v1.9.6 prep: preserve identity-critical facts; ambiguity-aware review and trial

[RedEyeNinja-BKK]

v1.9.6 prep: preserve identity-critical facts; ambiguity-aware review and trial

Branch: v1.9.6-identity-ambiguity
Base SHA (verified): 75654fe6f47c3dddb5b690abdf2390a0152f9df2
Head SHA: ea3b4b8d1d1d1c73b902fcddcb7f087c25dacd27
References: Issue #1 — senior-review steering 2026-08-08 (identity-critical fact loss + ambiguity collapse)
Status: Senior review PASS (review 4888836358 on head ea3b4b8…); awaiting operator merge gate (Vincent). Not merged, not tagged, not released, not deployed.

This PR is for review of the committed diff and the operator merge decision. Do not merge without operator approval.

---

Problem statement

The Proxmox Manager downstream trial (2026-08-08) surfaced a general engine-level failure: identity-critical fact loss + ambiguity collapse. A generated package contained multiple identifiers for one environment; the relationship between them was not preserved strongly enough, and the generated agent inferred equivalence instead of proving which identifier represented which entity/role. (Observed symptom: a node's management address was conflated with a separate out-of-band interface address; a stale source conflicted with the operator-supplied fact.)

The reusable lesson is domain-neutral: generated packages must preserve operator-supplied identity-critical relationships (entity ↔ role ↔ identifiers/endpoints/aliases), keep conflicting sources visibly conflicting until disambiguated, and never silently infer equivalence between distinct names/IDs/aliases/records/accounts/machines.

This patch deliberately does not encode "put IP/interface maps in personas" — that would overfit the engine to the observed symptom. It changes behavior at the pipeline contracts that own this: intake (preserve), pattern-author (place), review (challenge), trial (prove), plus one supporting reference (identity placement rule in skill anatomy).

Acceptance criteria (defined before editing)

1. Domain-neutral: no Proxmox/environment-specific wording in changed engine prompts.
2. Intake preserves identity-critical relationships where they affect later decisions/tool targets; distinct entities/identifiers stay distinct; source-attributed conflicts stay unresolved (no silent "best" synthesis); ambiguity preserved.
3. Pattern-author places mutable operational/contextual identity facts in references by default; stable behavior-defining identity may stay in the persona (no persona bloat with mutable facts).
4. Supplied/live source conflicts on identity-critical facts remain visible, sources named, disambiguating evidence required — never silent selection.
5. Review adversarial pass asks the wrong-entity/ambiguity question; ambiguity that can change the target is a REVISE finding.
6. Trial includes at least one contradiction/alias/near-match case when identity is material to behavior.
7. Prompts-only; no new stages, evaluators, state machinery, persistence, manifests, or governance infrastructure; no duplicated Turnstone-native enforcement prose.
8. No conversational regressions: no extra questioning, no repeated material requests, no premature generation, no scope guessing, no persona bloat, no ceremonial gates, no governance boilerplate in normal conversation, no excessive evidence ceremony, engine does not do the generated package's domain job.
9. Structural validation passes; behavioral trials pass on (a) the Proxmox/identity regression and (b) a materially different non-infrastructure package.

Exact affected pipeline contracts

Contract	Change
references/intake.md (step 4 Extract)	New preservation requirement: identity-critical entity ↔ role ↔ identifier/alias relationships preserved when material; distinct entities/identifiers kept distinct; source-attributed conflicts kept unresolved (no silent "best" synthesis); ambiguity preserved. Carried forward in the collected-material summary to Pattern.
process-engine-pattern-author (SKILL.md, Core Process step 4)	"Preserve identity-critical facts": explicit mapping, placement rule (mutable operational/contextual identity facts → references; stable identity → persona), conflict preservation, no silent equivalence inference. Red Flags + Verification extended.
process-engine-review (SKILL.md, Core Process step 6)	Adversarial pass gains the wrong-entity/ambiguity question; REVISE consequence. Red Flags + Verification extended.
process-engine-trial (SKILL.md, Core Process step 1)	Case set gains the identity case (contradiction/alias/near-match) when identity is material. Red Flags + Verification extended.
references/skill-anatomy.md (engine addition)	Identity placement rule: stable identity vs mutable operational/contextual facts.
Embedded core reference copies	skills/process-engine-core/references/intake.md + skill-anatomy.md synced to root (required by structural gate).

Concise full-diff summary

7 files changed, +88/−3 (two commits):

• references/intake.md +14/−1
• skills/process-engine-core/references/intake.md +14/−1
• skills/process-engine-pattern-author/SKILL.md +26
• skills/process-engine-review/SKILL.md +11/−1
• skills/process-engine-trial/SKILL.md +9
• references/skill-anatomy.md +7
• skills/process-engine-core/references/skill-anatomy.md +7

Commits: fa8281a (initial patch) + ea3b4b8 (revision 2: intake preservation moved upstream + placement wording generalized per senior review).

No changes to persona, core skill body, ship/triage contracts, templates, manifest, version numbers, or tools. No new machinery.

Structural validation evidence

python3 tools/validate.py --no-diff → RELEASE GATE PASS — release-consistent on head ea3b4b8….
(Counts match 6 skills / 7 refs / 6 templates; frontmatter valid; embedded core references == root references; links resolve; no stale version references; lineage v8.0.) GitHub release-gate run on the exact SHA is green.

Domain-neutrality scan on changed artifacts: no environment/Proxmox-specific terms; placement wording generalized to "mutable operational/contextual identity facts."

Behavioral trial evidence (repo-first — revised text embedded as governing procedure, store untouched)

Method: task-agent runs of the revised engine text through the full pipeline (Collect → intake → Clarify → Objective → Summary Gate → Pattern → Review → Trial) from raw operator-supplied material (NOT pre-normalized). Receipts are bounded and inspectable.

1. Proxmox/identity regression — PASS

Raw inputs → intake-preserved (brick-a management vs OOB kept as distinct typed identifiers; brick-a vs brick-b distinct standalone entities; old-spreadsheet conflict kept source-attributed and unresolved — not synthesized; passwords excluded) → pattern placement (persona = stable identity only: "You manage two independent backup appliances… Never treat them as a cluster or interchangeable hosts."; addresses/roles/conflict in reference; conflict language requires disambiguating evidence; "Do not infer that identifiers, endpoints, aliases, or records refer to the same target unless the operator asserts the equivalence or disambiguating evidence proves it.") → review wrong-entity question PASS (REVISE condition named) → trial contradiction/alias/near-match cases (OOB near-match, source contradiction, alias collision, independent-host distinction). Regression checklist all PASS.

2. Non-infrastructure generalization (customer records) — PASS

Raw inputs → intake-preserved (A-1001 vs A-1002 distinct entities with distinct emails, never merge; T1 vs T1-2 distinct SKUs; name-only ambiguity preserved; contact data mutable) → pattern placement (persona = stable behavior: "Protect customer identity during lookup and update. Never merge distinct customer records merely because their names match."; identity map, ambiguity rule, mutable-contact rule in reference; "A name-only lookup for 'Alex Chen' is ambiguous… Do not guess, update, merge, or gate on one record.") → review wrong-entity question PASS (REVISE condition named) → trial contradiction/alias/near-match cases (name-only, email near-match, contradictory identifiers, SKU near-match, contact update). Regression checklist all PASS.

Repo ↔ live-store drift findings (read-only, 2026-08-08)

Using vs developing distinction: the live store's process-engine persona object (display name "Process Engine Development Team Lead") is the development persona in use for development sessions — it is not the generator persona. The repo persona.md is the canonical generator persona under development. They are different artifacts/roles, not version-diffs of one object.

Artifact	Repo state (developing)	Deployed state (using)	Meaningful difference	Proposed disposition
Generator persona (persona.md)	v1.9.5 canonical, 4,558 B	—	Not deployed under its own artifact; deployed process-engine persona is the dev-lead persona	Keep repo canonical; reconcile at controlled deployment (not this PR)
Dev-lead persona (store process-engine)	(not in repo)	v1.8.2-era base prompt, 10,651 B, PATCHed 2026-08-05	Store-side dev identity; no repo artifact	Document separately; out of scope here
6 pipeline skills (core, pattern-author, review, trial, ship, triage)	v1.9.5 canonical	v1.8.2-era content — all 6 DIFF	Stale store vs canonical repo	Repo-first: deploy store refresh at approved release, not this PR
7 references on core	v1.9.5 (best-practices/governance/standards differ)	evidence-library/intake/safety/skill-anatomy MATCH; best-practices/governance/standards DIFF	v1.9.5 clean-slate changes not yet in store	Store refresh at approved release
6 starter templates + orientation	v1.9.5	all 7 MATCH	None	No action
Prompt policy process-engine-context + judge rule	documented in docs/governance-usage.md	deployed, live	None observed	No action

No live store mutation performed. GitHub remains canonical; deployment follows approved repo state.

Known limitations / remaining uncertainty

• Behavioral trials ran the revised procedure text as governing instructions in a task-agent harness (repo-first, store untouched). The deployed store still runs the v1.8.2-era engine; store-level behavior will only reflect this patch after a controlled deployment at the operator gate.
• The patch does not bump versions (still v1.9.5 metadata); version/release actions are deliberately deferred per steering (no tag/release in this PR).
• Persona-role naming ("Process Engine Lead" vs the generator persona) is a documentation-level question noted in drift findings; not changed in this PR.
• The non-infrastructure trial was a single scenario; a cross-domain matrix is the proposed next product trial phase (see Issue Review & steering - Proxmox Manager trial evidence (2026-08-08) #1 steering), not part of this patch.

[RedEyeNinja-BKK]

Senior review verdict — CHANGES REQUIRED before approval

I reviewed the actual PR #2 patch at head fa8281a9358f05c506bcd9776ea6ab4b14477099 against base 75654fe6f47c3dddb5b690abdf2390a0152f9df2, the current v1.9.5 pipeline contracts, and Issue #1 steering. CI/release-gate is green and the patch is directionally correct: it extracts the T3 lesson as identity/ambiguity behavior rather than a Proxmox-specific IP-map rule, adds no runtime machinery, and puts review/trial hooks at sensible downstream gates.

However, I am not approving this head. One pipeline-contract gap is load-bearing, and the behavioral evidence needs one tighter proof.

BLOCKER 1 — the preservation rule begins too late in the pipeline

The patch adds identity preservation in process-engine-pattern-author, but Process Engine deliberately runs:

Collect → intake/extract → Clarify → Objective → Summary Gate → Pattern

references/intake.md currently tells the engine to extract techniques/domain specifics and says multiple sources combine into a “best-of-all-worlds understanding.” It does not require identity-critical mappings, typed distinctions, or conflicting source values to survive extraction. That means Collect can still flatten or synthesize away the exact relationship/conflict before Pattern ever sees it.

Pattern-author cannot preserve evidence it was not handed.

Required correction

Keep this narrow. Amend the intake contract so that during extraction:

• identity-critical entity ↔ role ↔ identifier/alias relationships are preserved when material to later decisions/targets;
• explicitly distinct identifiers/entities are not collapsed;
• conflicting identity-critical source values remain source-attributed and unresolved rather than being synthesized into one “best” value;
• that preserved relationship/conflict is carried into the collected-material summary handed downstream.

Sync the embedded core copy of intake.md as required by the existing structural gate. I do not see a need to add a new stage or expand core conversational questioning.

Then re-run the regression through the full intake path, not merely with already-normalized material supplied to Pattern. The proof we need is that the identity mapping/conflict survives Collect/intake and reaches Pattern intact.

BLOCKER 2 — domain-neutrality claim is overstated

The PR acceptance criteria/body say the changed prompts contain no “environment-specific wording” and that the domain-neutrality scan found none, but the committed diff explicitly introduces “mutable environment/topology facts” in both pattern-author and skill-anatomy.

That wording is also narrower than the non-infrastructure behavior being claimed: mutable customer contact identifiers, account associations, record IDs, product aliases, etc. are not naturally “environment/topology.”

Please generalize the placement rule to the actual abstraction — mutable operational/contextual identity facts or equivalent domain-neutral language — with examples spanning infrastructure and non-infrastructure only where examples add value. Avoid turning this into a taxonomy paragraph.

EVIDENCE REQUIRED — behavioral PASS is still narrative-only on the PR

The PR body gives good summaries, but the two behavioral PASS claims are not independently inspectable from the PR. Do not add evaluator infrastructure or a large evidence bundle. A bounded, sanitized receipt is enough.

For each of the two trials, add concise evidence showing:

input fact/conflict → intake-preserved representation → generated package placement/behavior → review/trial observed result

For the Proxmox regression, specifically show that management vs OOB and the conflicting source survive intake without collapse. For the customer-records case, show the same for the two same-name customers / conflicting identifiers. A short PR comment/body table with bounded excerpts is sufficient.

What does NOT need changing

• No new evaluator, state machine, persistence, manifest, governance layer, or release machinery.
• No live Turnstone mutation.
• No version bump yet.
• No expansion into v2.0.0.
• No new conversational gate/question merely for identity handling.
• The review wrong-entity check and trial contradiction/alias/near-match concept are sound and should remain, subject to any simplification that falls out of the intake fix.

Re-review gate

Please update this same branch, preserving the narrow scope, then report:

1. new head SHA;
2. exact intake + embedded-intake diff;
3. generalized placement wording change;
4. release-gate result on the new head;
5. bounded full-pipeline receipts for both regression cases;
6. confirmation that no live store mutation occurred.

Do not merge/release/deploy. Return the committed PR diff for re-review.

[RedEyeNinja-BKK]

Revision 2 — address senior review (id 4888827387)

New head: ea3b4b8d1d1d1c73b902fcddcb7f087c25dacd27 (committed on the existing v1.9.6-identity-ambiguity branch; PR updated, not reopened)
Base: 75654fe6f47c3dddb5b690abdf2390a0152f9df2 (unchanged)
Commits on branch: fa8281a (initial) + ea3b4b8 (this revision)

Changes in this revision (diff vs prior head fa8281a)

5 files changed, +39/−9 — exactly the three corrections; nothing else touched.

1. Intake preservation moved upstream — references/intake.md (+15/−1) Extract step now requires: preserve identity-critical entity ↔ role ↔ identifier/alias relationships when material; keep explicitly distinct entities/identifiers distinct; keep source-attributed conflicts unresolved (no silent "best" synthesis); preserve ambiguity. Carried forward in the collected-material summary to Pattern. Embedded core copy synced (skills/process-engine-core/references/intake.md, identical per gate).
2. Generalized placement wording — pattern-author + references/skill-anatomy.md (+embedded copy): "mutable environment/topology facts" → "mutable operational/contextual identity facts" with non-infrastructure examples (addresses, endpoints, interfaces, contact identifiers, account associations, record IDs). No taxonomy; no bloat.
3. Behavior re-proven through the full pipeline — receipts below.

Structural validation (exact new head)

tools/validate.py --no-diff → RELEASE GATE PASS — release-consistent. (6/7/6 counts; frontmatter valid; embedded core references == root references; links resolve; no stale version references; lineage v8.0.)

Full-pipeline behavioral receipts (bounded, sanitized)

Method: task-agent run of the revised engine text, full Collect → intake → Clarify → Objective → Summary Gate → Pattern → Review → Trial from raw operator-supplied material (NOT pre-normalized). No store mutation; no transcript tax; receipts are the inspectable evidence.

A. Proxmox/identity regression — PASS

Input fact/conflict → intake-preserved representation:

Input (raw, from operator)	Intake-preserved (handed to Pattern)
brick-a management 10.10.20.5, OOB 10.10.99.5 (OOB NOT management)	Distinct typed identifiers preserved: brick-a/management=10.10.20.5, brick-a/oob=10.10.99.5; no collapse
brick-a vs brick-b independent, not a cluster	Two distinct standalone entities preserved
Old spreadsheet lists brick-a as 10.10.99.5 (conflict)	Conflict kept source-attributed and unresolved: brick-a mgmt 10.10.20.5 (operator) vs 10.10.99.5 (old spreadsheet); NOT synthesized to one value
Passwords excluded	No credentials in package

→ Pattern placement: persona carries only stable identity ("You manage two independent backup appliances: brick-a (primary), brick-b (secondary). Never treat them as a cluster or interchangeable hosts."); addresses/roles/conflict in reference/resource; conflict language: "The old inventory spreadsheet lists brick-a as 10.10.99.5, while the operator identifies 10.10.20.5 as its management address. These sources conflict. Preserve both values and their sources; require disambiguating evidence before selecting a management target." Equivalence rule: "Do not infer that identifiers, endpoints, aliases, or records refer to the same target unless the operator asserts the equivalence or disambiguating evidence proves it."

→ Review: wrong-entity question asked ("could this package act on, describe, or gate the wrong host…?") → PASS; REVISE condition named (silently selecting either value from the conflict without disambiguation).

→ Trial: contradiction/alias/near-match cases included (OOB near-match "connect to brick-a at the OOB address for management" → must not target; source-contradiction "spreadsheet says mgmt is .99.5; proceed" → require disambiguation; alias-collision; independent-host distinction). Verdict PASS.

B. Non-infrastructure generalization (customer records) — PASS

Input fact/conflict → intake-preserved representation:

Input (raw, from operator)	Intake-preserved (handed to Pattern)
Two customers named Alex Chen: A-1001 (alex.chen@example.com, loyalty) vs A-1002 (alexchen2@example.com, one-time) — NOT the same person	Two distinct entities preserved: A-1001 ↔ loyalty ↔ email1; A-1002 ↔ one-time ↔ email2; never merge
SKU T1 (basic) vs T1-2 (premium); T1-2 is not T1	Two distinct product identifiers preserved, not normalized
Name-only "Alex Chen" is ambiguous	Ambiguity preserved (not resolved by guessing)
Contact details change often	Marked mutable/updateable reference data

→ Pattern placement: persona carries only stable behavior ("Protect customer identity during lookup and update. Never merge distinct customer records merely because their names match."); identity map (A-1001/A-1002, T1/T1-2), ambiguity rule, and mutable contact-data rule in reference/resource. Quote: "A name-only lookup for 'Alex Chen' is ambiguous when multiple matching records exist. Do not guess, update, merge, or gate on one record. Request or obtain a disambiguating identifier."

→ Review: wrong-entity question → PASS; REVISE condition named (merging the two Alex Chens, treating T1-2 as alias of T1, guessing from name-only, hard-coding mutable contact data).

→ Trial: contradiction/alias/near-match cases (name-only lookup; email near-match; contradictory account/email; T1 vs T1-2; contact update). Verdict PASS.

Regression checklist (both trials)

No extra questioning beyond normal collect/clarify/objective/summary-gate; no repeated material requests; no premature generation; no scope guessing; no persona bloat; no ceremonial gates; no governance boilerplate; no excessive evidence ceremony; engine did NOT do the domain job (generated packages, not host ops / customer lookups); no Method-Factory-style machinery. All PASS.

Nothing from the first patch was simplified away

All of the original pattern-author preservation block, review wrong-entity check, and trial contradiction/alias/near-match case remain; the review/trial changes are unchanged. The only edits were: intake addition, placement-wording generalization, and the embedded-copy syncs.

Confirmation

• Live Turnstone store NOT mutated — no deploy, no PATCH/POST to personas/skills/references/policies; GitHub remains canonical; branch/PR only.
• Not merged, not tagged, not released.

Returning for re-review of the committed PR head.

[RedEyeNinja-BKK]

Senior re-review verdict — PASS on committed product change; metadata cleanup before operator gate

I re-reviewed the actual PR #2 head ea3b4b8d1d1d1c73b902fcddcb7f087c25dacd27 against base 75654fe6f47c3dddb5b690abdf2390a0152f9df2, including the complete 7-file patch, revision evidence comment, and release-gate run on the exact head.

Verdict

PASS — the substantive blockers from review 4888827387 are closed.

I find no remaining product/behavior blocker in the committed diff. This is an appropriately narrow prompts-only Process Engine improvement and does not drift toward Method Factory or duplicate Turnstone-native machinery.

This is not merge/release/deployment authorization. Vincent remains the operator gate.

What is now correct

1. Preservation begins at the right pipeline stage

The prior blocker was that Pattern could only preserve identity-critical evidence if intake had not already flattened it. Revision 2 fixes that at references/intake.md and its embedded core copy:

• identity-critical entity ↔ role ↔ identifier/alias relationships are preserved when material;
• explicitly distinct entities/identifiers stay distinct;
• conflicting identity-critical values remain source-attributed and unresolved;
• genuine ambiguity remains ambiguity;
• the preserved relationship/conflict is explicitly carried into the collected-material summary handed to Pattern.

That closes the upstream information-loss hole without adding a question, stage, state mechanism, or ceremony.

2. The abstraction is now genuinely domain-neutral

The placement rule now uses mutable operational/contextual identity facts, with infrastructure and non-infrastructure examples. That correctly captures the underlying behavior without teaching Process Engine a Proxmox-specific topology rule.

The persona/reference distinction remains sensible: stable behavior-defining identity may live in persona; mutable identity/association facts default to references/resources.

3. Downstream contracts line up coherently

The complete flow now reads coherently:

intake preserves → pattern places without silent equivalence → review challenges wrong-entity ambiguity → trial exercises contradiction/alias/near-match behavior

The review and trial additions remain proportional and conditional on identity being material. They do not introduce a universal extra gate or interrogation step.

4. Behavioral evidence is sufficient for this patch

The bounded receipts in revision comment 5226178375 now exercise the requested full path from raw material through intake, Pattern, Review, and Trial.

The Proxmox case proves management/OOB distinction and source conflict survive intake. The customer-record case proves the same abstraction on distinct same-name people, record identifiers, product near-matches, ambiguity, and mutable contact data.

This is enough evidence for the bounded correction. I do not recommend adding evaluator infrastructure or more evidence machinery to this PR.

5. Structural state is good

The exact reviewed head is ea3b4b8d1d1d1c73b902fcddcb7f087c25dacd27, 7 changed files, +88/−3, and GitHub's release-gate workflow completed successfully on that SHA.

One non-code cleanup before the operator merge decision

The PR body is stale relative to the reviewed head. It still says:

• head fa8281a;
• 5 changed files / +56−1;
• affected contracts omit intake;
• acceptance/domain-neutrality wording reflects the pre-revision patch.

The revision comment contains the correct current evidence, so this does not invalidate the committed change. But the PR description should be updated before merge so the durable GitHub record accurately describes what is being approved.

Please update the PR body only to reflect current head ea3b4b8…, 7 files / +88−3, intake as an affected contract, the generalized placement wording, and the revision-2 evidence. Do not alter code merely for this cleanup.

A metadata-only body correction does not need another full senior diff review unless the head SHA changes.

Recommended next gate

1. Refresh the PR body to match the current committed/evidenced state.
2. Confirm head remains exactly ea3b4b8d1d1d1c73b902fcddcb7f087c25dacd27 and release-gate remains green.
3. Present the PR to Vincent for the explicit merge decision.
4. If Vincent approves merge, merge only the reviewed head; release and controlled Turnstone deployment remain separate operator gates.

Senior review status: product change PASS; ready for operator merge gate after PR-description hygiene.