name: docs-strategy-update description: Gate + apply changes to the strategy docs (lean-canvas, hypotheses, theory-of-change, definitions, progress, research-programme) when strategy shifts — runs 9 anti-drift gates before writing. Owns the strategy-doc layer; /kdd owns decisions.md. when_to_use: "Before editing docs/lean-canvas.md, docs/hypotheses.md, docs/theory-of-change.md, docs/definitions.md, docs/progress.md, or docs/research-programme.md — especially after a strategic pivot in conversation. Sync mode (a described change) or audit mode (no arg). NOT for decisions.md (that's /kdd) or CLAUDE.md (that's /slava:maintain:claude-md)." version: 1.5.0

/slava:maintain:docs-strategy-update

Gate + apply changes to ClarityPledge's six strategy docs — lean-canvas.md, hypotheses.md, theory-of-change.md, definitions.md, progress.md, research-programme.md — so a strategic shift gets written down without the recurring drift failures: premature-fact, reversal-churn, cross-doc contradiction, unpruned bloat, claim-without-disproof, fragile refs, competing single-valued directives, degeneration-blindness.

Announce at start: "Running /docs-strategy-update."

This skill owns the strategy-doc layer. /kdd owns decisions.md + meta-reflection. /slava:maintain:claude-md owns CLAUDE.md and .claude/rules/.

When to use this vs other skills

| Situation | Skill | |---|---| | A strategic shift must land in lean-canvas / hypotheses / theory-of-change / definitions / progress / research-programme | /docs-strategy-update ← here | | "Is the programme progressing or degenerating?" — a verdict, not a doc edit | /slava:maintain:programme-health (writes nothing) | | A refutation appears to land on the hard core (K1–K7) | /slava:maintain:programme-health first — a core edit is not a sync | | Capture a decision + run session meta-reflection | /kdd (writes decisions.md) | | Change CLAUDE.md or a .claude/rules/ file | /slava:maintain:claude-md | | Output is a tracked feature, not a strategy-doc change | /slava:build:create-spec | | A stated fact in a strategy doc is wrong — a number measured to be false, a figure a later decision retired, a fired threshold, a broken pointer | Correct it DIRECTLY. Do not run this skill. See below. |

The nine gates guard strategy CHANGES, not defect corrections — this skill is not the router for both

Discriminator, one question: does any bet, position, priority or claim MOVE?

No → it is a defect correction: the doc simply stops asserting something false. Fix it directly, then record what and why in decisions.md. The log entry is the audit trail; the gates add nothing, because there is no drift to detect in replacing a wrong number with a measured one.
Yes → it is a strategy change. Run this skill, all nine gates.

Why this exists (2026-08-27). A prod measurement on 2026-08-14 refuted a completions figure that six strategy docs, three live feature specs and a blog post staged for external publication were quoting. The correction was written, committed, and never propagated for 13 days — because the follow-up task routed it here, and said so in its own words: "needs all six strategy docs read in full plus nine gates with quoted artifacts, which is why it was deferred rather than attempted at low context." Deferred three times. The gate's cost was paid by the error, not by the fix.

The boundary that ruling does NOT license — learned the same day, the expensive way. Direct means without these gates, never without checking. The first attempt at that very correction inverted the source: it re-measured one leg of an OR kill-criterion and declared the criterion void, which is the exact error the source entry had already caught in itself and warned about in writing — and it fabricated a quotation into the append-only log. Both were found by an adversarial reviewer, neither by the author, who had read the source and stopped six lines short of its self-correction. So:

Read the cited source to its end, including anything after the References line. Correction entries self-correct; the retraction is usually the load-bearing half.
A compound criterion (OR / AND) is refuted only by falsifying every leg. Re-measuring one leg says nothing about the criterion.
Never quote a string you have not grep -F'd. A fabricated quote in an append-only log outlives the correction, because later readers search for it and find it.
Have someone else attack the corrected diff before committing — one hostile reviewer is cheaper than nine gates and, measured here, caught 23 defects across two rounds where author self-review caught two. Wip-commit BEFORE that review, not after: the first completed version of this correction was destroyed uncommitted on the shared main checkout while the review ran, and uncommitted content has no reflog to recover from.
State propagation as incomplete when it is. Enumerate remaining sites from a fresh grep, never from a prior list — the entry documenting incomplete propagation is itself the likeliest place to under-scope it.

Modes (auto-detected — no flags)

Sync (invoked with a described strategic shift): apply it across the relevant docs, running every gate.
Audit (invoked with no argument): scan the six docs for accumulated violations; produce a punch list, including the Gate-9 rivals-reflection pass. No writes.

Workflow

Step 0 — Resolve branch + load state + record base

Resolve branch FIRST — before any write (Step 3 writes decisions.md). Strategy docs and their decision entry land on main or a dedicated docs branch, never an unrelated feature/fix branch. Run git branch --show-current; if it is not main or a docs branch, switch first using the .claude/rules/skills.md Branch Guard wip-pattern. Doing this before Step 3 keeps the decision entry and the doc edits on the same correctly-placed branch.
Read all six docs in full: lean-canvas.md, hypotheses.md, theory-of-change.md, definitions.md, progress.md, research-programme.md.
- Doc routing map: docs/CHARTER.md — one fact, one home (the strategy docs are homes 4–8, plus 4b). Read it when unsure where a fact belongs; this skill references it rather than restating the tree.
- research-programme.md (CHARTER rule 4b) is the home for programme-level content — the hard core (K1–K7 + core-hit signatures), the negative/positive heuristic, the rivals registry, the progressivity ledger, and the stopping rule. Routing test: is it about the bets, or is it a bet? A bet carries its own falsifier and priority and belongs in hypotheses.md — rule 4 wins, first-match. The hard core is not edited by a sync. Its only two retirement routes are a recorded founder decision or displacement by a rival with corroborated novel predictions the core cannot match; a sync that appears to require a core edit is a core-hit — stop, and hand off to /slava:maintain:programme-health before writing.
- README / CLAUDE.md reflection duty. Both describe the project publicly, and neither is in this skill's write scope. After any sync that changes the wedge, the market focus, the mission framing, or the programme's self-description, check whether they now misdescribe the project and say so in the Step-6 report. README is a direct edit; CLAUDE.md routes through /slava:maintain:claude-md, never a direct edit. A silent no-check is the failure this line exists to stop — the CLAUDE.md market-focus copy was stale for 17 days and misled three agents in one session.
- Launch/GTM posture source: docs/goals.md is the public, git-versioned home of tactical GTM/funnel execution (CHARTER rule 7) — read it for the current posture. .private/docs/gtm-launch-icp-worksheet.md holds only the named residue (private — contains names; reference by name only, never copy into public docs). goals.md is not a gated home (ungated/tactical) — free-standing tactical edits to it are outside this skill entirely. But when a sync you just applied invalidates goals.md (a hypothesis you demoted is still framed there as active, a page-lead you moved is still copied there), reconcile it in the same edit, under Gates 1 and 3 only (no SINGLE-VALUE markers exist in goals.md, so Gate 8 does not apply). Detect-and-hand-off was the prior rule; it produced two recorded hand-offs and zero edits (see Step 6).
Read the most recent ~50 lines of decisions.md for live context — and widen the read if Gate 2's git log -S later surfaces an older relevant decision.
Record the base commit: git rev-parse HEAD — used in Step 4 to detect concurrent drift before applying.
Resolve mode: argument present → Sync; absent → Audit.

Step 1 — State the delta (Sync) / scope the scan (Audit)

Sync: in 1–3 sentences, state the strategic change and exactly which docs + sections it touches. A pivot from conversation must be expressible as "X was true; now Y, because <evidence>." If you cannot name the evidence, stop — that is Gate 5/7 failing before you start.
Audit: skip to Step 2 against current doc contents.

Step 2 — Run the 9 gates

Run each gate against the proposed edit (Sync) or current docs (Audit). Produce a gate report (PASS / FIX / WARN per gate). A FIX blocks the write until resolved — with one carve-out below.

META-RULE (decisions.md 2026-08-05 [process]) — no gate may block a write, only require a label. A FIX may block a write that is wrong (a broken anchor, a mislabeled status, a stale number). It may never block a write for being unevidenced. Where a gate's only objection is "you have no field datum for this," the write proceeds carrying PROPOSED-PENDING-CONTACT (or UNTESTED + a falsifier) — the gate warns, flags, questions, and requires the label; it does not restrict. Rationale: a rule that stops the founder writing does not stop him acting; it only desyncs the docs from what he is already doing, which is worse than the drift the gate prevents. Any instruction in this skill that reads as "no evidence → no write" is void on sight. Sole exception: .claude/rules/pii.md (irreversible external harm). Every gate that runs a grep or command must quote its actual output in the report (matched lines, the anchor checked, or "0 matches") — a verdict with no quoted artifact is treated as not-run, not PASS.

Fix the axis before the gates — two orthogonal dimensions, neither is epistemic status, neither pre-blocks a write (this argument has recurred 3+ times; decisions.md 2026-07-03 [process]):

Edit permission (may I change this box?): destructive-vs-additive (Gate 1 / Gate 4; decisions.md 2026-06-30). Additive, UNTESTED-labeled, falsifier-carrying content is always recordable — validation is NOT a precondition.
Routing (which doc holds it?): structural/durable vs tactical/volatile — the structural bet (customer, value-prop, mechanism, and the channel bet of who delivers, e.g. coaches; change = pivot) → strategy docs; tactical mechanics (funnel choreography, prices, per-step targets, event run-of-show — including channel mechanics like a specific webinar format; change = every experiment) → decisions.md / goals.md / .private. Reason = churn-isolation, NOT epistemic rank. Advisory, and it never withholds content: default is to RECORD it (with a Gate-5 falsifier + honest UNTESTED label — that requirement is the churn/scope brake, so "record freely" is not "dump anything"); routing is a non-blocking note on which doc, never a reason to not write. If a piece is genuinely ambiguous between structural and tactical, record it where the founder wants and move on — do not argue placement (that argument IS the recurrence).

STOP — retired axis. If you are about to keep content OUT of a strategy doc because it is "unvalidated," "unrun," or "not yet validated" — that axis is DEAD (decisions.md 2026-06-30 + 2026-07-03). Everything in the strategy docs is an unvalidated bet. The only questions are destructive-vs-additive (edit permission) and structural-vs-tactical (routing). Recording is never blocked, only routed — and routing is advisory. Excluding on structural grounds is ALSO not a block: if the founder wants it in the canvas, it goes; a routing recommendation is a note, not a veto, and must never read to the founder as a refusal to record. Do not re-derive this; do not re-litigate it with the founder.

Gate 1 — Premature-fact (per-doc source of truth). Scope (read first): this gate governs the accuracy of epistemic status labels and numbers, not whether reasoned-but-untested content may be recorded at all. A box MAY be rewritten with a deductively-reasoned model so long as it is honestly labeled (UNTESTED / hypothesis) and carries a falsifier (Gate 5) — recording it is permitted; mislabeling it as Validated / Active / proven is what this gate blocks. The discriminator for may I update this box is destructive-vs-additive, NOT validated-vs-untested (decisions.md 2026-06-30 [process]) — empirical validation is not a precondition for recording. With that scope set: every status word (Active, Validated, proven, ✅) and hard number (price, R₀, %, count) must trace to real evidence:

lean-canvas.md / theory-of-change.md → must agree with that doc's Validation Status block.
hypotheses.md → must agree with the hypothesis's evidence rows / Transform-if.
definitions.md → no status words at all (it is a glossary). Flag any.
progress.md → numbers live ONLY inside the  block and regenerate via ./scripts/progress-refresh.sh --write; a hand-edited number anywhere in the doc is a FIX. Bets-and-kills rows must agree with hypotheses.md status + Transform-if. The human-stamped self-discipline line (untested-hypothesis date) must be re-checked against hypotheses.md on every sync.
A milestone label (e.g. C1) is NOT a customer. Unvalidated → label "theory / untested bet", never "Active".
A doc with no Validation block to check against → WARN (never a silent pass).
Agreeing with the Validation block is necessary but not sufficient — Gate 7 still requires an external source. If a number is internally consistent but unsourced, Gate 7 wins (label it "unverified").

Gate 2 — Reversal-lock (WARN-only for the unevidenced case; see the META-RULE above). Before rewriting a named construct or settled position:

# (a) in-file lock markers — case-sensitive + boundary-anchored so *unlocked* / *Unblocked* don't false-match
grep -rnE '(^|[^[:alnum:]])(LOCKED|DECISION-LOCK):' docs/lean-canvas.md docs/hypotheses.md docs/theory-of-change.md docs/definitions.md docs/progress.md
# (b) THE REGISTRY — rulings live in decisions.md, which (a) does NOT read. Skipping this is how the gate failed 2026-08-05.
grep -n 'PROPOSED-PENDING-CONTACT\|slot is frozen\|Rejected on the merits\|NOT applied' docs/decisions.md | head -20
# (c) history, keyed on the SLOT NAME — not on a phrase you chose
git log -S "<slot-name>" --oneline -- docs/ | head      # e.g. active-market-focus / page-lead / active-channel

Key on the slot name, never on your own construct phrase. If the edit region sits under a  marker, <slot> is the retrieval key — the marker names it, so the agent cannot pick wrong. A free-text git log -S "<construct phrase>" pass is circular and measured to fail: on 2026-08-05, git log -S "market focus" did not return the ruling that froze that slot, while git log -S "10-50" — the refusal's own vocabulary — did. Only the refusal's wording retrieves the refusal, which is useless when you don't already know it exists.

Zero hits on (a) means "no in-file lock, proceed" — not "gate passed". (b) and (c) run regardless. If any of the three surfaces a prior ruling that settled this the other way, read it in full before writing.

Then apply the META-RULE to what you found:

The prior ruling is factually superseded (its evidence was wrong, its anchor moved, its number was corrected) → FIX. Fix it, cite what changed.
The prior ruling stands and your only counter-argument is reasoning without a field datum → WARN, never FIX. Write the change, tag it [PROPOSED-PENDING-CONTACT <date>], keep the incumbent as the active answer, and name in the Step-3 decision entry what datum would settle it. Do not withhold the reasoning — an unrecorded conclusion is the failure this gate now exists to prevent, not the one it prevents.

When you settle a position you expect to hold, write LOCKED: <date> — <reason> next to it and record it in decisions.md — (a) alone is not durable, since a ruling written only in prose is invisible to this gate.

Gate 3 — Cross-doc. (a) Every anchor/section this edit links to must still exist — grep the target heading; a link to a missing heading is a FIX. (b) No sibling doc may assert the negation of what you're writing. Don't grep your new sentence — a contradiction is rarely a near-duplicate. Instead: list 3–5 key nouns in the new claim (e.g. coach, distribution, spread), then for each run grep -rniE '(not|without|never|no|isn.?t) .{0,30}<noun>' docs/ and read every hit in full. Fix or remove the contradiction in the same edit. (This is what catches theory-of-change "spreads NOT through coaches" vs a coach-distribution pivot — grepping the literal new wording never surfaces it.)

Gate 4 — Bloat (propose-only deletion). An addition that supersedes an existing dated block must delete or merge that block — but:

Never auto-delete. Show each proposed cut; get confirmation.
A block is "superseded" only if the newer block carries explicit supersede intent OR the user confirms it. Blocks marked as deliberate alternates ("alternates, not a sequence") are kept.
Never delete a block containing a falsifier / Transform-if — that content is load-bearing by Gate 5.
Additive UNTESTED updates are NOT bloat. Rewriting a box with current best thinking is permitted (see Gate 1 scope); only the destruction of a revivable fallback is a FIX. Do not block an additive, honestly-labeled rewrite.
Active-on-top / dormant-below check (mechanical FIX). When a box's active content is relabeled to a new (UNTESTED) market/construct, the prior version must survive as a ### Dormant (...) sibling in the same box. Grep the box for a Dormant heading: a box carrying a new active framing but no dormant-below sibling for a construct that previously had one = FIX (revivable fallback destroyed — the one thing the discriminator still guards). Quote the grep result ("0 matches" where one is expected IS the FIX).
Erosion guard (WARN). If across syncs the count of UNTESTED-labeled active blocks climbs while none are promoted (validated) or pruned, flag WARN — the label may be laundering scrutiny rather than carrying honest epistemic weight (the failure mode in decisions.md 2026-06-30 [process] falsifier).
(A coarse, line-granular net-bloat count is also flagged in scripts/pre-commit-checks.sh — a hygiene signal that over-counts multi-line blocks, not this gate's judgment.)

Gate 5 — Disproof. Every strategic claim being written carries a one-line falsifier — "how we'd know this is wrong" or a Transform-if. lean-canvas positioning claims are the usual offender; hypotheses.md is the model to copy.

Gate 6 — Fragile refs. No file.md:NNN or L123 line-number cross-references — use quoted-phrase anchors. (Also enforced mechanically in scripts/pre-commit-checks.sh.)

Gate 7 — Numbers cite a source. Any prod number (R₀, count, %, conversion, price-as-validated) must cite an evidence source — a Mixpanel query, a prod row, a dated session note — not "a number mentioned this turn." No source → label "unverified", never "Validated". Internal consistency is not truth.

Gate 8 — Single-valued-slot reconciliation. A single-valued slot is a strategy-doc heading that must hold exactly one current answer (the page hero, the active channel bet, the active market focus) — as opposed to a bet-list like hypotheses.md, where accumulation is correct. Gates 1/4 are blind to this: three "lead the page with X" directives can each be UNTESTED-labeled (Gate 1 pass), carry a falsifier (Gate 5 pass), and not negate each other (Gate 3's negation-grep misses non-contradicting directives) — so they silently pile up (the §UVP drift: 2026-07-04 "Get to PMF faster" + 2026-07-11 "divergent-AI hook" both marked "lead"). This gate is deterministic because single-valued slots are tagged in the doc:

<!-- SINGLE-VALUE: page-lead -->

placed immediately under the heading (convention documented in docs/CHARTER.md). A single-valued directive is a standalone dated blockquote callout > **Name (2026-07-04, …).** …; a > - **…** list bullet is an elaboration of the one answer, not a competing lead (excluded).

Run it:

python3 scripts/check-single-value-slots.py docs/lean-canvas.md docs/hypotheses.md docs/theory-of-change.md docs/definitions.md docs/progress.md

It counts unreconciled dated directive callouts under each SINGLE-VALUE marker (a callout carrying a structured SUPERSEDED / FALLBACK / dormant / parked / demoted token on its lead-in is treated as reconciled/subordinate, not counted). Exit 2 = a slot has ≥2 competing leads. Quote the script output (per this skill's "verdict with no artifact = not-run" rule).

On exit 2 → FIX: surface the competing directives as an A/B/C reconcile-to-one for the founder (never auto-resolve — "never fill in positioning without being told"), then mark the losers with [SUPERSEDED <date> → …] or subordinate them explicitly (primary + [FALLBACK …]). Adding a new directive to a SINGLE-VALUE slot without superseding/subordinating the incumbent is the exact FIX this gate blocks. This is the same logic the pre-commit check-single-value-slots.py canary enforces at commit time (defense-in-depth).

Gate 9 — Novel-prediction (degeneration guard). WARN-only, never a FIX. Gates 1–8 all check whether a claim is honest. None of them checks whether the programme is going anywhere: a hypothesis can be correctly labeled UNTESTED, carry a clean falsifier (Gate 5), cite its numbers (Gate 7), and still be a pure accommodation — a reframe that explains only the anomaly that killed its predecessor and predicts nothing new. A programme built entirely of those degenerates while passing every other gate on this list.

Applies to: any hypothesis being added or materially updated in hypotheses.md by this sync — a new bet, a reframe, a split, a wedge re-cut. Does NOT apply to: the existing backlog (the convention is forward-only), status-only edits, or typo/anchor fixes.

Scope the check to the ENTRY, never the file. A whole-file grep -n "Novel prediction" docs/hypotheses.md cannot fail — the Entry-conventions preamble contains the phrase, so it returns hits (measured: 4) no matter which hypothesis is being written, and the gate rubber-stamps PASS. Use the entry extractor:

# Extract ONE hypothesis entry, then check that entry alone.
entry() { awk -v id="$1" '$0 ~ "^#### " id "\\:" {f=1; next} f && /^#### / {exit} f' docs/hypotheses.md; }

entry H-YourHypothesis | wc -l                      # 0 lines => extractor missed it; fix before judging
entry H-YourHypothesis | grep -c "Novel prediction" # 0 => WARN;  >0 => PASS

# Ledger cross-check: anything corroborated but not yet promoted?
grep -n "corroborated" docs/hypotheses.md docs/research-programme.md

A zero-line extraction is "not checked", not "PASS" — same class as the whole-file grep, and the reason the entry count is printed first. (Verified both directions 2026-08-07: PASS on H-WTP-Pain and H-PopperianIncrement, WARN on H-PairsReturn and H-TopicDepthGate.)

Field present → PASS. If its status is corroborated, promote it into the progressivity ledger in the same edit, with its novelty grade (novel vs retrodiction — recorded before the test, or after?).
Field absent → WARN. Ask the founder the one question: "What does this framing predict that the one it replaces didn't?" — then write the answer, or write Novel prediction: none identified — accommodation and proceed.
"None identified" is a legitimate, recordable answer and must never block the write. Per the META-RULE above, this gate warns and requires a label; it does not restrict. An accommodation honestly labeled is a data point for the programme-health check; an accommodation silently written as a successor is the degeneration this gate exists to make visible.
Programme tag (business | scientific | both) and, where applicable, the core-adjacent flag with its ad-hoc-ness budget — same treatment: prompt, label, never block.

Gate 9 (audit mode) — rivals-reflection pass. In Audit mode only, additionally walk every row of the rivals registry and ask, per rival: did anything in the period make this rival's case stronger on our own data?

Define "the period" before you start — it is not otherwise defined in this skill, and two audits using different implicit windows are not comparable. Use the same resolution /slava:maintain:programme-health uses, so the two agree:

LAST=$(grep '^date:' ~/.claude_programme_health_last_run 2>/dev/null | awk '{print $2}' | tr -d '[:space:]')
[ -z "$LAST" ] && LAST=$(date -v-60d +%Y-%m-%d 2>/dev/null || date -d '60 days ago' +%Y-%m-%d)
echo "Rivals-reflection period: $LAST → $(date +%Y-%m-%d)"

State the resolved window in the punch list. An audit that does not name its period produces findings nobody can reproduce or diff against the last run. Output one line per rival in the punch-list shape. A rival whose evidence now looks better than ours is the single highest-value finding this skill can produce — and it is the one nothing else in the pipeline is looking for. Report no delta explicitly rather than omitting a rival; a silently skipped row is indistinguishable from a row with nothing to report.

Step 3 — Precondition: record the decision (Sync only)

Before writing any strategy-doc change, the decision + its falsifier must exist in decisions.md (on the branch resolved in Step 0). /kdd is user-triggered and cannot auto-fire, so a "capture it later" hand-off orphans the doc edit — it asserts a position with no recorded kill-condition (the exact failure Gate 5 prevents).

Already in decisions.md → cite it.
Not yet → write the entry now (the user invoked this skill = authorization), using /kdd's decisions.md format with a [product] tag and a one-line falsifier. /kdd's deeper meta-reflection still runs separately when the user calls it.

Step 4 — Apply (Sync) / report (Audit)

Concurrency check: re-run git rev-parse HEAD and git status --short docs/; if the six docs changed since Step 0's base, re-read before editing — a concurrent /kdd-at-merge or co-tenant edit must not be clobbered.
Sync — two classes of change (the auto-vs-surface contract):
- Mechanical → apply automatically: fix fragile line-refs (Gate 6), add falsifiers (Gate 5), merge exact-duplicate framings, and apply the spine already settled in the Step-3 decision entry.
- Judgment → surface to the founder as A/B/C options, never resolve silently (the "Founder decisions" rule — never fill in positioning / category / pricing without being told): every contradiction (Gate 3), every stale-fact / status reframe (Gate 1), every deletion (Gate 4, propose-only), and any positioning wording. Present each as: current text · why it's stale/contradictory · 2–3 resolution options; apply only the chosen one. Update the Validation Status block with the new evidence once resolved.
Audit: output the punch list grouped by gate, one line per violation in the shape <gate> | <doc> | "<quoted phrase or anchor>" | <suggested action> (matching the precision pre-commit Gate F already prints). A bare count is not actionable for a 50-item bloat list. No writes.

Step 5 — Self-check

Branch placement was resolved in Step 0 (decision + docs on main / a docs branch); confirm it held. Then verify every gate before writing.

Self-check (each box must cite its gate-report artifact — a ticked box with no quoted evidence counts as not-done):

[ ] Branch resolved before any write (Step 0.1) — decision + docs on main / docs branch
[ ] All six docs read this session (Step 0)
[ ] Sync delta expressible as "X was true; now Y, because <evidence>"
[ ] Gate 1: no unsourced status word / number; definitions.md carries none
[ ] Gate 2: lock grep run (output quoted); any override names new evidence
[ ] Gate 3: negation greps run (terms quoted); no anchor broken, no contradiction left standing
[ ] Gate 4: every deletion shown + confirmed; no falsifier-bearing block cut; relabeled box has a Dormant sibling (active-on-top check, grep quoted)
[ ] Gate 5: every new claim has a one-line falsifier
[ ] Gate 6: no line-number cross-refs introduced
[ ] Gate 7: every number cites a source or is labeled unverified
[ ] Gate 8: check-single-value-slots.py run (exit code + output quoted); no SINGLE-VALUE slot has ≥2 unreconciled leads
[ ] Gate 9: every added/updated hypothesis carries a Novel-prediction field (or an explicit none identified — accommodation) + programme tag; corroborated ones promoted to the progressivity ledger with a novelty grade. Audit mode: one line per rival.
[ ] README / CLAUDE.md reflection duty checked and reported (Step 0.2) — CLAUDE.md via /slava:maintain:claude-md, never directly
[ ] Decision + falsifier recorded in decisions.md (Step 3)
[ ] Concurrency re-check passed (Step 4)

Step 6 — Confirm + hand off

Report: docs touched, gate report (pass/fix/warn with artifacts), the decision entry written, and the next step (/kdd for meta-reflection; /slava:maintain:claude-md if CLAUDE.md needs a pointer).

GTM-drift detect. After applying, compare the active channel/posture in lean-canvas.md (the structural channel bet) against docs/goals.md §Active (the tactical funnel). If they name different active channels (e.g. the canvas says founder-direct while goals still runs coaches-first) that is drift — flag it, quoting the two active-channel phrases you compared.

Fix goals.md in this same edit (Gates 1 + 3 only) when the drift was caused by the sync you just applied — a status or lead you changed in a gated doc that goals.md still contradicts. If you leave it unfixed, say so explicitly in the Step-6 report and why. Hand off only .private/docs/gtm-launch-icp-worksheet.md (CHARTER rule 1 — private, contains names): "the private residue needs updating to match the strategy docs; no skill owns it — update it directly."

Why this changed (2026-07-23): the prior rule was detect-and-hand-off for goals.md too. decisions.md records two such hand-offs; neither produced the edit, and a stale page-lead copy survived in goals.md for nine days after the wedge flip. Honest limit: this binds only the path that runs inside this skill. /day and the kanban PATCH /api/goals endpoint write goals.md outside it, ungated.

Template (gate report + decision stub)

docs-strategy-update — <SYNC|AUDIT>
Branch: <main | docs-branch>    Base: <sha>

Delta: <X was true; now Y, because <evidence>>      # sync only
Docs touched: <list>

Gate report (every verdict cites its artifact — grep output, anchor, or "0 matches"):
  1 premature-fact ........ PASS|FIX|WARN  <status words/numbers + evidence row>
  2 reversal-lock ......... PASS|FIX|WARN  <lock grep result; git log -S result>
  3 cross-doc ............. PASS|FIX|WARN  <negation greps run (terms) + anchors + hits read>
  4 bloat (prune) ......... PASS|FIX|WARN  <proposed cuts, each confirmed y/n>
  5 disproof .............. PASS|FIX|WARN  <each new claim + its falsifier>
  6 fragile-refs .......... PASS|FIX|WARN  <line-ref grep result>
  7 numbers-cite-source ... PASS|FIX|WARN  <each number + its cited source>
  8 single-value-slot ..... PASS|FIX|WARN  <check-single-value-slots.py exit code + any competing leads>
  9 novel-prediction ...... PASS|WARN      <per added/updated hypothesis: the prediction, or "none identified — accommodation"; audit: one line per rival>

README/CLAUDE.md reflection: <no change needed | README edit applied | CLAUDE.md needs /slava:maintain:claude-md>

Decision recorded (decisions.md):
  ## <date> [product]: <title>
  Decision: <Y>
  Falsifier: <how we'd know this is wrong>

Applied: <files>    |    Punch list: <n items>   (audit)
Next: /kdd (meta-reflection) · /slava:maintain:claude-md (if CLAUDE.md pointer)

Quality Gates (Agent Self-Review)

[ ] Ran in the correct mode (sync vs audit) — auto-detected, not asked
[ ] Branch resolved before the decision write (Step 0.1)
[ ] Gate report produced with a verdict + quoted artifact per gate (no gate silently skipped or rubber-stamped)
[ ] No status word / number written without an evidence source (Gates 1, 7; Gate 7 wins ties)
[ ] No block deleted without showing it first (Gate 4 propose-only)
[ ] Judgment calls (contradictions, stale-fact / positioning reframes) surfaced as A/B/C options — never silently resolved; only mechanical cleanup applied automatically
[ ] Decision + falsifier in decisions.md before the doc write (Step 3)

Related Skills

/slava:maintain:programme-health — judges progressive/stagnating/degenerating from the docs alone; the hand-off when Gate 9 keeps returning accommodations, or when a refutation looks like a core-hit. Writes nothing, so it never competes with this skill for the doc layer
/kdd — owns decisions.md + meta-reflection; this skill writes the decision stub as a precondition, /kdd does the deeper pass
/slava:maintain:claude-md — gate for CLAUDE.md / .claude/rules/ changes
/slava:build:create-spec — when the output is a tracked feature, not a strategy-doc change
/slava:think:falsify — when a strategic claim itself needs first-principles stress-testing before it is written

Strategy Docs Update with Anti-Drift Gates

Recommended for