Version Pinning & Model Lock — OpsDeploy

Lock models, prompts, embeddings, rerankers, tokenizers, and analyzers so prod can’t drift.
Use this page to make “what ran” exactly the same as “what you tested,” and to detect any silent shift before a rollout.

Open these first

Readiness gate: rollout_readiness_gate.md
Boot & deploy traps: bootstrap-ordering.md, deployment-deadlock.md, predeploy-collapse.md
Retrieval integrity: retrieval-traceability.md, data-contracts.md
Embedding vs meaning: embedding-vs-semantic.md
Query split & reranking: pattern_query_parsing_split.md, rerankers.md

What to pin (and where)

Layer	Pin fields	Why it matters
LLM model	`MODEL_VER`, provider model ID, decoding params	Minor provider updates shift ΔS/λ; force exact version and params.
Prompt pack	`PROMPT_VER`, header checksum	Header order changes λ; lock and audit.
Embedding	`EMBED_MODEL_VER`, `EMBED_DIM`, `NORM`, metric	Any mismatch breaks neighborhood semantics.
Reranker	`RERANK_CONF`, model/ver, cutoff	Top-k order flips alter ΔS and coverage.
Tokenizer/Analyzer	`TOK_VER`, `ANALYZER_CONF`	CJK/diacritics/full/half width change token boundaries.
Chunker	`CHUNK_SCHEMA_VER`, overlap, window	Alters anchor alignment and traceability.
Index	`INDEX_HASH`, build time, doc schema	Guarantees the retriever sees the tested graph.
Tools/Functions	`TOOL_SCHEMA_VER`	JSON looseness raises ΔS and loop risk.

Acceptance targets

Prod = Staging for: MODEL_VER, PROMPT_VER, EMBED_MODEL_VER, EMBED_DIM, NORM, metric, RERANK_CONF, TOK_VER, ANALYZER_CONF, CHUNK_SCHEMA_VER, INDEX_HASH.
ΔS(question, retrieved) ≤ 0.45 and coverage ≥ 0.70 on gold set after pinning.
λ convergent across 3 paraphrases and 2 seeds with the pinned headers.
First call cold-start passes in ≤ 5 minutes.

60-second lock checklist

Freeze versions in config
Single source of truth (env or KV) for all fields above; deny deploy if any are unset.
Emit version headers on every request
Log and return: all pins + BUILD_ID and GIT_SHA. Refuse if missing.
Index integrity
Retriever must assert INDEX_HASH matches; otherwise hard-fail with fix tip.
Semantic sanity
Run 20–40 gold items; verify ΔS/coverage/λ targets before canary.

CI/CD gate (paste-ready)

# opsdeploy/version_lock.yml
checks:
  require_fields:
    - MODEL_VER
    - PROMPT_VER
    - EMBED_MODEL_VER
    - EMBED_DIM
    - NORM
    - metric
    - RERANK_CONF
    - TOK_VER
    - ANALYZER_CONF
    - CHUNK_SCHEMA_VER
    - INDEX_HASH
  assert_equal_envs:
    - { field: MODEL_VER, envs: [staging, prod] }
    - { field: PROMPT_VER, envs: [staging, prod] }
    - { field: EMBED_MODEL_VER, envs: [staging, prod] }
    - { field: INDEX_HASH, envs: [staging, prod] }
  deny_if_missing: true
  deny_if_changed_without_alias_swap: [INDEX_HASH]
decision:
  on_fail: block_rollout
  on_pass: proceed
artifacts:
  - logs/version_manifest.json
  - logs/index_hash.txt

Runtime header contract (client/server)

Client must send X-WFGY: MODEL_VER,PROMPT_VER,EMBED_MODEL_VER,EMBED_DIM,NORM,metric,RERANK_CONF,TOK_VER,ANALYZER_CONF,CHUNK_SCHEMA_VER,INDEX_HASH,BUILD_ID,GIT_SHA

Server must validate

All fields present.
Exact match to server’s pins.
If mismatch: return 409 + a JSON diff and links to fix pages.

Typical failure → exact fix

High similarity but wrong meaning after “minor” update → embedding-vs-semantic.md
Top-k order unstable between runs → rerankers.md, pattern_query_parsing_split.md
CJK/diacritics behavior changes post-deploy → tokenizer/analyzer pins; verify with retrieval-traceability.md
First call crashes or uses old index → bootstrap-ordering.md, predeploy-collapse.md

Version manifest example

{
  "MODEL_VER": "gpt-4o-2025-07-15",
  "PROMPT_VER": "p-2025.08.31-h1",
  "EMBED_MODEL_VER": "text-embedding-3-large",
  "EMBED_DIM": 3072,
  "NORM": "l2",
  "metric": "cosine",
  "RERANK_CONF": "bge-rerank-v2@top32",
  "TOK_VER": "tiktoken-2025.07",
  "ANALYZER_CONF": "icu-cjk-folding-v3",
  "CHUNK_SCHEMA_VER": "s128/o32",
  "INDEX_HASH": "a1d9f0...e7",
  "BUILD_ID": "2025.08.31.01",
  "GIT_SHA": "abcdef1"
}

Rollback note

If any pin drifts in prod, treat as outage. Block writes, flip to Blue/previous alias, then reopen with: blue_green_switchovers.md

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

🧭 Explore More

Module	Description	Link
WFGY Core	WFGY 2.0 engine is live: full symbolic reasoning architecture and math stack	View →
Problem Map 1.0	Initial 16-mode diagnostic and symbolic fix framework	View →
Problem Map 2.0	RAG-focused failure tree, modular fixes, and pipelines	View →
Semantic Clinic Index	Expanded failure catalog: prompt injection, memory bugs, logic drift	View →
Semantic Blueprint	Layer-based symbolic reasoning & semantic modulations	View →
Benchmark vs GPT-5	Stress test GPT-5 with full WFGY reasoning suite	View →
🧙‍♂️ Starter Village 🏡	New here? Lost in symbols? Click here and let the wizard guide you through	Start →

👑 Early Stargazers: See the Hall of Fame — Engineers, hackers, and open source builders who supported WFGY from day one.

⭐ WFGY Engine 2.0 is already unlocked. ⭐ Star the repo to help others discover it and unlock more on the Unlock Board.

10 KiB Raw Blame History Unescape Escape