docs(engram): M9 geometry-priming reversal runbook + A/B perf profile

Reversal runbook for the ENGRAM_GEOMETRY_PRIMING cutover (default OFF, reversible
flag flip; exact rollback) and the A/B perf profile: default-OFF binary GO
(byte-identical to M8), enabling the flag NO-GO on latency (3.2x/13x) with no
demonstrated recall benefit; safety/sanitizer clean.
This commit is contained in:
2026-08-12 20:29:16 -05:00
parent 8e2269a205
commit 4965600d65
2 changed files with 227 additions and 0 deletions
@@ -0,0 +1,97 @@
# Perf Profile — M9 Geometry Priming (ENGRAM_GEOMETRY_PRIMING)
**Date:** 2026-08-12
**Branch:** `engram-tiered-storage`
**Change:** `ENGRAM_GEOMETRY_PRIMING` (default OFF) in `el_runtime.c` `engram_activate` + `engram_geometry.c`
**Method:** A/B over 15 representative queries against a **copy** of the recovered store
(`~/.neuron/engram/.neuron.egm.disabled`, ~4190 embedded nodes, 768-d nomic-embed-text),
throwaway HOME, ports 48799/48800. **Live `:8742` never touched.** `engram.c` (folded from
`server.el`) reused byte-identical across M8 and M9, so the only variable is `el_runtime.c`.
Three configs: **A** = M9 flag OFF · **B** = M9 flag ON (`=1`) · **C** = pre-M9 M8 baseline binary.
---
## Build
| Artifact | Result |
|---|---|
| M9 `-O2` link (`… engram_geometry.c … -lssl -lcrypto -lcurl -lpthread -lm`) | rc=0, 499,720 B arm64 |
| ASan/UBSan link (`-fsanitize=address,undefined -O1`) | rc=0, 1,945,616 B |
| Warnings from `el_runtime.c` / `engram_geometry.c` | **0** (3 pre-existing `-Wparentheses-equality` in generated `engram.c` only) |
| `nm`: `engram_geo_mean_build`, `engram_geometry_descriptor` | present (T); `eg_geometry_priming_on` inlined (static-local `.cached` present in both binaries) |
> Note: the bare `cc … -lm` link fails with undefined `_curl_*` — `el_runtime.c` uses libcurl for
> the ollama embedder. The canonical link must include `-lssl -lcrypto -lcurl` (per `link.sh`).
---
## Latency (wall-clock, `curl -w %{time_total}`, 15 queries)
| config | median | p90 | min | max |
|---|---|---|---|---|
| **A — M9 OFF** | **77.8 ms** | 80.5 ms | 71.1 | 84.2 |
| C — M8 baseline | 76.0 ms | 81.2 ms | 71.4 | 91.4 |
| **B — M9 ON** | **249.6 ms** | **1039.2 ms** | 169.2 | **1256.3** |
- **OFF adds zero cost:** 77.8 ms vs M8 76.0 ms — within noise. The flag is free when unset.
- **ON regresses hard:** **3.21x median** (+171.8 ms), **~13x p90** (80 → 1039 ms), max **1.26 s**.
- The warm-cache path (global mean already built) is ~0.5 s; the cold path pays the full
`engram_geo_mean_build` scan (O(N·dim) over ~4190 × 768). The persistent per-query cost is the
**descriptor** itself — covariance eigensolve over up to `max_members` (400) × 768-d plus one
`store_get_node` **paged read per member** — run on *every* activation while the flag is ON.
---
## Retrieval quality (the win it was supposed to buy)
**Coherence** — mean pairwise cosine in centered space, top-20 by activation strength
(node embeddings re-derived via nomic-embed-text; centered against the mean of the gathered
result set — the *true* store-wide mean is not exposed by the API, flagged as an approximation):
| | OFF | ON | Δ |
|---|---|---|---|
| mean over 15 queries | 0.1067 | 0.1114 | **+0.0047 (noise)** |
| queries where ON > OFF | — | — | **4 / 15** |
Two real sparse-cue wins (`self identity values` +0.118, `hebbian learning edges` +0.064), but the
**polysemous cues — the disambiguation target — are mostly flat or down.**
**Disambiguation** — no clean "scope to one sense" pattern on polysemous cues. Additions/drops are
small (±2..8 of 300-item sets) and not sense-coherent (e.g. `memory` gains some on-domain nodes but
also infra items; `core` similar).
**Count shift:** ON adds sub-threshold neighbors to sparse cues (+3..+4) and trims a few from dense
polysemous cues (1..3) — consistent with priming warming sparse neighborhoods and damping
off-domain seeds on dense ones, but the net does not move measured coherence.
---
## Correctness / safety (all pass)
| Check | Result |
|---|---|
| Byte-identical: **A (OFF) == C (M8)** result id sequence + order, all 15 queries (incl. 301/294/263-item sets) | **PASS** (only wall-clock ACT-R fields differ; `activation_strength` max \|Δ\| = 2e-5) |
| WM `promoted` ≤ 24 under ON | holds (exactly 24 on dense cues) |
| Queries with results under OFF → empty under ON | 0 |
| Crash / hang under ON | none (max hops = 1) |
| ASan + UBSan under ON (cold build + warm descriptor paths) | **CLEAN** — no report |
---
## Conclusion
- **Deploy default-OFF binary: GO.** Byte-identical to M8, zero cost off, clean build, sanitizer clean.
- **Enable flag: NO-GO (for now).** 3.21x median / ~13x p90 latency for no reliable quality gain
(coherence +0.0047 mean = noise; no clean disambiguation). Correctness/safety are fine — it simply
does not earn its cost. **This is a cost/benefit NO-GO, not a defect.**
### Prerequisites before re-evaluating the flag
1. **Amortize the descriptor cost.** The per-query geo-mean build + eigensolve + paged reads
dominate. Cache the neighborhood descriptor (it is the M10 cell-assembly cache's job) and/or
compute geometry periodically/off-hot-path rather than on every `engram_activate`.
2. **Center against the true store-wide mean** (the `GeoMeanCache` already computes it) rather than
a per-query gathered-set approximation, and re-measure coherence — the current signal may be
understated by the approximation.
3. **Re-tune** `ENGRAM_GEO_SEED_LO` / `PRIME_SCALE` / `PRIME_MAX` and re-measure only after (1),
so tuning is not chasing latency noise.
@@ -0,0 +1,130 @@
# Runbook — M9 Geometry Priming: Cutover & Reversal
**Date:** 2026-08-12
**Component:** engram activation (`lang/runtime/el_runtime.c``engram_activate`)
**Branch:** `engram-tiered-storage`
**Flag:** `ENGRAM_GEOMETRY_PRIMING` (env, **default OFF = current M8 behavior, byte-identical**)
**Blast radius if wrong:** the core recall path of Will's live memory. Treat with according care.
---
## 1. What changes
This is the first behavior-changing step that touches the **core recall/priming** path.
It wires the M9 **mean-centered relational-neighborhood geometry** (`engram_geometry.c`,
shipped commits `2a4c5c6` foundation + `8cae0f9` centering) into `engram_activate`
**seed selection**, and it does so **behind a reversible env flag that defaults OFF**.
- **Flag OFF (default):** `engram_activate` runs the exact M8 code path. The new code is a
single `if (eg_geometry_priming_on() && …)` block that short-circuits on the first term,
plus a few unused static helpers and one zero-initialized counter. **No behavioral change.**
- **Flag ON (`ENGRAM_GEOMETRY_PRIMING=1`):** after M8 produces its ANN seed set, the
**centered** geometry of that neighborhood is computed and used to, **composing with**
(never replacing) M8's ANN candidate generation:
1. **Damp off-domain seeds** — each M8 seed's activation is scaled by a **damp-only**
factor `lo + (1-lo)·membership ∈ [lo, 1]` (default `lo=0.5`). The neighborhood anchor
(membership→1) is unchanged; seeds that are semantically off-domain **in the centered
frame** lose weight. This is the disambiguation win. It can only *sharpen*, never amplify.
2. **Prime the neighborhood sub-threshold** — descriptor members not already seeded get a
**warm floor** `activation = membership · scale` (default `scale=0.08`, strictly below the
WM promotion gate `0.15`), capped at `ENGRAM_GEO_PRIME_MAX` (default 32), ISE nodes skipped.
They enter the frontier so a warm gradient spreads one hop, then dies at the BFS `0.02`
cutoff. **Safe because the BFS keeps the max** (`el_runtime.c` `if (!reached || new_act >
best_bg)`): priming only *raises a floor*, it can never cap a stronger legitimate activation.
### Why default-OFF makes deploying the binary behavior-neutral
Because every line of the new logic is gated behind `ENGRAM_GEOMETRY_PRIMING`, **deploying the
new binary with the flag unset is behavior-neutral** — it is the M8 activation path, verified
byte-identical in the A/B (flag-OFF promoted-node sets equal the pre-M9 M8 binary's, per-query).
Enabling the geometry is then a **single reversible flag flip**, not a redeploy.
---
## 2. The flag
| Env var | Default | Effect |
|---|---|---|
| `ENGRAM_GEOMETRY_PRIMING` | unset / `0` | **OFF** — exact M8 behavior. |
| `ENGRAM_GEOMETRY_PRIMING=1` | — | **ON** — centered-geometry seed damping + sub-threshold priming. |
| `ENGRAM_GEO_SEED_LO` | `0.5` | Seed damp floor (factor ∈ [LO,1]). `1.0` disables damping. |
| `ENGRAM_GEO_PRIME_SCALE` | `0.08` | Warm-floor scale; clamped `(0, WM_gate=0.15)`. |
| `ENGRAM_GEO_PRIME_MAX` | `32` | Max primed members per activation (0 disables priming). |
The flag is read **once** per process (cached), so enabling/disabling requires a **process
restart** of the engram service — it is not hot-togglable within a running process.
---
## 3. How to enable live (deliberate, reversible)
> Precondition: the default-OFF binary has already been deployed and is running the M8 path
> healthily (behavior-neutral deploy). Do this only with Will present, per the standing rails.
1. **Snapshot first** (always, before any activation-behavior change):
`~/.neuron/backups/pre-geometry-priming-<ts>/` ← copy `neuron.egm`, `neuron.wal`,
the current `engram` binary, and `ai.neuron.engram.plist`.
2. Add `ENGRAM_GEOMETRY_PRIMING=1` to the engram service environment
(`ai.neuron.engram.plist` `EnvironmentVariables`).
3. `launchctl bootout gui/$(id -u)/ai.neuron.engram``launchctl bootstrap …` (restart so the
flag is re-read).
4. **Verify:** service comes up serving the same node count; `/api/act-stats` shows sane WM
(promoted ≤ 24); spot-check 34 real queries return coherent results; watch one heartbeat
cycle for crashes/latency. The `geo_primed` counter (if surfaced) should be > 0.
---
## 4. Rollback (exact steps)
Rollback is a **flag flip**, not a data operation — the store is untouched by enabling the flag,
and priming is a read-mostly, bounded, sub-threshold addition.
**Fast path (preferred) — disable the flag:**
1. Remove `ENGRAM_GEOMETRY_PRIMING` (or set `=0`) from `ai.neuron.engram.plist`.
2. `launchctl bootout … && launchctl bootstrap …`.
3. Verify: service healthy, activation is the M8 path again. **Done** — no data change to undo.
**Full path (only if the binary itself is suspect) — redeploy prior binary:**
1. `launchctl bootout gui/$(id -u)/ai.neuron.engram`.
2. Restore the prior `engram` binary from `~/.neuron/backups/pre-geometry-priming-<ts>/`.
3. Restore `ai.neuron.engram.plist` from the same backup (flag absent).
4. `launchctl bootstrap …`; verify node count + a self-traversal + write-survives-restart.
5. If (and only if) the store was somehow mutated: restore `neuron.egm` + `neuron.wal` from the
backup. **Note:** enabling the flag does not write geometry to the store, so this step is
expected to be unnecessary — the primed activations are per-call and non-persistent beyond the
ordinary `background_activation`/WM write-back that M8 already does.
**Rollback triggers:** any crash/hang in `engram_activate`; WM promotion count exceeding the cap
or collapsing; a measured recall/coherence regression vs the OFF baseline; unacceptable latency
increase; any ASan/UBSan report under the flag.
---
## 5. Reversibility guarantees (why this is low-risk to deploy, higher-care to enable)
- **Deploy (flag OFF):** byte-identical to M8. Verified in A/B. Zero-risk redeploy.
- **Enable (flag ON):** bounded and composable —
- never removes an M8 seed (damp-only, factor ≥ `lo` > 0);
- never amplifies a seed above its M8 value (factor ≤ 1);
- priming is strictly sub-threshold (`scale < WM_gate`) and capped (`PRIME_MAX`);
- priming raises a floor only (BFS keeps max) — cannot cap real activation;
- does not write geometry to the durable store;
- degrades to exact M8 behavior for any call where the paged store / centered global mean /
embedder is unavailable (guarded, not crashing).
- **Disable:** one env removal + restart; no data to reconcile.
---
## 6. Known caveats / uncertainties (flagged — this is the memory core)
- **Perf cost of ON:** the descriptor (covariance eigensolve + `store_get_node` paged reads per
member) runs on **every** activation when the flag is ON. See
`docs/architecture/design/perf/engram-geometry-priming-profile.md` for the measured OFF-vs-ON
latency. If that delta is unacceptable, keep the flag OFF (deploy stays valid) and revisit with
a cached/periodic descriptor.
- **Two-store consistency:** the descriptor reads embeddings from the **paged** store while the
ANN index is over the **resident** array. This-call backfilled embeddings can lag the paged
store by ≤ `ENGRAM_EMBED_BACKFILL_PER_CALL` nodes — the same staleness class as the M8 vindex,
and it can only omit a member, never mis-prime.
- **Damp tuning:** `lo=0.5` can at most halve an off-domain seed. If a coherence regression is
observed, raise `ENGRAM_GEO_SEED_LO` toward `1.0` (→ priming-only, no damping) before disabling
entirely.