docs(frank-archon): reconcile the page with the live 1.8.1 two-slot shape
frank-archon was migrated to radiance 1.8.1 with MAX_SEQS=2 and the --vram-weights-mib/--vram-kv-mib pair removed (PR #30, merge 88342c1,
deploy green, e2e 33/33). The env, serve script and bats carried the new shape;
the runbook and the AGENTS.md block were still describing the 1.0.13 3-slot
shape, so every number an operator would read at point of use was wrong.
A draft reconciliation existed but was written against a stale base (before PR #31's vault destruction), so it was re-derived rather than applied: every claim
below was re-checked against the live checkout (scripts/frank-archon.env, scripts/serve-frank-archon.sh) and the boot ledger of the live container
(2026-10-19 11:10, pin 1.8.1, MAX_SEQS=2).
serverAppConfigs-docs/LLM/frank-archon.md
Frontmatter + capacity contract: 1.0.13 → 1.8.1, iq4r → fp8-iq4r-moe
filename, capacity contract 3 → 2 conversations, and the note that at
1.8.1 that pair is the demand signal the resolver sizes against, not a KV
allocation.
Runtime block: digest 4c3f4845… → d590e300…; version oracle is the
boot's STARTUP REPORT — radiance 1.8.1 line (the --launch-sync line the
old text cited as a version tell is gone); records the two CLI breaks that
ride the pin — the VRAM pair died at 1.3.0, --prefix-cache-dir/ --prefix-cache-disk-mib with the disk tier at 1.8.1 — and that both are boot refusals, not ignored flags, which is why a rollback to 1.0.13 is a
trio rewrite; bats asserts the retired knobs are absent.
Shape table: slots/rows as demand signal; VRAM split derived from GPU_HEADROOM_MIB=96 + cache bound + host pool; Spec/batch row restored
(SPEC_TOKENS=2, MAX_BATCHED_TOKENS=2048 — the row had been blanked);
prefix cache host-tier only, STATE_DIR now exists only to hold the rad-tune
cache; new Step-sharing row (BUSY_PREFILL_CHUNK=auto, PREFILL_INTERLEAVE=0).
Container contract: the /data rw mount is gone — the serve script mounts /models ro, the tune cache and the template only.
Eras: new row for the retired 2026-10-15 → 10-19 era (1.0.13, pair bought
the third slot, replaced because of decode sludge + the two missing knobs,
rollback = digest plus trio rewrite).
Laws: pool-sizing and per-card-budget laws re-scoped to the explicit-pair
era and kept as causal evidence; new resolver law with the three-boot funding
table (93 % / byte-identical 93 % / 91 %) and the finding that host RAM below
a slab step (640 MiB) is inert and MAX_SEQS is a demand signal, not an
allocation; sludge law marked as measured at the 3-slot shape with the fix ordered but not verified at 1.8.1; "the knobs are not in this pin"
rewritten to "they arrived with 1.8.1 and only one is on"; slots law now
prices the third slot in KV (4.20 → 2.76 GiB) rather than rows; the EMULATION
double-billing law promoted to its own bullet with its dead crosslink tail
removed.
AGENTS.md
frank-archon block: digest, --max-num-seqs 2, --busy-prefill-chunk auto,
prefix cache host-tier only, and the explicit statement that the VRAM pair and
the disk tier are retired and naming either is a boot refusal.
§9 keeps the method and the ratios; the vendor's absolute t/s ladder and the
TTFT delta triple (measurements of a shape neither deployed nor compared) are
destroyed per the purpose test — what survives is "halving the shared chunk
costs roughly a fifth of prefill throughput" and "the span win is
long-prompt-only". Run on: now names the era the census was taken in.
Verification
./tests/run.sh — full seam suite (incl. docs link integrity, OKF, homepage,
LLM unit integrity, service-scoped planning).
Every number on the page traced to scripts/frank-archon.env, serve-frank-archon.sh
or the 11:10 boot ledger of the running container; nothing re-derived by
arithmetic.
No config, script or test changed — docs only, so deploy.sh plans nothing.
# docs(frank-archon): reconcile the page with the live 1.8.1 two-slot shape
`frank-archon` was migrated to radiance **1.8.1** with `MAX_SEQS=2` and the
`--vram-weights-mib`/`--vram-kv-mib` pair removed (PR #30, merge `88342c1`,
deploy green, e2e 33/33). The env, serve script and bats carried the new shape;
the runbook and the AGENTS.md block were still describing the 1.0.13 3-slot
shape, so every number an operator would read at point of use was wrong.
A draft reconciliation existed but was written against a stale base (before PR
#31's vault destruction), so it was re-derived rather than applied: every claim
below was re-checked against the live checkout (`scripts/frank-archon.env`,
`scripts/serve-frank-archon.sh`) and the boot ledger of the live container
(2026-10-19 11:10, pin 1.8.1, `MAX_SEQS=2`).
## `serverAppConfigs-docs/LLM/frank-archon.md`
- **Frontmatter + capacity contract**: 1.0.13 → 1.8.1, iq4r → fp8-iq4r-moe
filename, capacity contract **3 → 2 conversations**, and the note that at
1.8.1 that pair is the *demand signal* the resolver sizes against, not a KV
allocation.
- **Runtime block**: digest `4c3f4845…` → `d590e300…`; version oracle is the
boot's `STARTUP REPORT — radiance 1.8.1` line (the `--launch-sync` line the
old text cited as a version tell is gone); records the two CLI breaks that
ride the pin — the VRAM pair died at 1.3.0, `--prefix-cache-dir`/
`--prefix-cache-disk-mib` with the disk tier at 1.8.1 — and that both are
**boot refusals, not ignored flags**, which is why a rollback to 1.0.13 is a
trio rewrite; bats asserts the retired knobs are absent.
- **Shape table**: slots/rows as demand signal; VRAM split derived from
`GPU_HEADROOM_MIB=96` + cache bound + host pool; Spec/batch row restored
(`SPEC_TOKENS=2`, `MAX_BATCHED_TOKENS=2048` — the row had been blanked);
prefix cache host-tier only, `STATE_DIR` now exists only to hold the rad-tune
cache; new Step-sharing row (`BUSY_PREFILL_CHUNK=auto`, `PREFILL_INTERLEAVE=0`).
- **Container contract**: the `/data` rw mount is gone — the serve script mounts
`/models` ro, the tune cache and the template only.
- **Boot ledger** replaced with the measured 2-slot ledger: `elastic 26.73 GiB`
→ experts 23.73 / kv 2.76, `kv_attn` 26.00 KiB → 91,301 blocks = **365,204
tokens = 1.8 sessions (91 % of the 2 × 200,000 worst case)**, mover lends
2,146 slots, sampler 9 rows, template 28,234 B, vision ≤ 4,194,304 px.
- **Eras**: new row for the retired 2026-10-15 → 10-19 era (1.0.13, pair bought
the third slot, replaced because of decode sludge + the two missing knobs,
rollback = digest **plus trio rewrite**).
- **Laws**: pool-sizing and per-card-budget laws re-scoped to the explicit-pair
era and kept as causal evidence; new resolver law with the three-boot funding
table (93 % / byte-identical 93 % / 91 %) and the finding that host RAM below
a slab step (640 MiB) is inert and `MAX_SEQS` is a demand signal, not an
allocation; sludge law marked as measured at the 3-slot shape with the fix
**ordered but not verified** at 1.8.1; "the knobs are not in this pin"
rewritten to "they arrived with 1.8.1 and only one is on"; slots law now
prices the third slot in KV (4.20 → 2.76 GiB) rather than rows; the EMULATION
double-billing law promoted to its own bullet with its dead crosslink tail
removed.
## `AGENTS.md`
frank-archon block: digest, `--max-num-seqs 2`, `--busy-prefill-chunk auto`,
prefix cache host-tier only, and the explicit statement that the VRAM pair and
the disk tier are retired and naming either is a boot refusal.
## `serverAppConfigs-docs/LLM/LLM Performance Diagnosis.md`
§9 keeps the method and the ratios; the vendor's absolute t/s ladder and the
TTFT delta triple (measurements of a shape neither deployed nor compared) are
destroyed per the purpose test — what survives is "halving the shared chunk
costs roughly a fifth of prefill throughput" and "the span win is
long-prompt-only". `Run on:` now names the era the census was taken in.
## Verification
- `./tests/run.sh` — full seam suite (incl. docs link integrity, OKF, homepage,
LLM unit integrity, service-scoped planning).
- Every number on the page traced to `scripts/frank-archon.env`, `serve-frank-archon.sh`
or the 11:10 boot ledger of the running container; nothing re-derived by
arithmetic.
- No config, script or test changed — docs only, so `deploy.sh` plans nothing.
Build worktree: `/home/blank/.omp/wt/wt-archon-docs` (branch
`docs/frank-archon-181-two-slot`).
---
cwd: /home/blank/.omp/wt/wt-20261019-020309-6d1fc1f · session: 01a15366-7601-7175-bb69-c62186c3f8b5
The endpoint moved to radiance 1.8.1 with MAX_SEQS=2 and the
--vram-weights-mib/--vram-kv-mib pair removed (PR #30); env, serve script and
bats carried the new shape while the page and the AGENTS.md block still
described the 1.0.13 3-slot era, so every point-of-use number was wrong.
Re-derived from the live checkout and the 2026-10-19 11:10 boot ledger, not
from the stale pre-#31 draft: runtime block (digest, version oracle, the two
CLI breaks that are boot refusals), shape table (demand-signal rows, derived
VRAM split, restored spec/batch row, host-only prefix tier, step sharing),
container contract (no /data mount), the measured 2-slot ledger, a new Eras row
for the retired 1.0.13 era whose rollback is a digest plus a trio rewrite, and
the laws: pool sizing and the per-card budget re-scoped to the pair era, a new
resolver law (funds ~90% of the worst case, host RAM below a slab step is inert,
MAX_SEQS is a demand signal), the sludge law marked measured-at-3-slots with the
fix ordered-not-verified, the knob-availability law flipped, and the slots law
now pricing the third slot in KV (4.20 -> 2.76 GiB).
The t/s ladder and the TTFT delta triple measure a configuration neither
deployed nor compared; per the purpose test they are dead weight. What survives
is the knob pricing (halving the shared chunk costs roughly a fifth of prefill
throughput), the long-prompt-only shape of the span win, and the method. Run-on
now names the era the census was taken in.
dank
merged commit d42d053abe into main2026-10-19 14:56:59 -07:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
docs(frank-archon): reconcile the page with the live 1.8.1 two-slot shape
frank-archonwas migrated to radiance 1.8.1 withMAX_SEQS=2and the--vram-weights-mib/--vram-kv-mibpair removed (PR #30, merge88342c1,deploy green, e2e 33/33). The env, serve script and bats carried the new shape;
the runbook and the AGENTS.md block were still describing the 1.0.13 3-slot
shape, so every number an operator would read at point of use was wrong.
A draft reconciliation existed but was written against a stale base (before PR
#31's vault destruction), so it was re-derived rather than applied: every claim
below was re-checked against the live checkout (
scripts/frank-archon.env,scripts/serve-frank-archon.sh) and the boot ledger of the live container(2026-10-19 11:10, pin 1.8.1,
MAX_SEQS=2).serverAppConfigs-docs/LLM/frank-archon.mdfilename, capacity contract 3 → 2 conversations, and the note that at
1.8.1 that pair is the demand signal the resolver sizes against, not a KV
allocation.
4c3f4845…→d590e300…; version oracle is theboot's
STARTUP REPORT — radiance 1.8.1line (the--launch-syncline theold text cited as a version tell is gone); records the two CLI breaks that
ride the pin — the VRAM pair died at 1.3.0,
--prefix-cache-dir/--prefix-cache-disk-mibwith the disk tier at 1.8.1 — and that both areboot refusals, not ignored flags, which is why a rollback to 1.0.13 is a
trio rewrite; bats asserts the retired knobs are absent.
GPU_HEADROOM_MIB=96+ cache bound + host pool; Spec/batch row restored(
SPEC_TOKENS=2,MAX_BATCHED_TOKENS=2048— the row had been blanked);prefix cache host-tier only,
STATE_DIRnow exists only to hold the rad-tunecache; new Step-sharing row (
BUSY_PREFILL_CHUNK=auto,PREFILL_INTERLEAVE=0)./datarw mount is gone — the serve script mounts/modelsro, the tune cache and the template only.elastic 26.73 GiB→ experts 23.73 / kv 2.76,
kv_attn26.00 KiB → 91,301 blocks = 365,204tokens = 1.8 sessions (91 % of the 2 × 200,000 worst case), mover lends
2,146 slots, sampler 9 rows, template 28,234 B, vision ≤ 4,194,304 px.
the third slot, replaced because of decode sludge + the two missing knobs,
rollback = digest plus trio rewrite).
era and kept as causal evidence; new resolver law with the three-boot funding
table (93 % / byte-identical 93 % / 91 %) and the finding that host RAM below
a slab step (640 MiB) is inert and
MAX_SEQSis a demand signal, not anallocation; sludge law marked as measured at the 3-slot shape with the fix
ordered but not verified at 1.8.1; "the knobs are not in this pin"
rewritten to "they arrived with 1.8.1 and only one is on"; slots law now
prices the third slot in KV (4.20 → 2.76 GiB) rather than rows; the EMULATION
double-billing law promoted to its own bullet with its dead crosslink tail
removed.
AGENTS.mdfrank-archon block: digest,
--max-num-seqs 2,--busy-prefill-chunk auto,prefix cache host-tier only, and the explicit statement that the VRAM pair and
the disk tier are retired and naming either is a boot refusal.
serverAppConfigs-docs/LLM/LLM Performance Diagnosis.md§9 keeps the method and the ratios; the vendor's absolute t/s ladder and the
TTFT delta triple (measurements of a shape neither deployed nor compared) are
destroyed per the purpose test — what survives is "halving the shared chunk
costs roughly a fifth of prefill throughput" and "the span win is
long-prompt-only".
Run on:now names the era the census was taken in.Verification
./tests/run.sh— full seam suite (incl. docs link integrity, OKF, homepage,LLM unit integrity, service-scoped planning).
scripts/frank-archon.env,serve-frank-archon.shor the 11:10 boot ledger of the running container; nothing re-derived by
arithmetic.
deploy.shplans nothing.Build worktree:
/home/blank/.omp/wt/wt-archon-docs(branchdocs/frank-archon-181-two-slot).cwd: /home/blank/.omp/wt/wt-20261019-020309-6d1fc1f · session: 01a15366-7601-7175-bb69-c62186c3f8b5