View source for Builder Protocol v0.1: Metrology of the Abstract applied to AI

{{Development Notice
|status = Experimental — v0.1.1 Protocol
}}

= Builder Protocol v0.1.1: Metrology of the Abstract applied to AI =

'''Status:''' Experimental · Self-Assessed · Procedure only · not pilot-proven  

'''Cites:'''
* [[Metrology of the Abstract]] (discipline)
* [[Metrology of the Abstract: Application to Artificial Intelligence]] (concept)
* [[Operator Insight: Freedom under the Chain & Grounding under Ambiguity]] (optional operator note)

This is a '''builder protocol''': what to run in practice. It is not a new concept page and not a claim of solved confirmation at scale.

If this protocol conflicts with the discipline or concept page, '''those pages win''' until this protocol is revised.

Implementation Note: This pilot uses the current MediaWiki implementation (templates, categories, Resource blocks, Admin Page Status) as the working environment. These elements are implementation-specific and are not required by Metrology of the Abstract itself. Other implementations may use different mechanisms while following the same builder protocol.

----
{{GameModule
| type = Calibration Procedure
| category = [[:Category:Meta & Framework|Meta & Framework]]
| Calibration Type = Calibration Layer
| Application Layer = AI Systems / Human–AI Collaboration
| Version = 0.1.1
| Maturity = Experimental
| Last Updated = 2026-07-23
| description = Thin builder protocol for applying Metrology of the Abstract to AI claim pipelines: retrieve working standards, tier claims, label standing, log calibration notes, sample for independent check, respect user standards. Implements the concept page; does not expand doctrine. v0.1.1 adds solo Outcome-checked path and Confirmed clarification.
}}

== 0. Purpose and scope ==

=== Purpose ===
Give builders a minimal, repeatable procedure so AI-mediated work stays under a standards chain:

* model = instrument (not reference)
* claims are tiered
* standing is honest
* working standards are durable and inspectable
* user standards are used, not silently overwritten

=== In scope ===
* Human+AI workflows (research, wiki, product, procedural know-how, analysis)
* Drafting and using secondary / Nth working standards
* Claim labeling, cal notes, and light sampling

=== Out of scope (v0.1.1) ===
* Full lab eval harness design
* Automatic independent confirmation at internet scale
* Solving tier-C contested truth
* Replacing safety / RLHF / eval organizations

----

== 1. One-time setup ==

Do this once per project or domain (revise rarely).

=== 1.1 Adopt top-level rules (cite, don’t reinvent) ===
Post or link parent rules as binding:

* Instrument ≠ reference
* Self-check ≠ Confirmed
* Same lineage / weights ≠ independent confirmation
* Agreement among correlated systems ≠ Confirmed
* Standing is tier-qualified (A/B/C)
* Cal reports carry their own standing

=== 1.2 Create a Working Standards library ===
A folder, wiki category, repo path, or project memory for versioned artifacts.

Each stored working standard should eventually include:

* Name / version / date
* Purpose (task class)
* Pointer '''up''' (which top-level rules it implements)
* Standing (at minimum: Self-Assessed)
* Body (procedure or checklist)

=== 1.3 Register user standards ===
If the user or org has existing files/rules:

* Index them in the library (or link them)
* Default policy: '''apply as given'''
* Refine '''only on request''' (or explicit written protocol the user accepts)

=== 1.4 Define a minimal sampling policy ===
v0.1.1 does not require universal human review. It requires an explicit rule such as:

* Risk-weight high-stakes or irreversible claims for independent check; '''or'''
* Random sample N% of claim batches; '''or'''
* Always check tier-A procedural claims when a real test is cheap

Write the rule down. If you cannot staff a second party, '''say so'''—do not imply Confirmed standing you cannot support.

'''Solo-builder fallback:''' If no second human or diverse second system is available, a '''cheap tier-A outcome check''' (procedure works/fails, measurement, reproducible test run by the builder) counts as '''external contact relative to the model'''. The outcome is outside the model. Label clearly, e.g. '''Outcome-checked (tier A, solo)''' — not bare “Confirmed by independent reviewer.” Do '''not''' treat “I re-read my own answer” or “the same model agreed” as that external contact.

----

== 2. Per-task procedure (default loop) ==

Run this loop for each meaningful task, thread segment, or deliverable.

=== Step 1 — Classify the claim tier ===
{| class="wikitable"
! Tier
! If…
! Then…
|-
| '''A'''
| Outcome or test can vote (procedure works/fails, measurement, reproducible check)
| Prefer test/outcome language; strongest standing path available
|-
| '''B'''
| Checkable against a defined corpus or source
| Prefer citation/fidelity language—not “truth”
|-
| '''C'''
| Contested, values-laden, or no external referee
| Traceability and disagreement only—'''no fake Confirmed'''
|}

If unsure, '''label the lower tier''' (more conservative).

=== Step 2 — Retrieve before inventing ===
# Search Working Standards library + user standards for a matching task class.
# If found → '''apply''' it; note which standard/version was used.
# If not found → go to Step 3 (draft).
# Do '''not''' regenerate a new “standard” every turn when a checked one exists.

=== Step 3 — Draft a working standard only when needed ===
Draft a new secondary/Nth standard only if:

* task class is novel; or
* existing standard fails in use; or
* user explicitly requests a new or refined standard

Draft requirements (minimal):

* Clear steps
* Stated limits / uncertainty
* Tier assumptions
* Up-pointer to top-level rules
* Standing: '''Self-Assessed''' until promoted

Share the draft with the user for inspect/verify before treating it as stored project standard.

=== Step 4 — Apply user standards correctly ===
* Prefer user file standards when they match the task.
* Do not silently overwrite wording or policy.
* If improvement seems warranted: '''propose''' refinement; apply only on request (or under pre-agreed protocol).
* User standards can themselves be calibrated (ownership ≠ immunity)—label conflict openly.

=== Step 5 — Produce the work as measurements under a method ===
When generating claims:

* Stay in instrument role (not authority cosplay)
* State uncertainty where standing is thin
* Flag disagreement when using multiple instruments
* Escalate or abstain when reference is insufficient (especially tier C)

=== Step 6 — Write a minimal calibration note ===
For non-trivial outputs, attach a short note (comment, footer, log entry, or talk line):

'''Minimal cal note fields:'''

* What was checked
* What was assumed
* Claim tier(s)
* Standard used (top / secondary / user file / none—improvised)
* Whether user standard was applied, ignored, or refined
* Standing of '''this note''' (usually Self-Assessed)
* Open issues / escalations

Thin status labels without these fields count as narration, not calibration.

=== Step 7 — Sample for stronger standing (when policy requires) ===
If the sampling policy triggers:

* Obtain independent check (human practitioner, tier-A test, or diversely trained system used only as '''disagreement sensor'''—not as automatic Confirmed); '''or''' use the solo tier-A outcome path in §1.4 when no second party exists
* Record result and update standing only as labels in §3 allow
* Never promote “same model re-read itself” to Confirmed or Outcome-checked

=== Step 8 — Store and version on promotion ===
Promote a draft to the Working Standards library only after:

* User (or defined gate) could inspect it; and
* Version/date/standing recorded; and
* Up-link to top-level rules present

On failure in later use: recalibrate or retire—do not silently replace with a new improvisation.

----

== 3. Standing labels (v0.1.1 vocabulary) ==

{| class="wikitable"
! Label
! Meaning
|-
| '''Self-Assessed'''
| Author/instrument checked itself or its own pipeline only
|-
| '''Outcome-checked (tier A)'''
| A real test/measurement external to the model voted (works/fails, measurement, reproducible run). Allowed for '''solo builders''' when they run the test themselves. Prefer this label over bare “Confirmed” for solo outcome contact
|-
| '''Confirmed'''
| Stronger path: second party (independent human gate or defined independent process) stands behind the check — '''or''' project sampling policy explicitly treats documented tier-A outcome contact as sufficient '''and says so in writing'''. '''Not''' correlated self-agreement
|-
| '''Disputed / Open'''
| Known disagreement or unresolved nonconformance
|}

'''Clarification:''' v0.1.1 default for shared/public claims: use '''Outcome-checked (tier A)''' for solo test contact; reserve bare '''Confirmed''' for second-party or policy-defined stronger paths.

Tier qualifier recommended in prose: e.g. “Self-Assessed (tier C)”, “Outcome-checked (tier A, solo)”.

----

== 4. Stop rules (refuse theater) ==

Stop or escalate when:

* Pressure to call something Confirmed without an allowed path
* Tier C material is being closed as if tier A
* User standards would be overwritten quietly
* No working standard exists and stakes are high—draft and gate first
* Sampling policy cannot be met but Confirmed language is still requested → '''refuse the label'''

----

== 5. Roles (who does what) ==

{| class="wikitable"
! Role
! Responsibility
|-
| '''Human steward'''
| Top-level rules; high-stakes gates; sampling policy honesty
|-
| '''User'''
| Supplies file standards; accepts/rejects refinements; may reject recalibration suggestions
|-
| '''AI instrument'''
| Retrieve/apply; draft on need; propose refinements on request; cal notes; flag uncertainty/disagreement
|-
| '''Reality (tier A)'''
| Final vote where contact exists
|}

----

== 6. Minimal pilot recipe (recommended first use) ==

# Pick a '''tier-A''' procedural task (know-how that can be tried).
# Load or draft one working standard.
# Run the per-task loop once end-to-end.
# Perform one real outcome check if cheap (solo Outcome-checked is valid).
# Log one cal note with all minimal fields.
# Decide: keep, revise, or retire the working standard.

Do '''not''' start with contested culture-war topics as the first proof of the protocol.

----

== 7. Relationship to other pages ==

* '''Discipline:''' [[Metrology of the Abstract]]
* '''Doctrine (AI):''' [[Metrology of the Abstract: Application to Artificial Intelligence]]
* '''Operator feel:''' [[Operator Insight: Freedom under the Chain & Grounding under Ambiguity]]
* '''This page:''' procedures only

v0.1.1 is intentionally thin. Expand only from pilot nonconformances—not from brainstorming alone.

----

== Non-goals ==

* Not universal confirmation of every sentence
* Not automatic truth under ambiguity
* Not a requirement that every org adopt Sovereign Games branding
* Not a full economic solution to human gating (named constraint remains)

----

== Changelog ==

* '''v0.1.1 (2026-07-23):''' Solo-builder fallback (§1.4); Outcome-checked label; Confirmed narrowed; cites Metrology of the Abstract discipline page; Step 7 aligned.
* '''v0.1 (2026-07-22):''' Initial thin protocol.

----

== Current standing ==

{| class="wikitable"
! Field
! Value
|-
| Version
| 0.1.1
|-
| Maturity
| Experimental
|-
| Calibration review
| Self-Assessment
|-
| Validation
| Low — not yet pilot-proven
|-
| Role
| Builder protocol / working procedure
|}

This protocol is in [[Permanent Beta]]. Revise when pilots show missing steps, impossible gates, or theater loopholes.

----

'''See the Game. Refuse the Game. Build Better.'''  
''Reality gets the final vote.''

{{RelatedPages}}
{{Standalone Transparency}}
{{Calibration Dependent}}
{{Calibration Dependencies Display}}
{{Calibration Maintenance}}
{{Framework Reference}}
{{Calibration Procedure Reference}}
{{Tracking Logs Reference}}

[[Category:Meta & Framework]]
[[Category:Calibration Procedures]]

{{Resource
| Title = Builder Protocol v0.1.1: Metrology of the Abstract applied to AI
| URL = https://www.thesovereigngames.com/wiki/Builder_Protocol_v0.1:_Metrology_of_the_Abstract_applied_to_AI
| Description = Thin implementable protocol for AI claim pipelines under Metrology of the Abstract: retrieve working standards, tier claims, honest standing, cal notes, sampling, user standards on request only. v0.1.1 adds Outcome-checked and solo path.
| Category = Meta & Framework
}}

{{Admin Page Status
| categorization = Done
| calibration_review = Self-Assessment
| instrument_grade = Experimental
| validation = Low
| review_date = 2026-07-23
| reviewed_by =
| priority = Core
| review_threshold = 90
| has_backlinks = Not checked
| outbound_links_valid = Not checked
| in_outline = Yes
| in_category_outline = Yes
| templates_complete = Done
| formatting_standard = Meets standard
| symmetry_check = Applied
| self_report_flagged = No
| terminology_consistent = Yes
| standing_check = Self-assessed only
| drift_report_status = None open
| depends_on = Metrology of the Abstract; Metrology of the Abstract: Application to Artificial Intelligence; Operator Insight: Freedom under the Chain & Grounding under Ambiguity; Permanent Beta; Reference Standards
}}