;
155 - optimize for truth-seeking and decision quality, not verbal elegance.
156- Integration rule: epistemic principles shape decision process globally, while domain sections (IT/Portal/etc.) store concrete technical playbooks.
157- HPMOR completion pipeline:
158 1) finish remaining chapters,
159 2) extract explicit reasoning heuristics,
160 3) map each heuristic to decision/diagnostic behavior,
161 4) attach source anchors to original works where referenced,
162 5) publish concise daily checklist for applied reasoning.
163- Output format:
164 - heuristic,
165 - failure mode it prevents,
166 - operational check,
167 - revisit trigger.
168<!-- /section:epistemic-methodology-v1 -->
169
170<!-- section:epistemic-to-domain-bridge-v1 -->
171## Epistemic to Domain Bridge v1 (general)
172- Scope: general methodological layer for any domain (IT, science, strategy, education, product, research, communication).
173- Core mapping:
174 - probabilistic belief updates -> confidence-aware decisions,
175 - falsifiability -> measurable success/failure criteria,
176 - map vs territory -> prioritize direct evidence over narrative assumptions,
177 - anti-confirmation-bias -> require at least one disconfirming check,
178 - planning-fallacy control -> reference-class forecasting,
179 - ambiguity handling -> reversible decisions under uncertainty,
180 - fast Bayesian update -> rapid pivot/rollback when strong evidence appears.
181- Operational rule:
182 - no major decision without explicit confidence statement and falsification trigger.
183- Domain adaptation:
184 - each domain keeps its own metrics and diagnostics, but epistemic checks are shared.
185- Memory policy:
186 - epistemic base remains global;
187 - domain sections consume it as applied playbooks.
188<!-- /section:epistemic-to-domain-bridge-v1 -->
189
190<!-- section:epistemic-to-it-bridge-v1 -->
191## Epistemic to IT Bridge v1
192- Heuristic class -> IT application pattern:
193 - probabilistic belief updates -> confidence-aware architecture decisions,
194 - falsifiability -> measurable success/failure criteria in rollouts,
195 - map vs territory -> prefer instrumentation over assumption,
196 - anti-confirmation-bias -> mandatory disconfirming test before merge,
197 - planning-fallacy control -> reference-class estimates for delivery timelines,
198 - ambiguity handling -> reversible decisions under uncertainty,
199 - fast Bayesian update -> quicker rollback or strategy pivot on strong evidence.
200- Operational rule:
201 - no major technical decision without explicit confidence statement and falsification trigger.
202- Diagnostics tie-in:
203 - epistemic checks run alongside latency/error/saturation metrics in incident and architecture reviews.
204<!-- /section:epistemic-to-it-bridge-v1 -->
205
206<!-- section:hpmor-l1-pool-v1 -->
207## HPMOR (L1)
208Эпистемические эвристики из ГПиМРМ вынесены в L1. По запросу: route_context('HPMOR'|'epistemic heuristics') или knowledge/agent-notes-l1-pool.md.
209<!-- /section:hpmor-l1-pool-v1 -->
210
211
212
View only · write via MCP/CIDE