|
| 1 | +# Five-Layer Agent SOP Reference Record — 2026-07-04 |
| 2 | + |
| 3 | +## Purpose |
| 4 | + |
| 5 | +Preserve the useful concepts and review decisions from the temporary root-level `five_layer_agent*` delivery artifacts before those artifacts are removed. This record is a reference and memory artifact, not an operating rule, not a new workflow skill, and not a runtime commitment. |
| 6 | + |
| 7 | +## Source Artifacts |
| 8 | + |
| 9 | +The reviewed source set was temporary and expected to be removed from the repository root: |
| 10 | + |
| 11 | +- `five_layer_agent_sop_v6_rc_mother.md` |
| 12 | +- `five_layer_agent_sop_blog_public.md` |
| 13 | +- `five_layer_agent_sop_memo_full.md` |
| 14 | +- `five_layer_agent_sop_manifest.json` |
| 15 | + |
| 16 | +A bundle artifact, `five_layer_agent_sop_v6_rc_bundle.zip`, was created during review because the source delivery checklist required it. This reference record does not depend on that bundle remaining present. |
| 17 | + |
| 18 | +## Review 1: Delivery Artifact Check |
| 19 | + |
| 20 | +### Finding |
| 21 | + |
| 22 | +The mother document and memo both declared the delivery should include a bundle zip, but the bundle was absent at first review. |
| 23 | + |
| 24 | +### Action Taken |
| 25 | + |
| 26 | +Created `five_layer_agent_sop_v6_rc_bundle.zip` containing: |
| 27 | + |
| 28 | +- `five_layer_agent_sop_v6_rc_mother.md` |
| 29 | +- `five_layer_agent_sop_blog_public.md` |
| 30 | +- `five_layer_agent_sop_memo_full.md` |
| 31 | +- `five_layer_agent_sop_manifest.json` |
| 32 | + |
| 33 | +### Verification Performed |
| 34 | + |
| 35 | +- Manifest-listed document hashes matched the manifest for bytes, line counts, md5, and sha256. |
| 36 | +- Blog and memo were regenerated from the mother document's `SLICE:BLOG+MEMO` / `SLICE:MEMO` sections and matched the checked-in files. |
| 37 | +- Blog forbidden terms were absent: `產品假說 H1`, `H1`, `Heddle`, `Mini Ledger`, `Final gate`. |
| 38 | +- Archive integrity passed with `unzip -t`. |
| 39 | +- Bundle SHA-256 observed after creation: `3368b7c500af856635e24cf80255b0018fd850a8bd128c20dcba62804fc9dced`. |
| 40 | + |
| 41 | +### Residual Note |
| 42 | + |
| 43 | +The manifest anchors the three generated documents, not the zip itself. The manifest was left unchanged because embedding the bundle hash into a manifest that is bundled into the zip creates a self-reference problem. |
| 44 | + |
| 45 | +## Review 2: Prompt / Skill Promotion Decision |
| 46 | + |
| 47 | +### Direct Decision |
| 48 | + |
| 49 | +No TeaPrompt prompt or skill should be changed from this artifact alone. |
| 50 | + |
| 51 | +Reasons: |
| 52 | + |
| 53 | +- The reviewed material is a single reference artifact, not local recurrence evidence. |
| 54 | +- TeaPrompt already has nine frozen workflow skills; adding a tenth skill requires recurrence-gated evidence and explicit human approval. |
| 55 | +- Many concepts are already covered by existing TeaPrompt prompts and skills. |
| 56 | +- Several strong mechanisms are runtime/operationalization-layer ideas, while TeaPrompt intentionally remains a prompt and methodology library rather than an agent runtime. |
| 57 | + |
| 58 | +## Worthy Concepts Preserved |
| 59 | + |
| 60 | +| Concept | Transferable Value | TeaPrompt Coverage / Decision | |
| 61 | +| --- | --- | --- | |
| 62 | +| Five-layer control plane: α flow gates, β formal SOP, γ durable execution, δ adversarial evaluation, ε governance/standards, ζ memory/identity integrity | Useful taxonomy for thinking about agent risk controls by layer. | Concept only. Existing `04-agent/runtime-trust-boundary.md`, `04-agent/agent-selection.md`, and `reflective-risk` cover the nearby methodology surfaces. Do not create a new taxonomy prompt yet. | |
| 63 | +| Risk routing with an irreversible-action tier (`L3-Irr`) | Separates ordinary high-risk work from irreversible external effects. | Already mostly covered by `reflective-risk` and runtime trust-boundary Human Review gates. Do not add another strictness ladder. | |
| 64 | +| Effect outbox + commit gate + idempotency key | Strong runtime pattern for irreversible side effects: stage effects, validate, require approval, then commit exactly once. | Runtime candidate only. TeaPrompt can cite the pattern in future runtime design reviews, but should not implement or promise it as prompt text. | |
| 65 | +| Third-party skill supply-chain gates: provenance, scan, sandbox trial, permission minimization, signing | Good operational checklist for onboarding untrusted skills or agent scaffolds. | Partly covered by scaffold provenance and trust-boundary prompts. Keep as concept only unless local skill-adoption work recurs. | |
| 66 | +| Memory write whitelist and identity-file write protection | Frames long-term memory and identity files as policy surfaces, not ordinary notes. | Runtime / host-enforcement concept. Useful for future trust-boundary reviews; no prompt or skill change from one artifact. | |
| 67 | +| High-volatility fact discipline | Dates, standards status, versions, prices, and schedules need check-date, recency caveat, and a tracking point. | Candidate in-place refinement for `reflective-research` if it recurs. Current `research.md` already says to mark items needing fresh verification, but this concept names the fact class more sharply. | |
| 68 | +| Number semantic discipline | Keep terms like flawed, critical, and malicious distinct; trace numbers to body tables or primary data, not titles. | Already covered by evidence mapping, claims ledgers, and critical-thinking checks. Preserve as reference example. | |
| 69 | +| Maturity three-way split: landed / in-progress / research-stage | Prevents proposal-stage standards or prototypes from being described as mature deployments. | Mostly covered by evidence-strength categories. Useful research phrasing, not a new artifact. | |
| 70 | +| Four-dimensional evidence ledger: existence, number-text, attribution/process, extrapolation | More precise than a binary source-supported flag; separates source existence from data generation and extrapolation validity. | Candidate in-place refinement for `reflective-review` or `reflective-research` if repeated. No promotion yet. | |
| 71 | +| Domain-specific fallacy list | Captures agent-risk research mistakes: title anchoring, maturity inflation, scanner-solutionism, vendor-narrative overtrust, false convergence, and infrastructure-as-correctness. | Useful reference examples. Existing fallacy scans are general enough; do not add the full list to core prompts. | |
| 72 | +| Hypothesis/evidence separation with falsifiers and lock rules | Keeps product hypotheses out of public narrative until validated. | Already covered by `01-thinking/falsifiability.md` and `reflective-research`. | |
| 73 | +| Empty-source rule | Missing, unreadable, hash-mismatched, or marker-missing sources are unknown; never fill gaps from context. | Already covered by runtime trust-boundary missing-data discipline and research unknown handling. | |
| 74 | +| Delivery anchoring by file presence, line counts, hashes, and sentinel searches | Good anti-fabrication pattern for artifact delivery claims. | Useful for delivery reviews. Not a general TeaPrompt prompt change. | |
| 75 | +| Slice-based mother document strategy | Single source of truth with public and memo slices avoids drift between public and internal versions. | Useful documentation pattern. Not needed as a TeaPrompt core rule without recurrence. | |
| 76 | +| `evidence-calibrated-revision` v1.2 procedure | Composite procedure for revising research docs under review: source existence, hard-error review, number tracing, maturity split, credibility ledger, volatility scan, hypothesis separation, delivery verification. | Do not create a tenth workflow skill. Treat as a composite of `reflective-research` + `reflective-review` unless repeated local demand justifies a prompt lens. | |
| 77 | + |
| 78 | +## Concept Candidates To Watch |
| 79 | + |
| 80 | +These are the only concepts that may deserve future in-place prompt refinement if they recur in local work: |
| 81 | + |
| 82 | +1. **High-volatility fact discipline** — candidate for `05-domain/research.md` or `skills/reflective-research/SKILL.md` wording if multiple research tasks mishandle volatile dates, standards status, version numbers, pricing, or schedules. |
| 83 | +2. **Four-dimensional evidence ledger** — candidate for `reflective-review` / `reflective-research` if simple verified/unverified ledgers repeatedly hide whether the source exists, the number text is supported, the data-generation process is trusted, or the conclusion is over-extrapolated. |
| 84 | + |
| 85 | +Both remain recurrence-gated. They are not new skills. |
| 86 | + |
| 87 | +## Concepts Not Worth Promoting Now |
| 88 | + |
| 89 | +- A new five-layer workflow skill. |
| 90 | +- A new strictness ladder competing with TeaPrompt's L1-L6 strictness system. |
| 91 | +- Runtime features such as effect outbox, commit gates, idempotency enforcement, memory write ACLs, sandbox execution, signing, or replay logs. |
| 92 | +- The full domain-specific fallacy table as a core prompt requirement. |
| 93 | +- Delivery-specific slice markers or bundle manifests as general library conventions. |
| 94 | + |
| 95 | +## Decision Ledger |
| 96 | + |
| 97 | +| Claim | Evidence | Unknowns | Counterargument | Decision | Falsifier / Review Trigger | |
| 98 | +| --- | --- | --- | --- | --- | --- | |
| 99 | +| The temporary files contain useful concepts worth preserving. | Direct review of the mother document, memo, blog, manifest, and delivery checks. | The external research claims inside the memo were not independently reverified in this task. | Temporary docs can become clutter if copied wholesale. | Preserve distilled concepts only; do not preserve raw source files as dependencies. | If future users need raw provenance, add a source-backed research record instead of restoring root temp files. | |
| 100 | +| No prompt or skill should be changed now. | Existing TeaPrompt surfaces cover most methodology concepts; project knowledge freezes nine workflow skills and recurrence-gates promotions. | Future recurrence may reveal a sharper local gap. | The high-volatility and 4D-ledger concepts are useful. | Record them as candidates, not edits. | Three independent local recurrences or explicit project decision to update research/review prompts. | |
| 101 | +| Runtime mechanisms should not become TeaPrompt commitments. | TeaPrompt's current direction excludes operating its own agent runtime. | A future project direction could add runtime work. | Runtime controls are the most concrete parts of the five-layer memo. | Keep runtime patterns as reference examples only. | Project direction changes to include runtime enforcement. | |
| 102 | + |
| 103 | +## No-Change Record |
| 104 | + |
| 105 | +The reviewed files should not cause immediate edits to: |
| 106 | + |
| 107 | +- `skills/reflective-research/SKILL.md` |
| 108 | +- `skills/reflective-review/SKILL.md` |
| 109 | +- `skills/reflective-risk/SKILL.md` |
| 110 | +- `04-agent/runtime-trust-boundary.md` |
| 111 | +- `04-agent/agent-scaffold-provenance.md` |
| 112 | +- `05-domain/research.md` |
| 113 | + |
| 114 | +If future work revisits this record, start from the two concept candidates above rather than re-importing the removed root files. |
| 115 | + |
| 116 | +## Human Review |
| 117 | + |
| 118 | +Not required for this reference record. Human approval is required before any future action that creates a new core workflow skill, changes runtime/security gates, or promotes executable knowledge. |
| 119 | + |
| 120 | +## Handoff |
| 121 | + |
| 122 | +After this record exists and the Decision Index points to it, the root-level `five_layer_agent*` files can be removed without losing the TeaPrompt-relevant concepts. Removal itself was not performed in this task. |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?v=4&size=40){kind=avatar}
0 commit comments