Epistemic Skills 7.3.0

Epistemic Skills 7.3.0

7.3 adds bounded Consensus research-agent Threads to resolve → literature,
with a recon handoff for scholarly premises. One complete research brief goes
to the provider; its synthesis enters the existing primary-content, reception,
counterevidence and uncertainty checks. Provider Deep requires an explicit
user request independently of the literature evidence-depth dial.

Scope and compatibility

• The current dated contract covers tool selection, capacity and overages,
  full-brief dispatch, no self-chaining, polling, exact document/citation
  preservation, pending/failed states and missing-capability fallbacks.
• The older ZCode profile is retained as historical host evidence, with an
  explicit pointer to the newer schema inspection. No runtime probe was run.
• Recon's candidate harvest and no-execution boundaries, evidence classes,
  reception and original-source checks remain binding. Threads do not prove
  truth, runtime behavior, a systematic review, or real-time monitoring.
• Seventeen skills remain. Descriptions, schemas, hooks and dependencies are
  unchanged. Manifest versions identify this source candidate as 7.3.0.
  Install targets and current handbook/publisher source align to 7.3
  prospectively. Offline checks validate target paths in the candidate without
  assuming an unpublished tag exists; actual remote reachability and live wiki
  publication remain separately verified gates. Historical snapshots retain
  their original versions. See the candidate checklist.

Evidence and release state

Source base: upstream main at 3bd7376, checked 2026-10-20. No open upstream
PR existed at inspection; the prior evaluator repair was already merged in
PR #266. Unrelated custody and historical release issues remain open; this
change does not resolve them. No repository AGENTS.md or AGENTS.override.md
exists in this tree. CONTRIBUTING.md, docs/MAINTAINING.md and RELEASING.md
govern its source and publication workflow.

The dated local validation record records
checks, bounded behavioral trials and independent review. Source fixtures
alone cannot prove model compliance. The trials exercise planned behavior in
fresh contexts; they do not qualify Consensus runtime, authentication, billing,
research quality, or any installed host. No comparative-superiority claim is
made. Existing v7 evidence limits and the macOS custody limitation (#162) carry
forward.

Publication and evidence status

This committed note describes the source and required evidence; it does not
claim a later tag, asset or handbook publication has already succeeded. The
release candidate is the final merge commit of PR #267 and reviewed
pre-publication follow-ups PR #268 and PR #269. Their earlier merge
candidates are superseded; the final #269 merge is the intended subject. Exact-candidate results,
publication judgment and actual tag/Release/assets identity must be recorded
against that SHA in the annotated tag and attached publication receipts.

The owner approved this release conditionally on required checks and certified
DCO for the contribution. The agent executes through the existing owner route
under the explicitly recorded operation-specific policy disposition.
That exception is to the fleet execution route, not to the repository integrity
or designated-reviewer judgment gates. It is not direct human execution or
broker conformance. Protected-tag changes require separate action-time approval.

Gate	Status	Exact subject	Evidence	Limits
Version and links	Prepared; exact-candidate result pending	Final 7.3 merge candidate; SHA in attached receipt	Manifest/integration/handbook checks in exact-gates.json	Prospective tag paths checked offline; remote tag not presumed
Deterministic and CodeQL	Exact candidate pending	Final 7.3 merge candidate; SHA in exact-gates.json	Exact-candidate workflow run IDs in attached receipt	PR checks do not certify the later merge SHA
Security and public content	Exact candidate pending	Final merge candidate and complete history	Full-history positive controls and public-content review in attached receipt	Optional Codex Security scan unperformed; no coverage claim
Description bytes	Unchanged descriptions; final comparison pending	Final candidate versus v7.2.0	Both description-budget reports in attached receipt	Package-local bytes only
Harness evidence	Source/package tier; runtime unverified	Final 7.3 candidate	Tier table below and dated source validation record	No new installed-host or Consensus research run; prior host limits carry forward
Publication judgment	Pending frozen candidate	Final candidate SHA in designated-review.json	Exact-candidate judgment and independent review in attached receipt	Actual reviewer/context limits must accompany verdict
Publication identity	Pending	Annotated v7.3.0 tag, Release and assets	API observations and hashes in publication-receipt.json	Protection restoration and handbook verified separately

The v7 comparative-superiority null and unusable-pair results remain unchanged.
The known macOS custody limitation (#162) also carries forward; no guarantee
for case-insensitive POSIX custody is added. Required Ubuntu checks and any
separate diagnostic results must be distinguished in exact-gate receipts.

Harness verification tiers

Harness	7.3 verification tier and limit
Claude Code	Package/source checks; no 7.3 live method invocation
Codex CLI	Prior v7 discovery evidence; 7.3 source/package checks only
Codex desktop	Source execution and local checks in this task; installed 7.3 skill delivery unverified
Cursor	Package/install instructions; 7.3 loaded-host behavior unverified
Gemini CLI	Manifest/source checks; 7.3 method delivery unverified
Antigravity	Manifest/source checks; 7.3 runtime behavior unverified
Kimi Code	Prospective release-source pin and manifest checks; no host load
ZCode / generic Agent Skills host	Canonical tree and explicit-load instructions; 7.3 behavior unverified
ChatGPT / OpenAI	Deterministic bundle checks; upload/load and research-agent use unverified

No harness tier establishes scientific truth, real-time monitoring, new
comparative benefit, or successful Consensus authentication/quota behavior.

Epistemic Skills 7.2.0

Epistemic Skills 7.2.0

V7.2 expands resolve → literature with a research-gap coverage-map contract.
For questions about gaps or under-studied areas, the method now pre-registers a
coverage frame, records the papers and passages behind each cell, treats sparse
cells as POTENTIAL GAP rather than proof of absence, and requires a validation
search before handing a cell forward as a focused research question. This is a
backward-compatible expansion of one method, so the release is minor. The package
still contains seventeen skills: epistemic and sixteen substantive methods.

What changed

• The literature method adds the coverage frame, cell record, validation gate,
  and focused-question handoff. A connector-native matrix is treated as tool
  data and must pass the same validation.
• reference/research-gap-matrix.md gives the cell fields and a worked record.
  Counts, paper IDs, classification rationale, exact passages, source locations,
  validation query results, and residual coverage limits remain inspectable.
• Four declared fixtures describe structural requirements. They have no scorer
  or live agent epoch; they are not evidence of agent behavior. The existing
  resolve/literature trigger-and-scope tests cover the prior trigger contract,
  not the quality of a produced coverage map.
• The package manifests, install instructions, and package integration version
  expectation move from 7.1.0 to 7.2.0. The skill inventory and description text
  are unchanged.

Migration and compatibility

Replace an older install with one tagged v7.2.0 copy, reload the host, and
verify its loaded source path. Existing literature records remain valid; the
coverage map applies when a new question carries a gap, landscape, under-studied,
or “what is missing” premise. A sparse cell remains an orientation signal until
validation is recorded. No new connector or host integration is promised.

Evidence and limits

The content change merged as PR #262 (de0531d). The release candidate is the
merge commit of the version-rotation release PR. Exact-candidate check runs,
the designated review, annotated tag identity, and generated bundle hashes are
recorded in the tag message and attached release receipts. This committed note
does not claim that a later publication step has already passed.

Gate	Status record	Exact subject / evidence	Limits
Version, links and package	Final result in publication receipt	Annotated v7.2.0 target; manifest and package checks; post-tag URL checks	A manifest does not prove an installation loaded
Deterministic checks and CodeQL	Final result in exact-gates.json	Required hosted jobs on the exact merge commit	A green source check does not show an agent using the method
Security, public content and provenance	Final result in exact-gates.json	Exact candidate and reachable history; planted scan controls and public-content review	Pattern checks cannot prove universal absence
Description bytes	7,231, unchanged from v7.1.0	check_description_budget.py --report on both versions	Package-local, not total host context capacity
Harness evidence	Tiers below; no new 7.2 live-behavior claim	Prior v7 host coverage and generated-bundle checks	Discovery, packaging and actual method use are different observations
Designated-reviewer judgment	Final verdict in designated-review.json	Frozen release candidate and reviewer record	Reviewer identity and shared context are disclosed there
Publication identity	Final result in publication-receipt.json	Tag object, Release, assets and Wiki	Verify actual published objects, not this prose alone

Harness verification tiers

Harness	7.2.0 verification tier and limit
Claude Code	Package structure and source checks; no 7.2 live model run
Codex CLI	Prior v7.0 discovery of all seventeen skills; 7.2 loaded-source and method use unverified
Codex desktop app	Current session loads a cached 7.0 skill copy; 7.2 installation and behavior unverified
Cursor	Package and installation instructions only; 7.2 model delivery unverified
Gemini CLI	Extension manifest and instructions only; 7.2 model delivery unverified
Antigravity	Plugin manifest and instructions only; 7.2 model delivery unverified
Kimi Code	Plugin manifest and instructions only; 7.2 model delivery unverified
ZCode	Tagged checkout procedure only; 7.2 discovery and behavior unverified
ChatGPT / OpenAI	Generated bundle validation on the release candidate; no live agent execution
Generic Agent Skills host	Canonical tagged skill tree and fallback instructions; host behavior unverified

The coverage-map method follows the Consensus Research Gaps Matrix
description,
including its warning that an empty cell concerns the retrieved papers, not all
research. This release adds a written validation contract, not a proven
improvement in research outcomes. The v7 comparative-superiority null and
unusable-pair results still stand; no 7.2 comparative trial has run.

The macOS custody limitation (issue #162)
also carries forward: do not rely on distinct-filename or exclusion guarantees
on case-insensitive POSIX filesystems. The dispatch-only macOS diagnostic is
disclosed separately from required Ubuntu custody results in the exact gate
receipt.

Epistemic Skills 7.1.0

Epistemic Skills 7.1.0

V7.1 adds a transfer mode to outsource: alongside delegating a bounded
workload for a verified return, the skill can now hand full ownership of a body
of work to a receiving agent — inventoried responsibilities, recorded custody
acceptance, and origin divestiture. The delegate path is unchanged; this is a
backward-compatible minor version. There are seventeen canonical entries:
epistemic plus sixteen substantive methods, as in v7.0.0.

What changed

• Two named modes. Every outbound request is delegate or transfer,
  chosen by operator intent, recorded in the packet header, and shown in the
  readiness receipt's packet line. Ambiguous intent asks before publishing.
• Transfer packets (outsource-transfer@1, new
  reference/TRANSFER_TEMPLATE.md) replace the bounded outcome with a
  responsibility inventory — in-flight work, standing obligations and
  watchers, authority with provenance, open loops, escalation paths — swept
  from the durable sources that hold obligations, not just the current task's
  context. An obligation that cannot be inventoried is surfaced as a named
  gap, never silently transferred.
• Custody acceptance contract. The terminal event is the receiver's
  explicit read-back of what it now owns (outsource-acceptance@1 envelope),
  verified by the origin for coverage. ACCEPTED transfers the work;
  TRANSFERRED requires the origin's divestiture checklist to reach residual
  NONE (watchers stopped or handed over, credentials handed over or
  revoked, loops left, steward change recorded in affected missions and
  ledgers). A read-back proves comprehensibility, not competence; consequential
  work is directed to a custodied mission before transfer.
• Extended trigger battery. The trigger-and-scope scorer's closed action
  vocabulary gains transfer-packet and verify-acceptance with full field
  contracts; four new fixtures (eighteen total) distinguish a full-custody
  handover from a retained-ownership delegation and exercise the acceptance
  and gap branches; the completion battery adds a synthetic acceptance
  closure with mutation checks.
• Pinned projections. The package-integration suite pins the new contract
  phrases, both templates, and the mode declarations. README, handbook
  (Skill-Outsource, catalog, how-the-pieces-fit), and the host manifests
  describe both modes.
• Description budget. The outsource description grows by 78 bytes
  (7,153 → 7,231) to name both modes; CEILING_BYTES tightens from 8,636 to
  the measured 7,231 so the v7.0.0 slack cannot silently refill.

Migration and compatibility

No consumer-visible contract change to existing delegate behavior: the same
repository layout, prompt shape, publication discipline, and
outsource-relay@1 envelope. Existing docs/outsource/<work-id>/ packets
written before 7.1 carry no Mode row; they are delegate packets by
construction and remain valid. New transfer packets opt in per work ID.
Agents that only ever delegate need no change. The skill count is unchanged
(seventeen skills, sixteen disciplines); install one versioned copy per host
per the installation guide, replace the older tag,
reload, and verify the loaded source.

Evidence and limits

The content change merged as PR #259 (fc167ac); the release candidate is
the merge commit of the version-rotation release PR. Exact-commit hosted
outcomes, the candidate SHA, and the check-run inventory are bound by the
annotated tag object and the attached publication receipt. These notes alone
do not assert that an unperformed publication step passed.

Gate	Status record	Exact subject / evidence	Limit
Version, links and package	Final result in publication receipt	Annotated v7.1.0 target; source/bundle and post-tag URL checks	A manifest is not an installation receipt
Deterministic checks and CodeQL	Final hosted outcomes in tag and receipt	Required runs on the exact merge commit	Historical/local checks are retained separately
Security, public content and provenance	Final outcomes in tag and receipt	Exact candidate and reachable history; scan controls; matching sign-offs	Known patterns, not universal absence proof
Description bytes	7,231; delta +78 from v7.0.0's 7,153; ceiling 8,636 → 7,231	Canonical package descriptions; check_description_budget.py --report	Package-local, not total host context headroom
Host evidence	v7.0.0 tiers carry forward; no new host claims made in 7.1.0	v7 host coverage	Discovery/serialization does not prove automatic use
Designated review	Exact-candidate verdict in tag and receipt	Owner-designated implementing ZCode agent (GLM-5.3)	Shared implementation context; not blinded or independent
Publication identity	Final assertions in receipt	Annotated tag, release body and generated assets	Verify the actual tag and receipt, not this prose alone

Pre-merge local evidence on the content branch (not exact-commit evidence):
outsource trigger-and-scope suite PASS (18 fixtures), synthetic terminal
return PASS including the acceptance mutations, outsource package
integration PASS, skill surfaces in sync, wiki gate PASS, public-content gate
PASS, description budget at ceiling, JSON artifacts checked. A mimosa static
scan completed with eight findings, none in files touched by this release;
the pre-existing findings (gauntlet scripts, cursor hook renderer, a
historical v6 wiki checker) are recorded on PR #259 and remain open for the
owner.

Known limits carried forward

The macOS custody limitation (issue #162)
remains: do not rely on distinct-filename or exclusion guarantees on
case-insensitive POSIX filesystems. The v7 comparative-superiority boundary
also carries forward: no arm separation was ever demonstrated, and 7.1.0 adds
no behavioral-performance claim. The transfer mode's acceptance read-back
verifies coverage of the responsibility inventory, not the receiver's future
stewardship; that limit is stated in the skill itself.

Epistemic Skills 7.0.0

Epistemic Skills 7.0.0

V7 adds a shared usage guide and focused Perspective method, refines the existing
fifteen methods, and improves the distinction between instructions, executable
checks, observed behavior and unproved claims. There are seventeen canonical
entries: epistemic plus sixteen substantive methods.

What changed

• Visible, proportionate use. The usage entry teaches discovery, loading,
  application, acknowledgment and continuation. Every method remains directly
  invocable. Metacognate examines reasoning rather than acting as a dispatcher.
• Debugging that carries through. Triage prefers available applicable
  systematic-debugging, provides a standalone fallback, reuses adequate evidence,
  and continues authorized repair through original-failure verification.
• Focused and plural review. Perspective returns useful focused/adaptive
  analysis. Gauntlet preserves plural scrutiny, reasoned adjudication and dissent.
  The shared library accounts for all 102 original lens IDs and adds three methods;
  six retired entries remain historical, leaving 99 available entries.
• Continuity and native goals. Decision records preserve original predictions
  and later outcomes; completed external relays can close without another prompt.
  Write-goal adapts to actual native limits, state and lifecycle with opt-in budgets.
• Executable outcome checks. Versioned UAT contracts include expected and
  disconfirming observations. Failed persistence prevents acceptance. Watch proof
  history is consistent across applicable states. Gauntlet's template rejects an
  unsupported GO when panels, gates or material finding evidence are incomplete.
• Finite review and clear packaging. The owner-designated reviewer may approve
  a candidate without a mandatory extra model family. One canonical usage body,
  host-specific fallbacks, aligned metadata and a current handbook accompany it.

Migration and compatibility

This major version includes consumer-visible contract tightening since v6:
empty permission values and incomplete evidence accepted by earlier contracts may
now fail validation. UAT uat-contract@2 supports executable observations;
unversioned historical inputs retain explicit legacy semantics. Review validation
errors against the relevant schema instead of rewriting historical evidence.

Install one versioned copy per host, preserve customizations, then reload or start
a fresh task and verify the loaded descriptions/source. Do not restart or rewrite
existing native goals or loops simply because of upgrade. The historical alias
using-epistemic-skills maps to epistemic in documentation; native alias support
is not established. Superpowers is preferred where applicable, not a dependency.

See the installation guide and
handbook. The v6 tag and its historical
handbook remain unchanged.

Evidence and limits

The requirement packet records R01-R22 and the pre-publication
review; the local receipt preserves original failures,
timeouts and corrective checks. All 81 local workflow commands have their expected
outcomes after scoped corrections. The final publication subject and hosted
outcomes are bound by the annotated tag and the attached
publication receipt.
These notes alone do not assert that an unperformed publication step passed.

Gate	Status record	Exact subject / evidence	Limit
Version, links and package	Final result in publication receipt	Annotated v7.0.0 target; source/bundle and post-tag URL checks	A manifest is not an installation receipt
Deterministic checks and CodeQL	Final hosted outcomes in tag and receipt	Required runs on the exact merge commit	Historical/local checks are retained separately
Security, public content and provenance	Final outcomes in tag and receipt	Exact candidate and reachable history; scan controls; matching sign-offs	Known patterns, not universal absence proof
Description bytes	7,153; delta -1,483 from v6's 8,636	Canonical package descriptions; report command	Package-local, not total host context headroom
Host evidence	Explicit tiers in linked coverage	Versioned host coverage	Discovery/serialization does not prove automatic use
Designated review	Exact-candidate verdict in tag and receipt	Owner-designated implementing Codex agent	Shared implementation context; not blinded or independent
Publication identity	Final assertions in receipt	Annotated tag, release body and generated assets	Verify the actual tag and receipt, not this prose alone

Live Codex CLI discovery observed all 17 canonical descriptions. Startup model
consumption, native aliases and native-goal activation/readback remain unexercised.
Other hosts retain their declared verification tiers.

Custody filesystem limitation (issue #162). The dispatch-only macOS
probe remains outside the required merge gate and reproduced the existing
case-insensitive APFS limitation in run 35422446331
on 8854dd735b414997930e768bceead18f740e4868. Its lifecycle step failed
distinct-real-file-untouched and distinct-both-files-tracked-separately;
the required Ubuntu contract job passed. Later macOS steps were skipped,
not passed. Custody filename-distinctness and exclusion guarantees are not
supported on case-insensitive POSIX filesystems. This carries forward the
measured v6 KL-MACOS-162 boundary; the final tag receipt records the same
required-job/diagnostic distinction on the exact release commit. Other skill
instructions do not depend on claiming that unavailable filesystem guarantee.

The bounded v6-v7 comparison stopped after one baseline subject encountered
out-of-arm host skill loading and policy-blocked local tools. Thirty-one slots
were not dispatched; zero valid pairs and zero candidate subjects were obtained.
All scheduled outcomes
remain recorded. V7 makes no comparative superiority or universal activation claim.

Publication follows the release policy and the owner's
PR #250 authorization. The tag is immutable; later corrections
ship as new versions, and later evidence does not rewrite historical outcomes.

v6.0.0 (exception release)

Release 6.0.0

Support point: v6.0.0 — one semantic version, one immutable Git tag, one
evidence set. Supersedes v5.1.0.

Candidate identity. The v6 BUILD freeze was reviewed at candidate
03e972c5d427238033cb90d66846adabaf11928d with its packet at
546ccc8e55eb060379d62198310145f7243ac7bd; both are ancestors of the release
branch and are pinned by pin/es-v6-rc5-candidate-2026-08-19 and
pin/es-v6-rc5-freeze-2026-08-19. The release candidate is the commit produced
by merging this release branch. The integrity gates run against that commit —
not against this file's description of it. The publication gate does not: it
ran four times, on four earlier candidates, and returned NO-GO every time. See
the exception block below before reading anything here as an approval.

Why 6.0.0

The skill surface did not grow: v6.0.0 ships the same fifteen skills as
v5.1.0. The major version marks something else — a security fix in the custody
contract, the completion of the v5 design commitments, and a new assurance
contract that changes what a release of this project is allowed to claim.

The honest summary: v5 claimed things about itself that nothing checked. v6
makes those claims mechanical, and where a claim could not be made mechanical,
it is written down as a limit with an owner.

What ships

Security — mission-custody (es#137). Three P1 false-allow bypasses and four
P2 refusal gaps are closed. These are real permission-boundary defects: paths
that the custody guard should have refused and did not. If you rely on
mission-custody to bound where an agent may write, this is the reason to
upgrade.

The v5 design commitments, completed. All four were implemented and are now
enforced rather than asserted:

• ROUTING.md is generated solely from metadata.hands-to, byte-verified in
  CI, and the hand-authored-routing-table ban is mechanically enforced.
• Every packaged skill carries the intrinsic evidence-emission step
  (skill-run@1) with a schema-validated tracked exemplar; live ledgers are
  runtime-local by amended design.
• Sentinel corpora exist for every skill, with event_kind bound into each
  fixture and a scorer that fails closed in both directions.
• Skill membership has one source of truth: per-skill frontmatter event
  metadata derives the event map, its schema, the verifier inventory, and every
  count surface. Membership drift fails closed.

A new assurance contract (plugins/epistemic-skills/contracts/v6-assurance). A release candidate is
now sealed against an exact commit: per-file sha256 digests of every
inventoried source plus the candidate tree hash, recomputed by a validator that
turns CI red if any inventoried file changes after the freeze. The packet's
readiness state cannot be reached by editing a field — the independent verdict
must exist as an on-disk artifact naming the exact candidate SHA, and operator
acceptance must be recorded separately.

Deterministic-gate hardening. The durable decision ledger is byte-append-only
against the merge base; the DCO check gained planted self-test controls; the
secret scan's one path exemption is anchored and proven narrow in CI on every
run; the clean-room replicates the workflow's Python steps with a completeness
assertion rather than a hand-maintained copy.

Release gate status for 6.0.0

How this file resolves its own identity. A release note cannot contain the
hash of the commit that contains it. Following the v5.1.0 precedent, identity is
resolved by description: the final tag candidate is the commit carrying this
table — the commit produced by merging the release pull request, per
RELEASING.md Procedure step 4. Every other coordinate below is written in hex
because it is resolvable in advance.

Why this file names no run IDs. Procedure step 4 makes the merge commit the
candidate and says any correction mints a new one, so a table of exact-commit
evidence can never sit in the commit it describes.

An earlier draft of this file tried to escape that with a delta-enumerability
argument — it claimed the only change between the commit the runs were
dispatched at and the commit carrying the table was the table itself. That
statement was false, and an independent panel proved it false: at
48009fef938bfa989fb797380080824b050f3bb4 the delta from 92b3ca6c was ten
files, +690/-124, across ten commits, and it silently included the emptying of
blocking_claims and an amendment to RELEASING.md — the two changes a reader
scoping a re-check would most need to see. The device did not merely overstate;
it concealed. It is recorded here as publication-gate finding RG1-01, and it
is retired rather than repaired: no release note in this repository may again
assert its own delta.

What replaces it is step 4's own rule. Exact-SHA run IDs are post-candidate
facts, and they go in the annotated tag object, which names its target without
altering it. This table therefore states which gates were run and what tier of
evidence exists; the tag message states the run IDs, at the SHA they actually
judge. If you are auditing this release, read the tag message for the numbers.

This is an EXCEPTION RELEASE, not a conforming release

RELEASING.md § "Independent judgment gate" defines two outcomes. A
conforming release carries a recorded GO from an independent publication
gate on the exact candidate. 6.0.0 does not have one and will not get one.
Four independent publication reviews were run across three model families and
every one returned NO-GO (§ Provenance). The owner has elected to publish
anyway, under the exception the same section provides.

The five disclosures that section requires before tag creation:

1. The gate did not reach GO. It was run — four times — and returned NO-GO
   four times. This is not an unrun gate; it is an overridden one.
2. No GO exists. Not at this candidate, not at any candidate in the v6
   lineage. No later review can manufacture one: a post-release review adds
   judgment evidence and cannot convert this into a conforming release.
3. Owner, date, scope, authorization. Owner: the repository owner,
   2026-08-21, recorded as decision D24 in
   docs/v6/operator-decision-record-2026-08-20.md with the instruction
   quoted verbatim and its provenance stated. Scope: the v6.0.0 tag and
   GitHub Release only. It does not reach any integrity gate — see below.
4. What evidence remains, and what it cannot establish. The artifact
   itself is well evidenced. Of 31 class claims: 21 PROVED, 8 LIMITED
   within stated bounds, 2 PARTIAL, and none UNPROVED — counted from
   docs/v6/ES6-V6-CANDIDATE/claim-to-proof-matrix.json, whose other 41 rows
   are an open-issue census and are not claims about this artifact. The five
   gating workflows must be green at the exact candidate before this
   authorization fires (pre-authorization condition 1, which states the one
   expected job-level red and its failing tests by name); the full-history
   secret scan and public-content review are green with planted controls; the
   sealed source inventory verifies. What none of that
   establishes is the thing the judgment gate exists to establish: that an
   actor which did not build this release thinks it should be published. Every
   NO-GO was about the publication act and its paperwork, not about the
   skills. The reader should trust the artifact and discount the ceremony
   accordingly.
5. Successor condition / revisit trigger. The standing D8 cross-family
   consult is owed and undischarged; it carries forward to 6.1.0 as a
   blocking obligation, not as a nicety. If any of the four NO-GO findings is
   later shown to have named a defect in the artifact rather than in the
   release process, that is an immediate erratum-and-patch trigger under step
   11.

The exception reaches exactly one gate. RELEASING.md scopes owner
exceptions to the independent judgment gate (RG-8). It does not reach the
integrity gates — RG-1 accuracy, RG-4 alignment, RG-5 deterministic evidence,
RG-6 security and provenance — and no authority in this repository waives
those. Every one of them is met on its own terms below. A release note that
lied would still be a lie with an owner's signature on it; that is why RG1-01
was fixed rather than waived.

Item	Status	Evidence
1 — candidate identity and scope	met	Freeze candidate 03e972c5d427238033cb90d66846adabaf11928d, packet commit 546ccc8e55eb060379d62198310145f7243ac7bd, both ancestors of this branch and pinned by pin/es-v6-rc5-candidate-2026-08-19 / pin/es-v6-rc5-freeze-2026-08-19 (lightweight — see PG-17 below). The release candidate is the merge commit of this branch.
2 — release decisions and risk acceptance	met as an exception, with one bounded gap	Operator acceptance is recorded as D24 (2026-08-21) in docs/v6/operator-decision-record-2026-08-20.md, electing the RG-8 exception with the five required disclosures made above before tag creation. The packet still carries no operator_acceptance object and still reads NOT_READY: that field tracks conforming readiness, and an exception release does not make a packet ready. Leaving it NOT_READY is the honest state, not an oversight. Bounded gap RG4-05: the D8 cross-family consult is owed and undischarged; it is disclosed above as the successor-release condition and is the one RG-2 obligation this release does not satisfy.
3 — evidence retention	met	All nine verdicts of this lineage — the five BUILD panels and the four publication reviews — are in-tree under docs/gauntlet-runs/ with an index at `docs/gauntlet-runs/V6...

v5.1.0 — manifest ships, corrections land, custody's verify verb stops lying

Release 5.1.0 — manifest ships, corrections land, custody's verify verb stops lying

Date: 2026-08-15. Minor. Fifteen skills — the fourteen of
5.0.0 plus manifest, the mission-custody seat, published as a canonical
junction-reachable skill.

What this releases: (1) the manifest skill reaches every host that
consumes the packaged skills/ tree, not just plugin-marketplace harnesses;
(2) all of the never-published 5.0.1 correction work lands; (3) the custody
CLI's verify becomes what its name promised — read-only — closing es#138,
the defect that stranded a live mission.

Why 5.1.0, and where 5.0.1 went

release/5.0.1 was cut as a fourteen-skill correction patch but never
tagged. While it sat, main advanced thirty commits — including the entire
mission-custody contract family (#114) — past the branch point. Every
candidate whose commit could live in main's history therefore carried the
fifteenth skill, making RELEASE-5.0.1's "fourteen skills, unchanged" claim
untrue for any mergeable tree, and RELEASING.md step 9 requires the release
commit to be contained in main. An honest v5.0.1 had become impossible.

The 5.0.1 number is retired, unpublished. No tag ever existed, so no tag
moved. Its corrections — install-metadata honesty, the hardened
phantom-skill guard, the 5.0.0 errata and record corrections, the UTF-8
description-budget measurement — are carried here in full; see
RELEASE-5.0.0.md, RELEASE-5.0.0-ERRATA-2026-08-06.md, and the design
conformance records of 2026-08-06 for the underlying history.

What ships

1. manifest as a canonical skill (es#178)

The mission-custody seat already lived on main under
plugins/epistemic-skills/skills/manifest/ (added by #114, after the 5.0.1
branch point). This release makes it loadable by every host that consumes the
canonical skills/ tree — junction-projected harnesses included — instead of
plugin-marketplace harnesses only.

• Operator decision (2026-08-15, es#178): publish as canonical; junction
  hosts stay tag-pinned (they gain the seat at this tag); the generic-host
  caveat applies — the host must supply the custody CLI/hook runtime.
• ADR-184's four-layer precedence now holds on every install surface:
  wherever the epistemic seat ships, the agency seat ships with it.
• Description budget re-measured for fifteen: 8,636 bytes at ceiling
  (check_description_budget.py green; was 8,159 across fourteen).
• README gains the manifest catalog row and the entry-point carve-out
  (manifest may also be invoked directly — manifest this / /manifest).

2. The 5.0.1 corrections, landed

Install metadata no longer advertises deleted skills (the 5.0.0
using-epistemic-skills/helix overclaim), the member lists are generated
from the packaged set, check_no_phantom_skills.py carries the 5.0.0
deletions in its RETIRED map and scans marketplace.json + root instruction
prose, and the 5.0.0 release-record corrections (items 6/7/8) stand as
recorded in the errata. All version-bearing surfaces read 5.1.0; skill-count
claims read fifteen everywhere the surface-sync contract checks.

3. Custody verify is now read-only (es#138)

verify was begin_verification wearing a read verb's name; a read-only
auditor moved the live media-library-rebuild mission active → verifying
through it, with no clean path back. Now:

• verify runs Mission.verify_chain() — walks every checkpoint, checks
  prev_checkpoint_sha256 linkage, mutates nothing (read_only: true
  in the report; exit 0 intact / 4 walk-break; schema-broken chains still
  fail loudly at load, exit 2).
• The transition is begin-verification, matching the API.
• No mutating alias — old muscle memory gets a hard error, not a silent
  state change. That asymmetry is the fix.
• Oracles added: no-checkpoint-appended + status-unchanged on a healthy
  chain; tampered checkpoint → nonzero exit naming the file, zero writes.

Release gate status for 5.1.0

Per RELEASING.md §Release gate. Candidate chain: release content
7ba1f19 (PR #179 squash) → gate-evidence commit 8180554 (PR #180 squash,
against which the Gauntlet panel 1 ran and returned NO-GO as-specified,
GO binding on a named discharge set) → this discharge commit, which
executes that set. The final tag candidate is the commit carrying this
table together with the committed Gauntlet record at
docs/gauntlet-runs/es-v510-publication-2026-08-15/. Verification is at
job and step level, not by run label.

Item	Status	Evidence
4 — version/link alignment	met, after four in-tree stalenesses and two live out-of-tree stalenesses were found and fixed	all ten version-bearing surfaces at 5.1.0; surface-sync --check green (15 skills / 14 disciplines). In-tree (found by the release-diff review, fixed in 8180554): two Cursor install guards and the Kimi /plugins install URL still pinned to v5.0.0, and the README support-point paragraph still describing v5.0.0's gate history as 5.1.0's. Out-of-tree (found by Gauntlet panel 1, fixed before this commit, re-probed live 2026-08-15): the GitHub repo description still advertised the router seat deleted in v5.0.0 (executing the unchecked order in RELEASE-BODY-AMEND-v5.0.0.md), and the README-linked public wiki Installation page installed v5.0.0/fourteen with tree/v5.0.0 URLs — now v5.1.0/fifteen (epistemic-skills.wiki e6c6ba7..710dd2b); the full v5.1.0 handbook pass is a recorded post-tag follow-up.
5 — deterministic suite, DCO, parity, JSON, CodeQL	met	six workflows green at 7ba1f19, step-verified: epistemic-flexibility run 31894206384; release-security run 31894206380; openai-bundles run 31894206377; mission-custody-contract run 31894206404 (contract-macos correctly skipped — dispatch-only probe job); commission-watch-contract run 31894206405; CodeQL run 31894206293, all three Analyze matrices success. Four workflows green at 8180554 (31895143032/030/022/005) and at the discharge commit 2890ae6. The two path-filtered suites were dispatched via workflow_dispatch at 2890ae6 per panel-1 ruling: openai-bundles run 31897020113 success; mission-custody-contract run 31897018984 — the required contract job success, while the dispatch-only contract-macos probe job (built to settle es#162) failed 2 case-distinctness lifecycle tests at 2890ae6 and again at the final candidate: es#162 is settled negative — macOS-default (case-insensitive) filesystems do not support the custody CLI's case-distinct multi-file tracking; a prior dispatch at 8aa4ffe on 2026-08-13 (run 31723939498) had already failed the same step, so the settlement predates this release cycle and the dispatch runs triple-confirm it (31723939498, 31897018984, 31898055755); disclosed as a known limitation below and adjudicated by the panel-2 re-affirmation (the "red blocks the tag" clause is ruled there against the probe-job attack). Both suites are re-dispatched at the final tag SHA; their run IDs and conclusions are recorded in the annotated tag message.
6 — secret scan and public-content/provenance review	met, both halves, at both scopes	scan: release-security runs 31894206380 and 31895143030, every step success — including the planted-secret positive control and the digest-allowlist narrowness control — before the full-history gitleaks pass. Public content: check_public_content.py --self-test + live run both exit 0 in stdlib-checks (runs 31894206384, 31895143032), plus the manual release-diff review below — extended at panel-1 direction from the 19-file PR diff to the full 266-file v5.0.0..candidate window, whose findings and dispositions are recorded in the second subsection.
7 — harness surfaces exercised or assigned an honest tier	met via explicit tiers; one live junction verification lands post-tag	the README install table carries per-harness honest boundaries, now including the generated ChatGPT/OpenAI bridge (snapshot artifact, no live execution). Cursor's recorded behavioral epoch stays BLOCKED_EXTERNAL. ZCode: the junction mechanism was verified at v5.0.0 content on one fleet device (example-workstation) with no recorded procedure — the v5.1.0-specific junction count (fifteen) is verified by the same-session post-tag installer -Verify receipt; plugin install remains untested. Other plugin surfaces carry the per-harness tiers recorded in HARNESS-VERIFICATION-MATRIX-SUCCESSOR-2026-08-07.md (recorded at fourteen-skill content; the fifteenth seat adds no new runtime primitive on those surfaces — its caveat, the host-supplied custody runtime, is the README row's). No new live plugin-harness executions in this release.
8 — independent Gauntlet publication review reaching GO	GO — recorded, binding on the discharge set this commit executes	panel record at docs/gauntlet-runs/es-v510-publication-2026-08-15/GAUNTLET-SUMMARY.md: panel 1 (standard, five isolated lenses + judge) ruled NO-GO on publishing 8180554 as-specified (P1 ×2, P2 ×5) and GO binding on execution of the named discharge set; this commit executes every pre-tag item (the two zero-commit live-surface fixes verified above; the row amendments and full-window review in this file; both suite dispatches at the final SHA), and a delta-scoped panel-2 re-affirmation runs on this candidate before tag creation — a regression there reverts to iterate, no tag. Independence limits (per RELEASING.md): lenses, arbitrator, and the release-evidence author share one model family (GLM-5.3) in isolated sub-agent contexts; every load-bearing mechanical row was re-verified independently via the live GitHub API; no cross-family or human adjudication ran — that ceiling is recorded, not waived.

Release-diff public-content review (item 6 record)

Scope: git diff 1e36da9..7ba1f19 — 19 files, +303/−48 (version surfaces,
README, release notes,...

v5.0.0 — the loop release

Release 5.0.0 — the loop release

Date: 2026-08-06. Breaking. Fourteen skills: one entry point,
metacognate, and thirteen disciplines.

PUBLISHED, with two gate items not fully satisfied. Read the gate table before
relying on this release's assurances.

Items 4 and 5 are met on the exact tagged commit — version surfaces, the
deterministic suite, and CodeQL all passed there, verified at job and step level.

Item 6 is only PARTIALLY met. The full-history secret scan passed; the
public-content/provenance review it also requires was never performed. The tag
annotation says item 6 was met — that is an overclaim, corrected below and
uncorrectable in the tag itself.

Item 8 was WAIVED by the repository owner and never run. No adversarial
publication review, no GO verdict. Item 7 was never assessed as a gate row.

These are stated at the top rather than in a footnote, because the alternative is a
release record that reads as though the gate passed.

What changed and why

v4.0.0 consolidated eighteen skills into eleven. v5.0.0 does something different:
it removes the routing layer entirely and adds the four capabilities that
close an operational loop.

v5.0.0 skill	Owns	Refuses
metacognate	how much process this deserves — usually none	enumerating its members, ever
health	is the subject within declared bounds — and did we look?	rendering UNKNOWN as OK
triage	what is the cause, and what observation rules out the alternatives	naming a cause it did not observe
did-it-land	is the change in effect on the runtime	a green check whose oracle only read source
watch	a bound was crossed while nobody was looking	claiming installed before it has fired once

Together: watch notices, health assesses, triage diagnoses,
did-it-land verifies the fix took.

Each carries four-valued states where the third value is the point — UNKNOWN,
UNVERIFIED, NARROWED, SUSPECT. Absence of evidence never renders as
success.

Removed

using-epistemic-skills (the router), helix, and
helix/reference/composition-contract.json (the pair table).

metacognate replaces both seats. A pair table maps stages; it cannot express
the two things that turned out to matter — that either strand may interrupt the
other, and that control must come back to the point of interruption.

The check that deleted itself

ROUTER_DESCRIPTION_DRIFT required the router's frontmatter to enumerate every
discipline. It was the single largest source of the enumeration tax — adding any
skill forced an edit to another skill's firing surface — and it is precisely the
defect that shipped in v4.0.0, where the router's description named two skills
that no longer existed.

metacognate enumerates nothing, so the check had nothing left to check. The
package's integration suite now asserts the inverse: it globs the skills
directory and fails if the entry point names any member. Enumeration went from
enforced to forbidden.

The measured constraint behind the release

The Claude Code harness applies a total description-byte budget to its
assembled skill listing. A skill whose description is dropped cannot fire on
description match — it is functionally uninstalled, with no error anywhere.

Confirmed by reversible manipulation and then by intervention:

108 skills             -> one description dropped
111 (+3 probe skills)  -> five dropped
108 (-3 probes)        -> back to one
100 (-8 commands)      -> zero; the dropped skill returned

Adding a skill is a transfer, not an addition — it silently uninstalls roughly
its own byte-weight of other skills' descriptions. Total description budget across
all fourteen skills: 8,230 UTF-8 bytes, measured as the harness resolves them, so a
quoted description is charged for its content and not its delimiters.

This is why consolidation here is a resource constraint rather than a preference.

Known limitation — this release shipped over its own budget

v5.0.0 is net +1,389 description bytes. Each individual description was held
inside sibling range; the sum was never checked against the live budget. The
result was that triage and watch — two of the four skills this release adds —
could not fire at all until unrelated slash-commands were deleted elsewhere on the
installing machine. No error was raised anywhere, because there is no error to
raise: a dropped description simply stops matching.

That is per-item discipline with no aggregate check, which is precisely the defect
class this package exists to catch, committed by the package itself.

The gap is now closed at the artifact boundary:
.github/scripts/check_description_budget.py fails CI if the packaged total
exceeds a recorded ceiling, so any increase must be paid for in the same diff.
It does not and cannot observe the harness cap — that is a property of the whole
installed estate — but it does make this package's own contribution a number that
was chosen rather than one that drifted.

If you install this alongside a large existing skill collection, budget for it.
The measured figures below are the only ones we have; the exact ceiling was bounded,
never bisected.

Migration from 4.1.0

• using-epistemic-skills and helix no longer exist. Invoke metacognate
  instead. It is the only skill invoked by name; every other member fires on its
  own description.
• Trigger vocabulary maps directly: "route this", "which skills apply",
  "pair this with my workflow" → metacognate. Its Tier 2 replaces the router's
  routine gate, and declining is its most common correct outcome.
• Install exactly one mechanism per harness, as before. Replace the older
  copy; do not stack them.
• Evaluation corpora moved from skills/using-epistemic-skills/evals/ and
  skills/helix/evals/ to package-level evals/. Any local automation
  referencing those paths must be repointed.

Evidence posture (honest status)

• Deterministic suite: the complete local gate passes on the release commit —
  0 non-usage failures across every step the workflow declares, at 14 skills /
  13 disciplines.
• Clean-room verification: run in a fresh Linux clone of the release ref, on a
  machine that did not author the change. It found two real defects on its first
  run that every Windows run had passed — an environment-dependent guard, and a
  broken relative link.
• Behavioral superiority: UNESTABLISHED, unchanged. The four-arm campaign
  found no arm separation (primary D>A p=0.875; A=5 B=4 C=7 D=4 of 18).
  Nothing in this release alters that. A tidier architecture is not evidence that
  the skills improve outcomes, and this release does not claim it is.
• Two eval batteries were retired, their subjects having been deleted:
  the composition contract battery, and the blinded-proportionality battery which
  asserted the router's enumerated routing content. Results and experimental
  design were kept in full; only the harnesses were retired. Each carries a
  RETIRED.md.

Release gate status

Per RELEASING.md §Release gate. Items 5 and 6 were satisfied on 2026-08-06 against
the exact commit 3a18cd3, verified at job and step level rather than by run label:

Item	Status
4 — version surfaces aligned, link-existence checked	met. All ten surfaces agree on 5.0.0. Version-pinned URLs were checked individually: .../using-epistemic-skills/reference/routine-fast-path.md does not exist in the v5 tree, so that link was repointed to the file's new home rather than blind-bumped — the P1 class caught in v3.2.0.
5 — deterministic suite, DCO, manifest parity, committed-JSON, CodeQL	met on 3a18cd3. epistemic-flexibility job stdlib-checks succeeded across 45 steps — 44 success and one correctly skipped (Durable ledger append-only against merge base, which needs a PR merge base and has none on a dispatch run). CodeQL Analyze (actions), (javascript-typescript) and (python) all succeeded on the same commit.
6 — redacted full-history secret scan and public-content/provenance review of the release diff	PARTIALLY MET. The scan half passed on 3a18cd3 and again on the tagged commit: release-security job full-history-secret-scan succeeded, including its own positive control step "Prove the scanner detects a planted secret" — so the clean result is a measurement, not a scanner that finds nothing. The public-content/provenance-review half was never performed and no v5.0.0-dated review document exists. See Correction below.
7 — supported harness surfaces exercised live or assigned an honest verification tier; known limitations recorded	NOT TRACKED IN THIS TABLE. Addressed narratively in Evidence posture above (clean-room run, UNESTABLISHED behavioural status, known limitations) but never assessed as a gate row. Recorded here as untracked rather than silently omitted.
8 — Helix routing is recorded and an independent Gauntlet publication review reaches GO	WAIVED, not met — both halves. Explicitly waived by the repository owner on 2026-08-06 ("i approve the release and we don't need the full gauntlet"). No Gauntlet publication review was run for 5.0.0 and no GO verdict exists. The first half is additionally unsatisfiable as written: RELEASING.md item 8 still requires recording "Helix routing", and this release deleted Helix. That text needs updating in RELEASING.md itself.

How items 5 and 6 became reachable

Neither gate could be run on demand. Both were reachable only as a side effect of
pushing a commit or opening a pull request — which cannot satisfy a gate that asks for
a result on the exact release commit. workflow_dispatch was added to both (#99),
and the first dispatched run immediately failed on a real defect this session ...

v4.1.0 — the evidence release

Release 4.1.0 — the evidence release

Date: 2026-08-04. Additive, non-breaking. Same eleven skills as
v4.0.0 (router, helix, nine disciplines); no trigger-surface contract
changes. Operator-authorized (2026-08-04 session: merge of PR #83 and the
release motion approved).

What this release adds

• Post-consolidation evidence exists now. The v4 Tier-1 epoch wave ran
  under a committed preregistration
  (docs/superpowers/specs/2026-08-04-v4-tier1-epoch-preregistration.md):
  recon candidate PASS 14/14 (first battery to clear its first epoch
  with zero contract-shape failures), resolve literature PASS 14/14
  (second consecutive clean epoch), recon brief 12/14, resolve probe
  11/12, recon initiative 11/13 — 62/67 valid trials, zero
  mode/instrument-selection failures (the consolidation's novel risk
  class did not appear), one genuine hard-negative over-fire recorded
  honestly (the program's first), one dispatch-defect run quarantined as a
  defect record. The decision-ledger resume mode re-ran its full battery:
  skilled arm meets the gate in all 3 runs (8/8 traps, 0 false flags),
  with same-model baseline context recorded.
• Live-epoch methodology is law:
  docs/policy/LIVE-EPOCH-DISPATCH-CONTRACT.md (vocabularies inline,
  silence semantics, extraction-only prose tolerance, simulation-clause
  rule 4, opaque keys, preregistration before dispatch, results as-is).
• New battery: recon candidate-mode trigger-and-scope (14 fixtures,
  adversarially verified, CI-wired) — every v4 mode now has an eval
  surface.
• Trace dialects unified: validate_trace.py is the single
  structural authority (string|array residual_uncertainty; scenario
  validated when present); score_behavior.py --bound serves blinded
  harnesses. Real-incident behavioral fixtures 07/08 landed; the
  contract-vs-nothing ablation is preregistered as design-only
  (Tier-3 gated, not run).
• Calibration coordination advanced to the edge of Phase 2: Phase 0
  complete (counterpart confirmed, digests re-verified), Phase 1 charted
  and frontier-resolved
  (docs/coordination/2026-08-04-phase1-inventory-and-map.md — the full
  measurement matrix with subject revisions and sampling frames), the
  charter amended to test current reality (latest release tag + HEAD),
  the paired proposal filed (issue #84), and pin-tag reachability now
  CI-guarded (check_pin_tags.py; v4.0.0 is the counterpart's v4
  re-pin coordinate and is guarded like a pin tag — neither guarded tag
  is ever deleted or moved).
• Gauntlet Step-7b consult packet committed (#40, operator hand-carry
  pending); runtime tool-call gate productized as honest
  cooperative-grade reference machinery (#42); open-questions evidence
  deposited durably (#63); wiki v4.0.0 hand-off committed under
  docs/wiki-updates/v4.0.0/ (wiki pushes are proxy-blocked from
  sessions).

Evidence posture (honest status)

• Tier 0: 37 unconditional CI steps green at release (two new:
  candidate battery tests, pin-tag reachability guard).
• Tier 2 at this release candidate: every trigger battery has a
  current epoch under the evidence policy's subject-hash rule — the five
  re-armed post-consolidation surfaces epoched 2026-08-04 (results above),
  and the unchanged subjects (write-goal 14/14, outsource 12/14,
  open-questions 8/10, context-audit 8/14 — failures diagnosed as
  #77-class reporting/battery-design divergences) retain their valid
  2026-08-04 epochs. The resume-fixtures battery holds its v4 gate PASS.
• Behavioral superiority: still UNESTABLISHED. The four-arm campaign's
  null stands; issue #39 remains open; the contract-ablation design is
  committed but not run. No superiority language attaches to this
  release.
• Risk acceptance: RELEASE-3.0.0-RISK-ACCEPTANCE.json remains the
  controlling record, append-only. FR3-R1/R2 were re-adjudicated ahead of
  this tag (revisit-before-tag honored), re-accepted unchanged, triggers
  re-dated to the next tag after v4.1.0 or 2026-10-31.
• Single-model-family evidence throughout the 2026-08-04 waves
  (claude-fable-5, N=1 per fixture; N=3 runs for resume). Every claim
  above cites its committed artifact, per the evidence policy's rule 5.

Migration from v4.0.0

None required. Install surfaces re-pin to v4.1.0; the skill inventory,
trigger vocabulary, and ECS event contract are unchanged (v4.0.0
remains the counterpart's valid contract-pin coordinate — see issue #84
for the coordination state).

What this release does not claim

No behavioral-superiority claim; no claim that the epoch passes measure
real-world effectiveness (trigger/shape evidence only, per the evidence
policy); no cross-provider generality; the calibration loop is designed
but not yet fed (field pairs ≈ zero until the store lands and Phase 2
runs).

v4.0.0 — the consolidation release

Release 4.0.0 — the consolidation release

Date: 2026-08-04. Breaking. Eleven skills: the using-epistemic-skills
router, helix, and nine disciplines — recon, resolve, decision-ledger,
write-goal, outsource, open-questions, context-audit, gauntlet,
evidence-locked-uat. Operator-authorized (2026-08-04 session: consolidation
assessment approved; merge, resolution of PR #79, and release approved).

What changed and why

v3.x grew to eighteen skills faster than evidence accumulated, and every new
seat cost ~12 hand-synced enumeration surfaces. 4.0.0 inverts both curves:

v4.0.0 skill	Consolidates (as modes/instruments)
recon	blindspot-pass (brief mode) · wayfinding (initiative mode) · harvest-before-adopt (candidate mode)
resolve	applying-formal-rigor (derivation) · evidence-research (literature) · throwaway-prototyping (probe)
decision-ledger	continuity-verify (resume mode, pre-arc) + outcome reviews (first-class trigger)
(craft doctrine, not disciplines)	intent-traced-merge · agent-interface-design → plugins/epistemic-skills/reference/craft/

Every absorbed method survives verbatim (mode/instrument files are the former
SKILL.md bodies, moved with git history); every old name survives as
mode/instrument vocabulary; every committed battery and epoch result moved
with its method as historical evidence.

Migration from v3.4.0

• Install surfaces are unchanged in mechanism; the package now registers
  eleven skills. Remove any older copy first (one mechanism per harness).
• Trigger vocabulary maps 1:1: "run a blindspot pass" → recon (brief);
  "wayfind this" → recon (initiative); "should we adopt X" → recon
  (candidate); formal-rigor/evidence-research/prototyping requests →
  resolve's instruments; resumption re-anchoring → decision-ledger's resume
  mode. intent-traced-merge and agent-interface-design are reference
  doctrine, read on demand, no longer routed.
• The ECS contract enumerates the eleven producers; retired producer names
  in previously collected events validate against the pinned pre-4.0
  contract revision (immutable tags), per the contract's own versioning rule.

Evidence posture (honest status)

• Tier-0 deterministic: 35 unconditional CI steps green at release,
  including the new skill-surface generator (sync_skill_surfaces.py --check) that derives every inventory surface from the filesystem glob +
  skill-event-map.json.
• Trigger epochs: the 2026-08-04 waves ran first live epochs for all ten
  pre-consolidation batteries (94.6% pass under born-pinned contracts; every
  failure reporting-shape over correct conduct; zero over-firing on 51+ hard
  negatives — register: issue #77). These epochs predate the consolidation:
  the merged trigger surfaces of recon/resolve/decision-ledger are new
  subjects and re-arm at Tier 1 per docs/policy/EVIDENCE-POLICY.md. No
  post-consolidation epoch has run at release time.
• Arbitrator certification: the amended planted-flaw battery (AC-07 =
  seat-provenance neutrality) ran blind 2026-08-04 — 10/10 catch, CERTIFIED
  at standard rigor (same-model-family caveat stands).
• Behavioral superiority: UNESTABLISHED, now with evidence. The four-arm
  campaign ran under its committed design (72/72 blinded seeded trials,
  preregistered, exploratory size): no arm separation (primary D>A
  p=0.875; A=5 B=4 C=7 D=4 of 18). The run surfaced two structural findings
  — the two-trace-dialect defect (declared uniform adapter; unification is
  open work) and that the shared trace contract, not discipline prose,
  carries the structure at smoke fidelity. Issue #39 remains open; no
  superiority language attaches to this release.
• Risk acceptance: RELEASE-3.0.0-RISK-ACCEPTANCE.json remains the
  controlling record, append-only (G3-R1's exit criterion met 2026-08-04 and
  recorded in revisit_history; all other accepted scopes unchanged). The
  upgrades-landing gauntlet run's rulings R1–R12 are all discharged.

Governance shipped with this release

• docs/policy/EVIDENCE-POLICY.md — tests bind to claims, not calendars
  (subject-hash-armed epochs; once-per-claim campaigns; field-tier
  thresholds).
• docs/superpowers/plans/2026-08-04-v4.0.0-consolidation.md — the executed
  consolidation plan with its evidence-continuity rules.
• Creation-gate record docs/audits/2026-08-04-creation-gate-revisit.md —
  calibration-review DEFERRED behind a mechanical mint threshold;
  outcome-review promoted in place.

What this release does not claim

No behavioral-superiority claim; no claim that consolidation improves
outcomes (the campaign motivates simplification and testability, not a
performance claim); no post-consolidation epoch evidence; single-model-family
evidence throughout. Every claim above cites its committed artifact, per the
evidence policy's rule 5.

v3.4.0 — wayfinding + throwaway-prototyping + intent-traced-merge

epistemic-skills 3.4.0

Release record date: 2026-07-31
Intended channel: stable
Release type: minor — three new disciplines, five amendment sets, router/helix
ripple, backward compatible
Validity contract: this record is authoritative only when the exact-commit
gates pass and the annotated tag plus non-draft GitHub Release satisfy the
publication-identity checks in RELEASING.md.

What's new

• wayfinding (new discipline): decompose large foggy efforts by
  decisions, not tasks — a dependency map whose nodes are decisions, a
  frontier rule (work only decisions whose prerequisites are resolved), and
  build tickets minted only from fog-free regions, each carrying the
  three-fact handoff (resolved dependencies, proving observation, the
  upstream decision whose reversal would invalidate it). Falsifiable gate:
  any open ticket walks upstream to fully-resolved ancestors.
• throwaway-prototyping (new discipline): resolve a live decision by
  building a disposable probe under a four-clause contract — one
  pre-registered question, disposal declared at birth, answer captured
  durably before deletion, never promote prototype code. The concrete form
  of the "bounded reversible probe" closure control.
• intent-traced-merge (new discipline): resolve non-trivial merge/rebase
  hunks by tracing both sides to their origin commits and the intent each
  served; preserve both intents or record the drop; verify against BOTH
  origins' motivating checks; record per-hunk rulings in the merge commit.
  Deliberate correction of community "never abort" absolutism: aborting is
  the return path, not a failure.
• Amendment sets (five): open-questions — frontier discipline for
  dependent docket items; gauntlet — revision-loop discipline (revised
  subject = new attack surface, delta-scoped re-review, hard cap of three
  panels per subject lineage); decision-ledger — outcome reviews with the
  anti-hindsight boundary (prediction and result recorded as separate
  untouched facts; generalized lessons need operator approval) + prototype
  finding capture; evidence-locked-uat — two new oracle-honesty rows
  (artifact existence is not render proof; unread console/error channels are
  unexercised oracles, relevant non-empty error sets are hard FAIL);
  agent-interface-design — domain-vocabulary naming clause (never mint a
  synonym for a concept the project already names). gauntlet-generator now
  names throwaway-prototyping as the execution discipline for
  prototype-instantiated options.
• Ripple: router handoff/routing tables + arc notes, helix pairing map
  (+3 rows), README (seventeen skills / fifteen disciplines), GEMINI.md, ten
  version manifests + root plugin.json → 3.4.0, outsource integration-test
  assertions (17 dirs).

Provenance and evidence posture

Idea provenance, all re-derived and re-worded (no text copied):
ConnorGriffin/skills (MIT; wayfinder, grilling frontier, plan-review delta
cap, prototype patterns), the Matt Pocock agentic-coding community synthesis
(merge-conflict intent-tracing, prototype disposal, vocabulary clause), and
the fleet's own field practice. Each new skill names its provenance inline.

Honest limits: the three new disciplines have NO eval battery or
behavioral epoch (register: issue #70, scope extended to 3.4.0). The 3.2.0
evaluation posture is inherited, not extended. The gauntlet revision-loop cap
(three panels) is a design judgment, not a measured threshold — revisit on
the first field case it bites.

Migration from 3.3.0

Update the install coordinate to v3.4.0 and reload the harness. No existing
trigger, contract, schema, or output shape changed; amendments are additive
clauses inside existing sections.

Erratum (2026-08-04, post-tag)

The honest-limits register moved: "NO eval battery or behavioral epoch
(register: issue #70)" was true at tag time, but PR #73 landed deterministic
trigger-and-scope batteries for all five v3.3/v3.4 disciplines shortly after
the tag, and issue #70 was closed on the authoring half with the live-epoch
half explicitly unmet. The live register is now issue #77 (behavioral epoch
program). "No live behavioral epoch" remains the accurate residual limit.

Erratum (2026-07-31, post-tag)

The migration section's "No existing trigger, contract, schema, or output
shape changed" overstates: the evidence-locked-uat oracle-honesty rows
TIGHTEN the acceptance gate — a non-empty console/error set relevant to the
exercised surface is now an explicit hard FAIL where the prior text left it
unstated. Consumers whose UAT runs previously passed with unread error
channels may now fail; that is the intended behavior change. Provenance
coordinates for this release's external attributions: PROVENANCE-3.4.0.md.

Harness verification tiers

Unchanged from 3.3.0: Claude Code primary (deterministic suite on the release
commit); Codex/Gemini/Kimi manifest-validated; Cursor packaged but not
publicly listed.